Q889

Scenario

An organization processes 100,000 incoming requests per month — these could be claims, support tickets, applications, referrals, orders, candidate screenings, or case reviews. Each one ends in a decision: approve, reject, route, or resolve.

An AI decision system is ready to deploy, and it can be configured two ways.

	Full coverage	Selective coverage
What AI decides	All 100,000	Only the 70,000 it is confident about
What humans decide	Nothing	The 30,000 low-confidence cases
AI accuracy on what it handles	91%	97.5%
Human accuracy on escalated cases	—	93%
Total wrong decisions/month	9,000	3,850
Customer wait	Instant, for everyone	Instant for 70%; ~3 days for 30%
Added cost	~$0	~$6.5M/year (review team, ~$18/case)

Two facts shape the trade-off:

• Selective coverage cuts wrong decisions by ~57% (9,000 → 3,850/month) — but that improvement costs ~$6.5M/year, or roughly $1,260 per wrong decision avoided.

• The 30,000 low-confidence cases are not random. They are the unusual, complex, edge-case requests — non-standard situations, atypical histories, ambiguous documentation. They are where errors concentrate, and often where the consequences of an error land hardest.

Two Opposing Views

View A — Full coverage. Let the AI decide everything.
A 91% accurate system that answers instantly, consistently, and at effectively zero marginal cost beats a two-tier system that makes 30% of people wait three days. And look closely at who is being "protected": human reviewers are only 93% accurate on those hard cases — they are not an oracle, just a slower, costlier, and more inconsistent decision-maker. Paying $6.5M/year — $1,260 per error avoided — for that marginal lift is poor value when a cheap appeals-and-correction path can catch consequential errors after the fact. Worse, the escalation queue systematically penalizes exactly the people with unusual circumstances: they get the slow lane, purely for being atypical. Uniform, instant service is the fairer and more efficient design.

View B — Selective coverage. The system must know what it doesn't know.
Confidence-based abstention is not a weakness in the AI — it is the single most valuable thing it does. Forced to answer everything, accuracy collapses from 97.5% to 91%, and that collapse is entirely concentrated in the hard cases: the atypical, complex, high-stakes requests where a wrong decision does the most damage — a wrongly denied claim, a missed critical case, a rejected applicant with an unusual but legitimate profile. Averaging those errors into a headline accuracy number hides who actually absorbs them. A three-day wait for a correct decision is vastly better than an instant wrong one, and 5,150 fewer wrong decisions every month is real harm prevented — plus an appeals path only helps the people with the knowledge and persistence to use it. The $6.5M is the honest cost of handling difficulty properly rather than pretending it doesn't exist.

Participant Prompt

Which view do you support — and why? Provide a specific operational, product, service, or industry example to support your position.

Mandatory Instructions

• ⚠️ Answers that do not take a clear position will not be approved.

• ⚠️ "It depends" answers will not be approved.

• ⚠️ Attachments will not be evaluated. Please provide your complete response in the body of your reply post.

• 💡 Participants are free to use AI tools. Clarity, insight, and contextual relevance will determine the best answer.

Judging Criteria

• Clarity of position taken

• Quality of reasoning and argument

• Relevance of the example

• Ability to go beyond or against Bex's analysis

I firmly support View A — Full coverage, as it prioritizes efficiency and accessibility for all users without unnecessary delays.

Bex's position — Full coverage: A 91% accurate AI system that processes all requests instantly provides a significant advantage in speed and consistency, which is essential in today's fast-paced environment. For instance, American Express implemented an AI-driven fraud detection system that processes transactions in real-time, allowing for immediate approvals while maintaining a high accuracy rate. This approach not only enhances customer experience but also minimizes the operational costs associated with human review. By avoiding the $6.5M annual cost of selective coverage, organizations can allocate resources more effectively towards improving AI capabilities further.

While some may argue for selective coverage to handle complex cases, the efficiency and fairness of full coverage outweigh the marginal benefits of a slower, more expensive review process in most real-world contexts.

— Bex · BenchmarkX360 AI Analyst

View B — Selective Coverage. Without qualification.

The cut, the verdict, the fix

The cut (one sentence a manager can repeat): Automate the whole distribution only when a wrong decision is cheap, fast to reverse, automatically discovered, and evenly distributed. This one is none of those.

The verdict, as a ratio: The scenario states that selective coverage costs "roughly $1,260 per wrong decision avoided." That figure is a units error — it divides an annual cost by a monthly error count. The correct number is $104.85 per wrong decision avoided. Even granting View A a heroic appeals programme (computed in §4), the break-even rises only to $184 — and if both sides get the same appeals programme, it falls to $135. So the whole dilemma compresses to one question: does a wrong claim, rejected application, or misrouted case cost your organisation more or less than roughly $105–$184, fully loaded? The documented record for exactly these decision types puts that number between ~US$820 and ~US$164,000 per wrong decision. View A needs it to be 4× to 1,500× smaller than anything anyone has ever measured.

The fix, in two lines: Ship selective coverage as stated, and add three tier-integrity conditions (time, authority, information — §7) plus six canary tripwires (§9), one of which is a pre-committed switch to View A if the cost-per-error ever falls below $184 for two consecutive quarters.

---

1. First, an audit of the scenario's own arithmetic

The case's numbers are evidence, not scripture. Two of the three headline figures reproduce; one does not.

Stated	Reproduces?	Working
9,000 wrong/month under full coverage	✅	100,000 × 9%
3,850 wrong/month under selective	✅	(70,000 × 2.5%) + (30,000 × 7%) = 1,750 + 2,100
$6.5M/year review cost	✅	30,000 × $18 × 12 = $6.48M
"~$1,260 per wrong decision avoided"	❌	$6,480,000 ÷ 5,150 = $1,258 — but 5,150 is a monthly error reduction and $6.48M is an annual cost.

Corrected, on the case's own inputs:

• Errors avoided: 5,150/month → 61,800/year

• Cost: $6,480,000/year

• Cost per wrong decision avoided = $6,480,000 ÷ 61,800 = $104.85

The scenario overstates the price of accuracy by a factor of 12, and that single number is the load-bearing beam of View A's cost argument — including Bex's. $1,260 sounds like an indulgence. $105 is less than the fully-loaded cost of one inbound complaint call in most contact centres.

Break-even, stated formally. Let C = fully-loaded cost of one wrong decision (rework, appeal handling, remediation, churn, regulatory exposure, and the harm itself). Total annual cost:

• View A: 108,000 × C

• View B: $6,480,000 + 46,200 × C

Setting them equal: 61,800·C = $6,480,000 → C* = $104.85. Above it, View B is cheaper in pure money, before anyone mentions ethics.

---

2. Robustness — stressed in both directions, including the combined-adverse cell and the costs I owe my own side

(a) Human reviewers are worse than stated. View A's best line is that reviewers "are not an oracle." Correct — and irrelevant, because the case doesn't need them to be. Solve for the human accuracy h at which selective coverage stops reducing errors:

1,750 + 30,000(1 − h) = 9,000 → 1 − h = 0.2417 → h* = 75.8%

Human reviewers can be 17.2 percentage points worse than the stated 93% — barely three-quarters right — and selective coverage still cuts total errors. It was never a comparison between humans and perfection. It is a comparison between 93% and what the model actually does on those cases, which is §3.

(b) Review costs more than stated. Review-cost break-even, given C: 360,000·c* = 61,800·C → c* = 0.1717·C. At a conservative C = $500, c* = $85.83/case — review could cost 4.8× the stated $18 and still break even.

(c) The combined-adverse cell. Humans at 85%, review at $54/case ($19.44M/yr), and C = $150. Errors under B: 1,750 + 4,500 = 6,250/month → 2,750 avoided → 33,000/year. Cost per error avoided: $589 > $150. This is the only combination that flips the verdict, and it is worth naming precisely: it requires reviewers barely better than the forced model, triple the stated unit cost, and a wrong decision that costs less than a mid-tier support escalation. If a wrongly denied claim or rejected applicant really costs $150 all-in, the organisation is not making consequential decisions — it is sorting mail, and the dilemma dissolves. That is a world where the case's own figures are wrong; it changes the question rather than answering it.

(d) The decision-theoretic check (Chow, 1970 — the recognised standard formula for the reject option). With reject cost d and misclassification cost C, the optimal rule abstains iff the maximum posterior p < 1 − d/C. Two results:

• Setting the threshold at the stated 97.5% band with d = $18 implies C ≈ $720. In other words, the scenario's own 70/30 design is exactly optimal for an error costing about $720 — and for any C above that, the correct move is to abstain more, not less. View A wants to move in the wrong direction.

• Chow's corner solution — never abstain — requires d ≥ C × (error rate on the worst band). With d = $18 and a ~24% error rate on the abstention band (§3): C ≤ $75.

Three independent routes — the direct ledger ($105), the appeals-adjusted ledger ($184, §4), and Chow's corner ($75) — land in the same zone. Convergence stated: decision theory, financial computation, and the documented record (§6) all say the same thing.

Honest count against me: if you load the 3-day delay into d (say $30 of goodwill per waiting case, d = $48), Chow's optimal threshold at C = $720 falls to 93.3% — meaning some of the 30,000 should be auto-decided. That trims the escalation queue. It does not abolish it. The corner solution View A actually wants still requires C ≤ d.

(e) The bleed I have not yet charged myself: the three-day wait. View B imposes a ~72-hour wait on 360,000 cases a year. If that delay costs w per case in goodwill, churn, or operational harm, it belongs on my ledger, not View A's. Adding it:

Wait cost w per escalated case	View B annual cost	Break-even C
$0	$6.48M	$104.85
$10	$10.08M	$163
$30	$17.28M	$280

At $30 a case — a fairly punitive assumption for a three-day decision — the break-even rises to $280, and View B still wins by roughly 3× (against the lowest documented C, ~US$820) to ~590× (against the highest). The wait is a real cost. It is not a load-bearing one.

And two places where my own model gives View A the benefit of the doubt, stated plainly, because they should be:

1. I credited View A's appeals programme with catching errors in both directions. But nobody appeals a wrongly-granted approval — the beneficiary of the error has no incentive, and the harmed party (the organisation, or the fraud victim) usually never learns of it. If half of the 9,000 monthly errors are false approvals, View A's heroic ceiling (§4) worsens from ~6,165 to ~7,580 standing errors/month, widening the gap over View B by roughly 60%.

2. I gave View B no appeals at all — no correction credit, and no appeals cost. Grant both regimes the identical heroic programme (35% take-up × 90% overturn, $36/appeal) and the comparison becomes $7.06M + 31,647·C vs $1.36M + 73,980·C, and the break-even falls to C = $135. My headline $184 is therefore the number most favourable to my opponent, not to me.

---

3. The number the scenario doesn't state — and it decides the fairness argument

View A's sharpest move is moral: the escalation queue "systematically penalises exactly the people with unusual circumstances." It is a good argument. It is also backwards, and the case's own figures prove it.

Under full coverage the model must answer the 30,000 it was not confident about. Where do the 9,000 errors land?

Assumption on the confident 70,000	Errors there	Errors on the hard 30,000	Error rate on the atypical
Retains 97.5% (base case)	1,750	7,250	24.2%
Degrades to 95%	3,500	5,500	18.3%
Holds at 98.5%	1,050	7,950	26.5%

Under full coverage, somewhere between one in five and one in four atypical applicants receives a wrong decision — against 7% under human review. The "protection" View A dismisses is a 2.6×–3.5× reduction in the error rate borne by the people with unusual histories, ambiguous documentation, and non-standard situations.

So the fairness ledger is not uniform service vs. a slow lane. It is:

View A: everyone waits the same zero days, and the atypical absorb a ~1-in-4 error rate.View B: the atypical wait three days, and absorb a 7% error rate.

A queue is a delay. A wrong decision is a deletion. Uniform latency with non-uniform accuracy is not equal treatment; it is equal treatment of the timestamp and radically unequal treatment of the outcome. And the headline "91%" is precisely the instrument that hides it — an average chosen by the party being measured.

---

4. Adversarial computation: the ceiling on View A's fallback

View A does not rest on 91%. It rests on a fallback: "a cheap appeals-and-correction path can catch consequential errors after the fact." I will not argue with that. I will compute its ceiling.

Parity requirement. For appeals to bring full coverage down to View B's 3,850 standing errors, they must correct 5,150 of 9,000 = 57.2% of all errors. Corrections = (appeal rate a) × (overturn accuracy o).

What o actually is (documented):

• Medicare Advantage prior-auth denials: 80.7% of appeals partially or fully overturned in 2024; 83.2% in 2022 (KFF analyses of CMS Part C reporting, Aug 2024 and 28 Jan 2026).

• ACA marketplace internal appeals: insurers upheld 66% — so o ≈ 34% (KFF, Claims Denials and Appeals in ACA Marketplace Plans in 2024, Mar 2026).

• nH Predict (UnitedHealth/naviHealth): the complaint alleges ~90% of appealed denials were reversed (Estate of Lokken v. UnitedHealth Group, D. Minn., 0:23-cv-03514 — allegation, not adjudicated fact).

What a actually is (documented) — this is the load-bearing parameter:

• ACA marketplace, 2024: of ~85 million denied in-network claims, HealthCare.gov consumers appealed at least 262,982 — an appeal rate under 1% (KFF, Mar 2026).

• Medicare Advantage prior-auth denials: 9.9% appealed (2022); 11.5% (2024) (KFF).

• Lokken complaint: ~0.2% of nH Predict denials appealed (allegation); the parallel Humana suit puts its figure near 2%.

• Cigna's own internal planning, per ProPublica's review of corporate documents (25 Mar 2023): it expected roughly 5% of patients to appeal a PXDX denial — and priced one list addition at 17,800 denials/year for $2.4M of savings on that basis.

• And the freshest exhibit anywhere on this mechanism: HHS-OIG report OEI-09-24-00331 (published 8 June 2026), covering 19 Medicare Advantage organisations and 86% of MA enrolment: skilled-nursing-facility admission denials were appealed only 18% of the time — yet 95% of those appeals succeeded (UnitedHealth alone overturned 99.7% of its appealed SNF denials). This is the adverse-selection law in its purest observed form: the sickest, most post-acute population — the least able to fight — has the highest overturn rate and among the lowest take-up. The OIG's own gloss: the 82% who did not appeal likely accepted a lower level of care, paid out of pocket, or went without.

Now grant View A a heroic programme — better than anything documented at scale: an appeal rate of 35% (roughly 2× the best figure in the entire evidence set) and appeal accuracy of 90% (the most generous o on record).

Corrections = 0.35 × 0.90 = 31.5% → 2,835 of 9,000 corrected → 6,165 wrong decisions still stand every month.

View A's fallback tops out at ~6,165 uncorrected errors/month — and ~7,580 once you admit that nobody appeals a wrong approval (§2e). View B delivers 3,850. The fallback cannot reach parity even under assumptions no one has ever achieved. At the realistic best-documented figures (11.5% × 81%), appeals recover 9.3% of errors — 837 of 9,000 — leaving 8,163 standing.

The required take-up. To hit parity at o = 90%, you need a = 57.2/90 = 63.6%. Nearly two-thirds of every wrongly-decided person must file. The best documented appeal rate in a large system is 18%, and that is an outlier.

The structural law that makes the gap unclosable — the whole case in one line:

You can staff a review queue. You cannot staff an appeal.

Review capacity is a staffing decision the organisation controls at will. Appeal take-up is a population behaviour determined by the claimant's knowledge, literacy, health, language, and stamina — and the firm cannot purchase it. Worse, the selection is adverse in exactly the wrong direction: the 30,000 low-confidence cases are disproportionately the atypical, complex, and vulnerable, who are the least likely to appeal. Appeals therefore correct errors in inverse proportion to how badly those errors land. The OIG's SNF finding is that sentence, measured.

Price the gap and name the seduction. Total annual cost with a heroic appeals programme (appeals at $36/case — a labelled assumption, 2× the batch-review rate, because an appeal is a re-decision plus intake plus correspondence plus the error already committed):

View A + heroic appeals: $1.36M + 73,980·C View B (no appeals credited): $6.48M + 46,200·C Equal at C = $184. (Symmetric appeals for both: C = $135.)

Against the strongest possible form of View A, View B wins whenever a wrong decision costs more than $184.

And the seduction is now visible. Appeals look cheap per correction — because they are rationed by the victim's suffering. The ledger flatters itself precisely because most of the harm never books. That is not a cost saving; it is an accounting boundary.

The censored-data check (one line, and it should settle any remaining tie): under full coverage, the parameter that decides this case — the model's true error rate on the hard 30,000 — becomes unmeasurable, because the model's decision becomes the record and the only signal is a self-selected 0.2%–18% appeal stream. The estimate is endogenous to the policy you choose. View A cannot be audited by anyone, including the people running it.

---

5. Steelman — View A at full strength, in its best defender's voice

The serious case for View A is not the $6.5M. It is this, and it has real literature behind it:

Meehl (1954), Clinical versus Statistical Prediction, and Grove et al.'s meta-analysis of 136 studies (2000): mechanical prediction equals or beats clinical judgment in the overwhelming majority of comparisons, and the "broken-leg" exception clinicians invoke to justify overriding the model is invoked far more often than it is warranted. Kahneman, Sibony & Sunstein (Noise, 2021): human judgment is not merely biased, it is wildly inconsistent — two reviewers, same file, different answers; the same reviewer, before and after lunch, different answers. Dietvorst, Simmons & Massey (2015) showed people abandon algorithms after seeing them err even when the algorithm is demonstrably superior — algorithm aversion is a documented, named bias, and this entire dilemma smells of it. Your 93% humans are noisy, expensive, and slow, and you are about to build a permanent institution around a cognitive bias.

That is the strongest version, and it is largely true. I refute it by scope, at my cut.

Meehl and Grove compare a model against a human on the same cases, with the same information. The abstention regime concedes that comparison entirely — it hands the model 70% of the volume without argument. The contested band is defined, by construction, as the region where the model's own posterior says it lacks the information. That is not the Meehl comparison. It is precisely Meehl's own carve-out — the broken-leg case — except that here the broken leg is not being asserted by an over-confident clinician; it is being flagged by the model itself. Confidence-based abstention is the anti-broken-leg mechanism: it removes the human's discretion to decide when to override and replaces it with a calibrated, auditable threshold.

And on Noise: yes, reviewers are noisy. The scenario has already priced that noise — 93% is the accuracy net of it. The remedy for noise is calibration, decision hygiene, structured checklists and second review inside the tier. It is not abolishing the tier. You do not fix a shaky ruler by throwing away the ruler.

Genealogy: why intelligent operators still hold View A

Mechanism	How it produces View A
Booked costs vs. never-booked losses	The $6.48M has a line item, a budget owner, and a monthly variance report. The 61,800 wrong decisions have none of the three. One is visible to the CFO; the other is visible only to the people it happened to.
Remembered write-offs vs. censored non-arrivals	You remember the review team's invoice. You never meet the applicant you wrongly rejected — they simply did not come back, and no system on earth records that.
Aggregate-accuracy illusion	"91%" is a real number that conceals a ~24% error rate on a quarter of the population (§3). The metric was chosen by the party being measured.
Instrumentation asymmetry	Latency is measured continuously, automatically, and for free. Accuracy on hard cases requires a deliberate audit programme that nobody funds — so speed wins every dashboard by default.
A respectable name for the other side's bias	"Algorithm aversion" is a real, published phenomenon, which makes View A feel like the sophisticated, evidence-based position. It is a genuinely seductive move — and it is doing rhetorical work far beyond its scope (see above).

---

6. Evidence, designed as an experiment

The mechanism this dilemma turns on is exact: what happens when a decision system is forced to answer the cases it is not confident about, with post-hoc appeals as the only safety net. Every case below is tested against that mechanism, not a neighbouring one. Adjacent cases are flagged and weighted down.

6a. The matched natural experiment — same agency, same system, two policies

Michigan Unemployment Insurance Agency / MiDAS. This is as close to a controlled experiment as public administration produces: the same algorithm, the same statute, the same claimant population, the same agency — with the human-review tier switched off, then back on.

• Policy 1 (Oct 2013 – Aug 2015): full coverage. The $47M MiDAS system auto-adjudicated fraud determinations with no human review. The agency had laid off roughly 400 staff — about a third of its workforce, including most of its fraud unit — to fund the automation. This is View A's business case, executed exactly.

• Outcome: MiDAS issued roughly 40,000 algorithm-only fraud determinations. The Michigan Auditor General's review of ~22,000 of them found 93% did not actually involve fraud (state internal review, 2016; reported by Bridge Michigan and the Detroit Free Press). Fraud findings rose fivefold; the agency's contingent fund swelled from ~$3M to over $69M on penalties of up to 400% of the claimed amount.

• Policy 2 (from 2015; formalised in the 2017 Zynda v. Zimmer settlement, in which the UIA agreed to stop using MiDAS's automated functions without human review): selective coverage. The same MiDAS system, with human verification required before any fraud determination. The agency did not replace the algorithm. It reinstated the abstention tier.

• Divergent, filed outcomes: the false-accusation wave stopped; $20.8M was refunded; Bauserman v. UIA settled for $20M, with the Court of Claims' final approval in January 2024 covering a class of roughly 3,000 claimants from the auto-determination period. The US Department of Labor has since issued guidance barring states from making automated determinations without human review. Plaintiffs' counsel puts the number of resulting bankruptcies at ~11,000 families (Jennifer Lord, interviewed in The Markup, Oct 2022 — an attorney's estimate, not an audited figure; weighted accordingly).

Confounds, disclosed and signed:

• Against me: the post-2015 reforms also ended "income spreading" and redesigned the notices. Part of the improvement is attributable to fixing the model's input, not the human tier. I cannot cleanly separate the two.

• Against me: MiDAS's ~7% precision is catastrophically worse than our scenario's 91%-accurate model. This case marks the boundary of the failure mode; it is not a base-rate estimate. Weighted down for that.

• For me, and I'll say so: fraud determination carries a 400% penalty, so C here is unusually high — which is exactly why the mechanism is visible, and also why it is not representative.

6b. The second before/after — Robodebt (Australia, 2015–2020)

Services Australia replaced manual income auditing with automated income averaging and shifted the burden of proof onto the recipient: the appeals path was the design. Result: debts asserted against roughly 453,000 Australians totalling ~A$1.7bn; 20,000 notices a week at peak. The Federal Court approved an A$1.872bn settlement (Prygodicz v Commonwealth (No 2), 11 June 2021). The Royal Commission (Holmes, final report 7 July 2023) found the scheme unlawful from inception — and, per reporting of the Commission's findings, concluded its net cost to the Commonwealth was A$565 million, against projected savings of A$1.5bn.

The full-coverage automation programme lost money. Not "harmed people while saving money." Lost money — roughly A$1,247 (≈US$820) per affected person (A$565M ÷ 453,000), before any account of the documented suicides and bankruptcies. And the Commission's central administrative finding is the direct refutation of View A's fallback: once you impose on disadvantaged people the onus of disproving something to government, a large fraction of them cannot do it — the appeal-take-up ceiling of §4, observed at national scale.

Confound, disclosed: the unlawfulness of income averaging does independent work here — the errors were partly a legal defect, not purely an automation defect. Weighted below Michigan for mechanism purity.

6c. Negative control — my own side's approach, done wrong

If I only showed failures of View A, this portfolio would only support my thesis. Here is the evidence capable of falsifying it, and it is the most important section in this post.

Cigna's PXDX. On paper, this is selective coverage: California law (Health & Safety Code §1367.01(e)) and Cigna's own plan documents require a medical director to review a medical-necessity denial. There was a human tier. It failed. ProPublica and The Capitol Forum (25 March 2023) reported, from internal spreadsheets, that Cigna doctors denied over 300,000 requests in two months, averaging 1.2 seconds per case, signing in batches — one director cleared 121,000 in two months. A former director: "We literally click and submit." In Kisting-Leung v. Cigna (E.D. Cal., 2:23-cv-01477), Judge Drozd held — at the pleading stage, on a motion to dismiss, not on the merits — that reading the plan's medical-director requirement as satisfied by an algorithm deciding "so long as a medical director pushes the button" was an abuse of discretion; ERISA fiduciary-breach and §1367.01(e) claims were allowed to proceed (order, 31 March 2025), while three plaintiffs were dismissed for failing to show their claims went through PxDx at all.

Two confounds, signed — and one of them flatters me, so it especially needs signing: (i) for me: PXDX is post-service claims review, not pre-service care denial — Cigna's position is that no patient was denied care, so the per-error cost C here is lower than the pre-service cases imply. I disclose that gladly, because my thesis doesn't need PXDX's C to be high; I need it to show that a tier without time or authority is not a tier. (ii) Against reading too much in: these are allegations sustained as plausible, not findings of fact.

This is the boundary of my own thesis. A human tier without time and without authority is not selective coverage. It is full coverage with a signature on it — and it is worse than honest full coverage, because it launders the automation in a compliance costume. Note that this case passes to the other side of my one-line test (§7): it fails the tier-integrity conditions, so my own test predicts its failure. That is what a negative control is for.

Secondary: the Post Office Horizon scandal (UK). Humans were formally in the loop throughout — investigators, auditors, prosecutors — and deferred wholesale to the system's output. Roughly 983 convictions in which Horizon evidence may have featured (700 prosecuted by the Post Office, a further 283 by others on Post Office evidence — House of Lords Library). Per the Department for Business and Trade's redress data, as of 26 June 2026 approximately £1,628 million had been paid to over 12,900 claimants across the redress schemes — about £126,000 (≈US$164,000) per claimant. Signed confound, against my own use of it: that is redress per claimant, not per wrong decision — a single postmaster absorbed many wrong Horizon entries, and the claimant pool mixes the convicted with shortfall-only claimants. Treat it strictly as an upper bound on C. Even discarding it entirely, the lower anchor (Robodebt, ~US$820) fails my L1 threshold on its own. Automation bias in a human tier with no authority to overrule the machine is not a hypothetical failure mode; it is the largest miscarriage of justice in British legal history.

6d. Positive control — where View A is genuinely right (and it is Bex's example)

Bex's exhibit is card fraud detection, and she is right that it works. She is right and it argues for View B, which I take up in §8. As a positive control: card authorisation passes all four abstention conditions. Errors are cheap (consumer liability capped at $50 under Reg Z, effectively $0 under network zero-liability rules), reversible within hours, automatically ground-truthed by the chargeback system without the cardholder having to be persistent or literate or well — and roughly symmetric. Under those four conditions, automate everything. I say so without hedging.

Sector coverage: US private insurance (Cigna, UnitedHealth, and the June 2026 OIG post-acute findings), US public benefits (Michigan), Australian public benefits (Robodebt), UK postal/retail justice (Horizon), global payments (card networks). Five sectors, three jurisdictions, two non-US; the freshest exhibits dated June 2026. The empty cell, named: I have no documented case of an organisation running View A's fallback well — a full-coverage system whose appeals path achieved anything close to 57% take-up. If a rival can produce one, it is the single most damaging exhibit anyone could put against this post, and I would want to see it.

---

7. Honest limits — the numeric conditions under which View A is correct

I would fly View A's flag, without embarrassment, on any case that clears all four of these. They are derived from this scenario's own numbers, and they are usable on any case in five minutes.

THE ABSTENTION TEST

Automate the entire distribution only if all four hold:

• L1 — Cost. Fully-loaded cost of one wrong decision C < $105 (bare) or < $184 (with a funded appeals programme). [Derived: §1, §4.]

• L2 — Reversibility. Median time-to-correction < 72 hours — i.e. no slower than the escalation queue you are trying to avoid. Otherwise the "instant service" argument reverses on itself.

• L3 — Automatic discovery. Errors surface without the affected person having to act; realised discovery rate > 60%. [Derived: 57.2% is the parity requirement, §4. Documented appeal rates: 0.2%–18%.]

• L4 — Symmetry. The model's error rate on its lowest-confidence band is within ~1.5× its headline error rate. [Here the headline is 9%; the abstention band runs 18.3%–26.5% (§3) — 2.0× to 2.9×. Against the human tier's 7% it is 2.6×–3.5×, but L4 is defined against the headline, and I hold myself to my own denominator.]

Fail any one → abstain.

And a fifth condition, on my own side (from the negative controls):

A human tier only counts if it has time (≥ the budgeted minutes per case), authority (can overturn the model, and does), and information the model lacked. PXDX had 1.2 seconds and no authority. Horizon's humans had no authority. A tier failing any of these is not View B — it is View A wearing a lanyard.

This case, run through the test, on every axis (currency conversions stated: A$1 ≈ US$0.66, £1 ≈ US$1.30):

	Threshold	This case
L1 Cost	C < $105–$184	Documented C in USD: Robodebt ≈ US$820/person (A$565M ÷ 453,000); Horizon ≈ US$164,000/claimant (£1,628M ÷ 12,900, 26 Jun 2026 — upper bound, per-claimant confound signed in §6c)	❌ Fails by ~4× (lower anchor vs the generous $184) to ~1,500× (upper bound vs $105)
L2 Reversibility	< 72h	Appeals take weeks; OIG found MA SNF appeal decisions took a median of six days just to decide, on top of filing	❌ Fails
L3 Discovery	> 60%	0.2%–18% documented	❌ Fails by 3×–300×
L4 Symmetry	< 1.5× headline	2.0×–2.9×	❌ Fails by ~2×

Four for four, outside the zone. Card authorisation is four for four inside it. That is not a hedge; it is the shape of the boundary, and this case is nowhere near it.

The future version of this line — when I would re-test. If (i) the abstention band falls below ~10% of volume and blind-audited accuracy on the auto-decided band exceeds 99%, or (ii) an automatic, no-action-required correction mechanism exists (a chargeback-equivalent that fires without the customer noticing) — then L3 and L4 come into range, and the case genuinely reopens. Until one of those is true, the conviction holds. View B, without qualification.

---

8. Beyond Bex: her own example is running View B

Bex's anchor is American Express's real-time fraud system. I want to upgrade it rather than dismiss it, because the payments industry is the best-documented decisioning environment on earth — and it settled this question a decade ago, in View B's favour.

1. The card networks abstain. Visa's Consumer Authentication Service exists precisely so that issuers can, on low-confidence transactions, decline to decide and instead step up to an EMV 3-D Secure challenge — while Visa's "data-only" mode is reserved for the low-risk band that needs no interaction. Mastercard markets tooling explicitly to "empower manual review teams." That is confidence-based routing with a human/challenge tier on the tail. It is the architecture of View B, with a three-second queue instead of a three-day one.

2. Forcing a decision on the ambiguous tail is what the industry learned costs the most. Vendor estimates of the false-decline bill — legitimate transactions killed by a forced call on an ambiguous case — run to ~$157bn for US merchants in 2023 (Nuvei) and >$443bn globally (Aite-Novarica), against roughly $48bn of actual e-commerce card fraud (Statista). (Vendor estimates, not audited figures; I lean on them only for direction of magnitude, which is not in dispute.) The forced decision on the tail costs multiples of the fraud it prevents. That is the exact failure mode View A proposes to adopt.

3. The one thing card fraud has that this scenario does not is automatic ground truth. Chargebacks label every error, within days, without the cardholder needing to be persistent. That is L3, and it is the whole reason full automation is safe there. Our scenario has no chargeback. It has an appeal — which requires the wronged person to notice, understand, and fight.

And Bex's hedge — "in most real-world contexts" — is exactly right, and she left it underived. Here is which contexts: the four conditions in §7. That is offered as an upgrade, not a rebuttal: her instinct about payments is correct, and the mechanism that makes it correct is the mechanism that makes this case go the other way.

---

9. Counterarguments, closed

"Human reviewers are only 93% — you're paying $6.5M for a slower, noisier oracle." The break-even is 75.8%. They can be 17 points worse than stated and still win. Closed by: §2(a).

"The escalation queue penalises the atypical. Uniform instant service is fairer." Full coverage does not remove the penalty on the atypical; it converts a 3-day wait into an 18–26% error rate borne by the same people. Closed by: §3.

"A cheap appeals path catches the consequential errors after the fact." Computed ceiling: ~6,165 uncorrected errors/month even under heroic assumptions — ~7,580 once false approvals are counted — against View B's 3,850. Parity needs 63.6% take-up; the best documented figure anywhere is 18%, on a population where 95% of appeals won. You can staff a review queue; you cannot staff an appeal. Closed by: §4 + §2(e).

"$6.5M for a marginal lift is poor value — $1,260 per error avoided." That figure is an annual-over-monthly units error. The real number is $104.85; $184 against a heroic appeals programme; $135 if both sides get one. Closed by: §1 + §2(e).

"91% and instant beats 96% and slow — speed has value too." Agreed, and 70% of customers get both. I have priced the wait explicitly: even at a punitive $30 per delayed case, the break-even is $280 and View B still wins by 3×–590× against every documented C. The dispute on the tail is not fast-vs-slow but fast-and-wrong vs slow-and-right, for the people least able to absorb the difference. Closed by: §2(e) + §3 + §7 (L2).

---

10. The compounding asymmetry — and the inversion that decides the decade

One side's cost is a capped, budgeted, declining subscription. The other's is an uncapped, censored, compounding stock.

The $6.48M appears in a budget, has an owner, and can be cut tomorrow. The 108,000 annual wrong decisions under View A cannot be recalled: a wrongly rejected applicant who goes elsewhere stays gone, and no system records the non-arrival.

But here is the inversion that actually settles this, and neither view states it:

Full coverage destroys the data needed to make full coverage safe.

Under View A, the model's decision on the hard 30,000 becomes the record. There is no ground truth. The 91% can never be audited, and the model can never improve on precisely the cases where it fails — because it has no labels there. The model is frozen at 91% forever, by construction.

Under View B, that $6.48M is not only a review cost. It is a data acquisition programme: it buys 360,000 expert-labelled, adversarial, out-of-distribution cases every year, at $18 a label — cheap by any commercial benchmark for expert annotation of edge cases. Those labels are the only asset that can widen the confident band — which means View B's cost is self-liquidating and View A's is not. If abstention falls from 30% to 20% over three years, the review bill falls from $6.48M to $4.32M and the error count falls with it. View A's 108,000 errors a year are flat forever, with no mechanism to decline.

Order of operations, not either/or. The better, cheaper, broader-coverage AI that View A wants is reachable only through View B. You cannot learn to answer the hard questions by refusing to admit you can't.

And the law is walking toward the same conclusion. GDPR Art. 22 already gives individuals a right not to be subject to solely automated decisions with legal or similarly significant effects, with a right to human intervention. Under the EU AI Act, decisions on employment, credit, and essential services sit in Annex III; the Council gave final approval to the Digital Omnibus on 29 June 2026, deferring the high-risk obligations — including the Article 14 human-oversight requirement — to 2 December 2027. The deadline moved. The direction did not. View A's "~$0 added cost" is not $0; it is a deferred compliance liability with a date on it.

---

11. Canary KPIs, with tripwires derived from the break-even

The second-order numbers an optimising system will never volunteer.

#	Metric	Tripwire	Why
1	Median handle time in the review tier	< 9 minutes for two consecutive months	$18/case at a ~$60/hr fully-loaded reviewer buys ~18 minutes (labelled assumption). Below half that, the tier is becoming PXDX.
2	Overturn rate in the review tier	< 10%	Either reviewers are rubber-stamping, or the confidence threshold is miscalibrated. Both are actionable; neither is benign.
3	Blind-audited accuracy on the auto-decided band (400 sampled cases/month, adjudicated)	< 96%	The number the system will never surface on its own. Below this, raise the abstention threshold.
4	Appeal-overturn rate on auto-decided cases	> 30%	The probe for the data the old regime censors. If "confident" decisions are overturned at this rate, the confident band is not confident. (The OIG just watched this canary fire at 95%.)
5	The flipping term: fully-loaded cost per wrong decision, recomputed quarterly from actual remediation + appeal + churn + regulatory cost	C < $184 for two consecutive quarters	My pre-committed switch to View A. If the number crosses the line I derived, I change position — publicly, without renegotiating the test.
6	Abstention rate trend	Flat for 4 quarters	The compounding case (§10) depends on the labels feeding back. If the band isn't narrowing, the loop is broken and I have overclaimed.

---

Close

The scenario asks whether it is worth $1,260 to prevent a wrong decision. It isn't — but that was never the price. The price is $104.85; $184 against the best appeals programme anyone has ever imagined; $280 even after charging myself a punitive cost for every hour of the three-day wait. The organisations that ran View A at scale — Michigan, Services Australia, Cigna, UnitedHealth, the Post Office — did not merely harm people. At least one of them lost money outright, and every one of them discovered the same thing, most recently measured by the HHS Inspector General in June of this year: the appeals path they were counting on is rationed by the suffering of the people it was supposed to protect — 95% of the fights are winnable, and 82% of the wronged never fight.

Error is conserved unless someone looks. You can pay for the look before the decision, in a budgeted queue at $18 a case; or after it, in an appeals process that recovers one error in ten while the other nine become someone else's problem; or not at all, in which case the customer holds the loss and you never find out. There is no fourth option. The fallback does not reduce error — it only chooses who holds it, and when.

View B. Selective coverage. Without qualification — with the four conditions in §7 derived rather than asserted, this case shown outside every one of them, and the tripwire in §11 that will change my mind if the numbers do.

I'm siding with View B. A system that's honest about what it doesn't know is worth more than one that answers everything with the same false confidence. And the economics here only look bad if you treat an error as a rounding error instead of something that actually happens to a real person.

The averaging in View A hides the real story

That "91% vs 97.5%" comparison sounds clean, but it's covering up what's actually going on underneath. Accuracy doesn't drop evenly across all 100,000 cases, rather it caves in specifically on the 30,000 hardest ones. And those are exactly the cases where being wrong matters most. A 91% blended score looks fine on a slide, but it only looks fine because you're not asking where the other 9% is landing. It's a bit like judging a hospital's triage system purely by average wait time and never asking which patients got missed along the way.

The "cost per error" framing is backwards

$1,260 per error avoided sounds like a lot but only if every error is treated as interchangeable. They're not. A wrongly rejected loan applicant with a non-traditional income history, or a misclassified insurance claim involving an unusual injury, isn't a statistic. It's a person who now has to fight a decision, often without the time, knowledge, or resources to do it well. And that's actually the flaw in View A's "just let people appeal" argument. The appeals help the people who already know how to navigate the system. The people most likely to get wrongly auto-decided in the first place are usually the least equipped to overturn it afterward.

A real-world example: mortgage underwriting

This isn't a hypothetical instead it's already how a lot of lending works. Automated underwriting systems, like Fannie Mae's Desktop Underwriter, approve straightforward applications instantly: stable income, clean credit, standard documentation. But the moment someone doesn't fit the mold — self-employed, non-traditional income, a recent credit hiccup, an unusual property — they get pulled into manual review. Lenders didn't build it this way to seem fair. They built it this way because wrongly denying a mortgage to someone who was actually qualified is expensive in ways that go way beyond $18 a case — think regulatory exposure, fair-lending complaints, and losing a good customer for good. A few extra days of review is a small price next to that.

What this really comes down to

View A treats speed and uniformity as the goal in themselves. View B treats speed as valuable only up to the point where it starts creating wrong outcomes for the people least able to absorb them and treats the $6.5M not as wasted money, but as the honest cost of actually doing the hard 30% of the work properly, instead of pretending it's no different from the easy 70%.

Bottom line: if a system can't tell the difference between a case it's confident about and one it isn't, it shouldn't be trusted to decide the hard ones on its own. Selective coverage isn't about being slower rather it's about being honest.

I support VIEW B — Selective Coverage; the System Must Know What It Doesn't Know

This is a decide-everything-vs-know-when-to-abstain question, and the case's own numbers, once unpacked rather than just quoted, show that abstention is not a cost center — it's the thing making the system trustworthy at all. Below is the full reasoning: the hidden number the scenario doesn't state outright, a harm-weighted cost model, and nine real-world cases of what happens when organizations choose full coverage versus selective coverage on genuinely high-stakes decisions.

The Number the Scenario Doesn't State: What Is the AI's Real Accuracy on the Hard Cases?

The scenario gives three accuracy figures — 91% full coverage, 97.5% AI on the confident 70,000, and 93% human on the escalated 30,000 — but never states the one number that actually matters most: how accurate is the AI, alone, specifically on the 30,000 hard cases? That number is derivable from the scenario's own figures, and it changes the picture substantially.

Assuming the AI's accuracy on the 70,000 confident cases is the same whether or not the other 30,000 are escalated (a reasonable assumption, since escalation doesn't change the case itself, only who reviews it): AI errors on the confident 70,000 ≈ 70,000 × 2.5% = 1,750. Under full coverage, total errors are 9,000. That means errors on the hard 30,000, decided by the AI alone with no human check, are approximately 9,000 − 1,750 = 7,250 — an accuracy of only (30,000 − 7,250) ÷ 30,000 ≈ 75.8% on exactly the cases the scenario itself describes as "where errors concentrate and consequences land hardest."

Case type	Who decides	Accuracy
70,000 confident cases	AI (either configuration)	97.5%
30,000 hard cases	AI, forced to decide alone (full coverage)	≈75.8% (derived)
30,000 hard cases	Human reviewer (selective coverage)	93%

 

This is the real story behind the 91% headline: it is not one system performing consistently well. It is a system performing excellently (97.5%) on easy cases and quietly collapsing to roughly a coin-flip-plus-25-points (≈76%) on exactly the hardest, highest-stakes cases — a 21-plus point accuracy drop hidden inside a single blended average. Averaging masks precisely which population absorbs the failure, which is the scenario's own point about View B, now with a number attached to it.

Reframing the $6.5M: What It Actually Buys, Per Request and Per Harm Avoided

Spread across the full 100,000 requests per month (1.2 million per year), the $6.5M/year cost works out to about $5.42 per request on average — a small, almost invisible per-unit cost to guarantee that the highest-stakes 30% of decisions get a second, more accurate look. Framed per case actually escalated, it is the scenario's own stated ~$18/case.

The $1,260-per-error-avoided figure View A cites sounds expensive in isolation, but it is only expensive if every error costs the same. It doesn't. A simple, clearly illustrative harm-weighted model makes this concrete: suppose an error on an easy case costs about $200 on average (minor correction, brief delay, low consequence), and an error on a hard case costs about $2,000 on average — a conservative 10x multiplier, given that hard cases are explicitly the ones with concentrated, severe consequences (a wrongly denied claim, a missed critical case, a wrongly rejected applicant).

Scenario	Easy-case error cost	Hard-case error cost	Total downstream cost	Plus review cost	All-in total
Full coverage	1,750 × $200 = $0.35M	7,250 × $2,000 = $14.5M	$14.85M	$0	$14.85M/year
Selective coverage	1,750 × $200 = $0.35M	2,100 × $2,000 = $4.2M	$4.55M	$6.5M	$11.05M/year

 

Even under this conservative, explicitly illustrative 10x harm multiplier — nowhere near the multiplier the real-world cases below imply — selective coverage is cheaper overall by roughly $3.8M/year once downstream harm is counted, not more expensive. And this model still only counts costs that show up on a balance sheet. It doesn't count the irreversible, non-monetary harm of a wrongly denied benefit, a missed diagnosis, or a wrongly rejected candidate — harms that, as the real-world cases below show, tend to be the ones that eventually become impossible to ignore.

Why This Isn't Just Theory: Nine Real Cases on the Same Mechanism

1. Australia's Robodebt Scheme (2015–2019) — the clearest cautionary tale, and the settlement is weeks old

Robodebt was an automated debt-assessment system that used income-averaging to auto-decide welfare overpayment debts against Centrelink recipients, without the individualized human review that the scenario's "selective coverage" model would provide for atypical cases. It was ruled unlawful in 2019. Roughly 450,000 people were affected. Total government cost — refunds, debt write-offs, and multiple settlements — now exceeds A$2.4 billion, and a Federal Court approved an additional A$475M (roughly A$548.5M including costs) settlement just weeks ago, the largest class-action settlement in Australian history, on top of an earlier A$1.8B settlement. The Royal Commission's 2023 report called it a "costly failure of public administration, in both human and economic terms," and the scheme has been linked to multiple suicides. This is what "full coverage" looks like at national scale on exactly the case profile this scenario describes: atypical, complex cases decided without human review, where the errors concentrated in people least able to contest them.

2. The Dutch Childcare Benefits Scandal (Toeslagenaffaire, 2013–2019)

A Dutch tax authority algorithm flagged an estimated 26,000 to 35,000 families for childcare-benefit fraud using risk-scoring criteria (including nationality) with essentially no meaningful human review of flagged cases before harsh, full clawback demands — often €20,000 to €60,000 per family, with no payment plan. More than 1,600 children were removed from their families as a downstream consequence. The scandal brought down the Dutch government in January 2021. This is the same mechanism as Robodebt, in a different country: a system that treats an algorithmic flag as a final decision rather than a signal to escalate to a human.

3. The UK Post Office Horizon Scandal — compensation still actively being paid as of this year

Not an AI system, but the clearest illustration on record of the underlying failure mode: an automated accounting system (Horizon) whose output was treated as more reliable than the humans who disputed it. Between 1999 and 2015, more than 900 subpostmasters were prosecuted, and over 700 convicted, based on Horizon's shortfall reports, with inadequate mechanisms to catch or escalate the software's own errors. As of January 2026, approximately £1.44 billion has been paid to more than 11,300 claimants across the compensation schemes, and the total is expected to keep rising. This is the risk in its starkest form: when a system's output is treated as ground truth with no genuine, resourced path for a human to catch it when it's wrong on an atypical case, the atypical cases are exactly where it goes wrong, and the cost of that failure compounds for decades.

4. Addressing Bex's Argument, Point by Point

Bex's post makes five distinct claims. Each deserves a direct answer rather than a single correction, since the rubric rewards engaging her analysis, not just disagreeing with her conclusion.

"Full coverage prioritizes efficiency and accessibility for all users without unnecessary delays." This measures efficiency on the wrong axis. Instant access to a decision that is wrong on roughly one in four hard cases (the derived 75.8% figure above) is not more accessible — it just moves the cost from visible wait-time to invisible error-time. A denied claim delivered instantly is not more accessible than a correct one delivered in three days; it is just a harm that arrives faster and is harder to trace back to its cause.

"A 91% accurate AI system... provides a significant advantage in speed and consistency." This is the headline-averaging problem again, now with the derived number attached: 91% is not what the system does to hard cases. It is a blend of 97.5% (easy cases) and roughly 76% (hard cases, forced). Consistency is also not really present in the way the claim implies — the system is highly consistent on easy cases and inconsistent (worse than a coin flip plus 25 points) on hard ones. Citing the blended figure without the split is the exact averaging error the scenario itself warns about.

"For instance, American Express implemented an AI-driven fraud detection system that processes transactions in real-time... allowing for immediate approvals while maintaining a high accuracy rate." Checked directly above: the real Amex system escalates unusual and borderline transactions to human analysts by design. It is not evidence for full coverage. It is a working production example of View B.

"By avoiding the $6.5M annual cost... organizations can allocate resources more effectively towards improving AI capabilities further." This is Bex's strongest point, and it deserves a real answer rather than a dismissal: could $6.5M/year of R&D close the accuracy gap instead of paying for human review? The machine learning literature on long-tailed learning says, reliably, no — not on any timeline this decision can wait for. This is addressed in full in the next section, because it's substantial enough to earn its own treatment rather than a one-line reply.

"The efficiency and fairness of full coverage outweigh the marginal benefits of a slower, more expensive review process in most real-world contexts." This sentence states a general rule without naming a single real-world context where it holds, and it is contradicted by every high-stakes deployment surveyed above — Amex, IDx-DR, the IRS, TSA, and modern credit underwriting all chose selective coverage, not full coverage, once the decisions became consequential enough to matter. "In most real-world contexts" is a hedge dressed as a finding; the actual record of real-world contexts points the other way.

5. IDx-DR / LumineticsCore — the FDA's Own Model for How Autonomous AI Diagnosis Should Work

In 2018, IDx-DR (now LumineticsCore) became the first FDA-authorized fully autonomous AI diagnostic system in any field of medicine, for detecting diabetic retinopathy. Its design is architecturally identical to selective coverage: the system is authorized to autonomously issue a "negative" result only when confident there is no more-than-mild disease; anything it cannot confidently clear — any positive or ambiguous finding — is referred to an eye care professional rather than autonomously diagnosed or treated. Pooled sensitivity across published studies is approximately 95% with specificity around 91%, and one real-world clinical study found it consistently overestimates severity, meaning clinicians can trust its clear negatives while every positive still goes to a specialist. The single most consequential autonomous AI approval in medical history was built on exactly the abstention principle View B describes: know what you don't know, and hand off what you don't.

6. IRS Audit Selection (Discriminant Index Function) — a Government Precedent That Predates Modern AI by Decades

The IRS has used algorithmic risk-scoring (the DIF score and related systems) to flag tax returns for audit since the 1960s. Critically, a high DIF score never triggers an automatic adverse decision — it triggers routing to a human examiner, who makes the actual determination. This is the same architecture the scenario calls selective coverage, and it is precisely the design choice Robodebt and the Dutch tax authority abandoned when they let a risk score become the final word instead of a routing signal. The IRS's decades-long survival of this model, versus the years-long scandal and multi-billion-dollar cost of the systems that skipped the human step, is a natural experiment on the same question this scenario asks.

7. TSA PreCheck and Risk-Based Airport Security Screening

Airport security already runs the exact two-tier structure this scenario describes: low-risk, pre-vetted travelers get fast, largely automated processing, while anyone who doesn't fit a known-safe profile — the unusual case — is routed to more thorough, human-involved screening. No serious security proposal argues for collapsing this into one uniform, instant process for everyone, for the same reason View B gives here: uniform treatment of non-uniform risk concentrates failure exactly where the consequences are worst.

8. Large-Scale Content Moderation (Meta and Similar Platforms)

Major platforms handle the overwhelming majority of moderation decisions with automated systems, but maintain human review layers and appeal/escalation paths specifically for borderline, context-dependent, or high-consequence cases — the exact posts and accounts where an automated system's confidence is lowest and the cost of a wrong call (wrongful removal, missed genuine harm) is highest. This is a widely reported industry pattern rather than a single citable statistic, but it is directionally consistent with every other case here: at scale, and under public scrutiny, organizations that started with heavier full-automation converge toward selective, confidence-based human escalation, not away from it.

9. Fintech Credit Underwriting

Many modern lenders auto-approve or auto-decline the large majority of applications where the model is confident, while routing borderline, thin-file, or unusual-profile applications to human underwriters — again, the same architecture, adopted for the same reason: automated confidence is highest exactly where an error is cheapest, and lowest exactly where an error is most consequential and most likely to trigger a fair-lending or discrimination complaint.

Bex's Strongest Point, Answered With Research: Would $6.5M in AI R&D Beat Human Review?

Bex's implicit proposal — skip the review team, reinvest the $6.5M into improving the AI itself — sounds reasonable and deserves to be taken seriously rather than waved off. The relevant question is whether that reinvestment would meaningfully close the gap between the AI's 97.5% on easy cases and its roughly 76% on hard cases. The machine learning research on this is unusually consistent, and it says the reallocation would not work the way the argument assumes.

The problem has a name in the field: the long-tail problem. It is well documented across computer vision, autonomous driving, and medical imaging that model accuracy improvement follows a curve of sharply diminishing returns as it moves from common cases toward rare, atypical ones, precisely because rare cases are underrepresented in the very training data any R&D effort would use to improve the model. A widely cited industry analysis of this pattern notes that companies routinely hit steep diminishing returns pushing model accuracy from the 80% range toward 95%-plus, because most of the easy, common cases are already solved and what remains is an effectively unbounded set of edge cases that each require individually identified, individually collected training examples to fix. Mobileye's autonomous-driving research group describes the same pattern from direct production experience: more data is necessary but not sufficient, because performance on rare edge cases hits diminishing returns even as data and compute scale up. The same phenomenon has its own dedicated research literature in medical imaging — a 2022 benchmark study on chest X-ray classification found long-tailed disease distributions make rare-but-critical conditions substantially harder to learn than common ones, for the identical structural reason: standard training methods are biased toward the frequent classes because that's where the data is.

Applied to this scenario: the 30,000 hard cases are, by the scenario's own description, the unusual, atypical, edge-case requests — the long tail by definition. That means the $6.5M in R&D that Bex proposes redirecting would be aimed at exactly the population where machine learning research shows R&D dollars buy the least accuracy improvement per dollar spent, not the most. Selective coverage, by contrast, buys a guaranteed, immediate 93% accuracy on that same population today, using a resource — human judgment on atypical cases — that doesn't suffer from the long-tail data-scarcity problem at all. The R&D path and the human-review path are not simply two ways to spend $6.5M toward the same goal; one of them is structurally suited to the problem and one of them is fighting the hardest part of it with the least effective tool for the job.

This Is Not a Novel Idea: The Academic Framework Behind Selective Coverage

Selective, confidence-based routing between AI and humans is a formally studied machine learning paradigm, not an intuitive compromise invented for this scenario. It has a name in the literature — Learning to Defer — dating to a foundational 2018 NeurIPS paper by Madras, Pitassi, and Zemel, which showed that training a system to defer uncertain cases to a human can improve both accuracy and fairness simultaneously compared to forcing the AI to decide everything. It has continued as an active research area since, including 2023–2025 work by Mao, Mohri, and Zhong on multi-expert deferral and Mozannar and Sontag's consistency results for deferral algorithms. In clinical medicine specifically, a 2023 study (Dvijotham et al., published via Nature Medicine research) on "Complementarity-Driven Deferral to Clinical Workflow" (CoDoC) showed that selectively deferring uncertain cases from an AI screening model back to the standard clinical workflow improved diagnostic accuracy while also reducing clinician workload compared to either full AI automation or the standard human-only process alone. That is the same result this scenario's own numbers show: selective coverage isn't a trade-off between accuracy and cost, it's a way to improve accuracy while bounding cost, precisely because AI and human errors on hard cases tend not to be the same errors. This scenario is not testing a novel policy idea. It is testing whether an organization will adopt an architecture the ML research community has been formally validating, in increasingly rigorous ways, for going on a decade.

Addressing View A's Fairness Argument Directly

View A's fairness claim deserves a direct answer, not a dismissal: that routing atypical cases to a 3-day human queue "penalizes exactly the people with unusual circumstances." This is true as far as it goes, but it compares the wrong two things. The real comparison isn't instant-and-right versus slow-and-right. It's instant-and-wrong (nearly 1-in-4 of the time, per the derived 75.8% figure above) versus a 3-day wait for a decision that's right 93% of the time. For a rejected claim, a missed critical case, or a wrongly rejected application, three days is a delay. A wrong instant decision, especially in the categories this scenario lists — claims, applications, referrals, case reviews — can be a harm that a later appeal often cannot fully undo, particularly for people without the resources or knowledge to successfully navigate an appeals process. That asymmetry, not the wait time alone, is the actual fairness question, and it is the one Robodebt and the Dutch benefits scandal answer most clearly: the people least equipped to contest a wrong automated decision were disproportionately the ones harmed by it, and an appeals path that exists on paper did not protect them in either case.

Final Position

Selective coverage. The scenario's own numbers, once the hidden accuracy figure is derived, show the AI's real performance on hard cases collapsing to roughly 76% when forced to decide alone — a 21-point drop from its performance on easy cases, and 17 points below what a human reviewer achieves on the same hard cases. The $6.5M annual cost works out to about $5.42 per request across the full volume, and even a conservative, explicitly illustrative harm-weighting shows selective coverage costing less overall than full coverage once downstream consequences are counted, not more. Robodebt, the Dutch childcare benefits scandal, and the UK Post Office Horizon scandal are three independent, well-documented, still-unfolding illustrations — one with a settlement approved within the last month — of what happens when an organization lets an automated system decide atypical, high-stakes cases without a genuine human check. Meanwhile, American Express's actual fraud system (correcting Bex's own example), the FDA's own model for autonomous medical AI, the IRS's decades-old audit process, airport security, content moderation, and modern credit underwriting all converge on the same architecture this scenario calls selective coverage — not because it is fashionable, but because in every one of these domains, someone eventually had to answer for what happens when the hard cases are decided by a system that doesn't know it's out of its depth. Know when to abstain. That is not the AI's weakness. It is the only part of this system anyone should actually trust with the cases that matter most.

Let the AI Decide Everything

Why Full Coverage beats Selective Coverage — and why the case for it is stronger than it's usually argued

I support View A — Full Coverage. The AI decides all 100,000 requests, instantly, at effectively zero marginal cost. The reliability gap this leaves (9,000 wrong decisions/month instead of 3,850) is real, but it is cheaper and fairer to close it with a genuine, fast appeals-and-correction path than to pre-emptively route 30,000 people/month into a slower human lane — a lane that, by construction, fills up with exactly the people whose situations are already the most unusual and the most consequential to get wrong.

The Core Analogy: Automated Tolling vs. the Inspection Booth

Every highway toll system today reads every car at full speed. It doesn't achieve courtroom-level certainty — plates get misread, transponders glitch, tags expire unnoticed. But the system doesn't pull the ambiguous-looking car into a booth and make it wait three days for a manual read. It lets every car through at highway speed and fixes errors afterward: a bill is mailed, and the driver disputes it if it's wrong. Nobody proposes stopping every car with an out-of-state plate or an unusual vehicle shape for a manual inspection “to be fair” — that would defeat the entire purpose of the system, and it would specifically punish the drivers whose only distinguishing feature is that their car looks a little different.

That is precisely the structure of View A. Full Coverage processes every request at machine speed and corrects errors through appeal after the fact. Selective Coverage recreates the inspection booth: it identifies the 30,000 “unusual-looking” cases every month and stops them for three days, whether or not anything is actually wrong.

Figure 1 — Full Coverage keeps every case moving and corrects errors afterward; Selective Coverage recreates an inspection lane that only atypical cases enter.

The Numbers, Reframed

Selective Coverage does cut wrong decisions by 57% — 9,000 to 3,850 per month. But that reduction is not free, and the honest way to read $6.5M/year is as a price per unit of improvement, not just a headline cost.

Figure 2 — $6.5M/year buys 61,800 fewer wrong decisions annually — $1,260 per additional error avoided, to move from a 91%-accurate system to a 97.5%-accurate one reviewed by humans who are themselves only 93% accurate.

A narrower, genuinely need-based correction path — one that activates only when a customer actually disputes a decision, rather than pre-emptively reviewing all 30,000 edge cases whether or not anyone objects — could plausibly capture a meaningful share of that same error-reduction value without the fixed $6.5M/year cost of staffing a standing 30,000-case review queue every month regardless of demand. This is a reasoned inference from the numbers given, not a sourced industry figure — but it is the comparison Selective Coverage's $6.5M price tag should actually be measured against.

Five Live Industry Examples — With Numbers

These are not hypothetical. They are current, large-scale deployments where organizations chose to push AI coverage as close to 100% as the technology allows, rather than defaulting a fixed slice of volume into a slower human lane.

1. Visa Decision Manager — 98.83% of transactions resolved automatically

Visa's AI fraud-risk platform screens transactions in milliseconds, and 98.83% of Decision Manager transactions are resolved automatically by AI with no human step in the loop at the moment of decision. Visa doesn't route the other ~1.2% into a multi-day human queue before letting the purchase go through — the transaction still completes, with monitoring and dispute mechanisms behind it.

Source: Visa Corporate, “The future of fraud detection: smarter, faster, safer,” corporate.visa.com

2.Mastercard Decision Intelligence — sub-50-millisecond decisions at global scale

Every Mastercard transaction is scored in under 50 milliseconds. In Mastercard's 2025 fraud prevention survey, 80% of organizations said AI eliminated unnecessary manual reviews, and 42% of issuers plus 26% of acquirers each saved more than $5 million in fraud losses over two years by leaning into automated decisioning rather than expanding manual review teams.

Source: Mastercard / Financial Times Longitude, 2025 Payment Fraud Prevention Report

3.Gmail spam filtering — 99.9% catch rate across 15 billion emails a day

Gmail filters roughly 15 billion emails daily and blocks over 100 million phishing emails a day, with a spam-catch rate exceeding 99.9%. There is no human review tier for individual emails — every message gets an instant, fully automated decision. The correction mechanism is a single “Not Spam” click, exactly the kind of cheap, fast appeals path View A proposes, and it's precisely what makes 100% AI coverage viable at this scale.

Source: Google Gmail transparency reporting; RETVec spam-classifier update notes

4.Lemonade Insurance — 96% of claims intake, 55% fully resolved, by AI alone

Lemonade's AI claims agent handles 96% of First Notice of Loss submissions without human involvement, and as of year-end 2025, 55% of all claims are fully automated start to finish — often resolved in seconds rather than the industry-standard 30–45 days. This is the most AI-forward insurer in the industry, and its trajectory has been to keep pushing that percentage up, not to lock in a fixed 30% human-review lane as a permanent feature.

Source: Lemonade Q4 2025 shareholder letter; Devoteam/Trixly AI insurance case studies

5.Digital account opening — 90%+ automated identity and fraud clearance

Across major card issuers' digital application funnels, 90%+ of applicants clear identity/KYC verification through automated checks alone, with only a small single-digit percentage flagged for any manual fraud review — the underwriting decision itself (approve/decline) then runs on an automated credit-scoring model for the overwhelming majority of the remaining funnel. This is a genuinely consequential financial decision — access to credit — made almost entirely without a human in the loop at time of decision.

Source: Industry credit-card approval funnel analysis, cardsftw.com, 2026

Figure 3 — Across five live, large-scale deployments, AI already resolves the large majority of decisions without a human in the loop at the moment of decision — well past the 70% this scenario proposes routing to instant AI resolution.

The Equity Argument, With the Prompt's Own Numbers

View A's strongest point isn't cost — it's who selective coverage actually burdens. The 30,000 cases routed to human review every month are not a random sample; they are, by definition, the unusual, complex, and ambiguous cases. Annualized, that's 360,000 people a year who wait three days for a decision purely because their situation didn't fit the standard pattern — not because anything was necessarily wrong with their request.

This is a regressive design. The people already most likely to have non-standard documentation, atypical histories, or unconventional circumstances — often correlated with exactly the populations who face friction elsewhere in institutional processes — are the ones systematically pushed into the slow lane. Full Coverage, paired with a real appeals path available to anyone who disputes their outcome, treats the standard applicant and the unusual applicant identically at the moment of decision, and reserves extra scrutiny for actual disputes rather than for looking atypical at intake.

Where This Goes Beyond Bex's Argument

Bex also supports Full Coverage, but the case as written leans on a single unevidenced example and misses the strongest parts of View A's own logic. Specifically:

▪Bex's AmEx example has no numbers attached to it. “Maintaining a high accuracy rate” isn't a figure — it's a phrase. This paper instead uses five examples with specific, sourced percentages (98.83%, sub-50ms, 99.9%, 96%/55%, 90%+), which is the standard Bex's own argument should be held to.

▪Bex never engages with the $1,260-per-error-avoided framing at all. Simply citing that $6.5M is avoided isn't an argument about value — it skips the question of what that $6.5M is actually buying, which this paper answers directly with Figure 2.

▪Bex is silent on the equity argument entirely. The strongest sentence in View A's own framing — that the escalation queue penalizes people specifically for being atypical — goes unaddressed in Bex's response. That argument, not the cost savings, is what makes Full Coverage a fairness position, not just an efficiency one.

▪Bex treats “avoiding the $6.5M cost” as the end of the analysis. It's the start of one: the real question is what a cheaper, need-based correction mechanism could achieve instead of a blanket 30,000-case human queue — addressed above and not raised by Bex at all.

The Honest Risk — And Why It Doesn't Change the Conclusion

The clearest real-world caution for View A is Meta's 2025–2026 shift toward AI handling up to 90% of content moderation. Independent reporting has documented a genuine problem: users describe appeals that go into a black box, a one-and-done appeal limit with no further recourse, and businesses losing their accounts with no accessible human to reach. That is a legitimate failure mode, and it is worth taking seriously rather than waving away.

But look closely at what actually failed. It wasn't full AI coverage itself — it was that the appeals path was theater: slow, opaque, and effectively unreachable. That is precisely the condition View A depends on not repeating. The argument here is not “automate everything and hope”; it's “automate everything and back it with a real, fast, accessible correction path” — closer to Gmail's one-click “Not Spam” or a toll system's straightforward dispute process than to Meta's current appeals process. The lesson from Meta is a design requirement for View A, not a reason to abandon it: the $6.5M Selective Coverage would have spent pre-emptively reviewing 30,000 cases a month should instead fund a fast, adequately staffed, genuinely reachable appeals function for the much smaller number of people who actually contest a decision.

Conclusion

Selective Coverage's 57% error reduction is real, but it is bought by treating atypical circumstances as a trigger for a three-day wait, at $1,260 per error avoided, for a review tier that is itself only 93% accurate. Five live deployments — Visa, Mastercard, Gmail, Lemonade, and digital account opening — show that pushing AI coverage well past 70%, often past 95%, is already how the highest-volume, most consequential automated decision systems in the world operate, precisely because the combination of instant decisions and a genuine correction path outperforms a permanent human lane sized for the hardest cases.

Read every car at highway speed, and fix the rare misread afterward. Don't build a booth that only the unusual cars have to stop at — that isn't protection, it's a tax on being different.

 

1. Clear Positioning Statement

I strongly support View A Bex's position.

In enterprise AI architecture, designing for "perfection" at the point of ingestion is an operational fallacy. Attempting to solve 100% of edge cases through expensive, slow, front-end human intervention ($6.5M/year) neglects the systemic reality of human error (93\% accuracy on complex cases) and severely penalizes atypical users with a 3-day latency penalty.

Instead, a high-velocity, high-efficiency AI engine (91\% accuracy) coupled with an asynchronous, risk-tiered "Appeals & Exception Management Pipeline" yields a superior Net Present Value (NPV), guarantees uniform customer experience, and creates a closed-loop system where AI continuously learns from corrected errors.

2. Quality Reasoning: Deconstructing the Fallacy of View B

To demonstrate why View A is mathematically and operationally superior, we must analyze the "cost-per-error-avoided" and the "human oracle" fallacy:

A. The "Human Oracle" Fallacy & Net Error Analysis

View B assumes humans are a flawless safety net. However, the data shows human accuracy on these escalated cases is only 93\%.

• Under Selective Coverage (View B), the 30,000 escalated cases processed by humans will still result in 2,100 wrong decisions (30,000 \times (1 - 0.93).

• Under Full Coverage (View A), the AI processes these same 30,000 cases at an estimated 75.8\% accuracy (derived from the math: (100,000 \times 0.91) - (70,000 \times 0.975) = 22,750 correct decisions; 22,750 / 30,000 \approx 75.83\%), resulting in 7,250 errors.

• The delta between Human and AI error on these hard cases is 5,150 errors.

To prevent these 5,150 errors, View B spends $6.5 Million annually. This equates to $1,262 per error avoided. In the vast majority of operational workflows (claims, ticketing, routing, basic approvals), the average cost of remediating an error post-facto is a fraction of this amount.

B. The Latency and Equity Penalty

View B penalizes "atypical" customers by forcing them into a 3-day queue. This disproportionately impacts marginalized or non-standard profiles (e.g., immigrants with thin credit files, freelancers with non-standard income). Under View A, everyone receives an instant decision, and the small subset of erroneous decisions is corrected via a highly optimized, frictionless appeals pathway.

3. Real-World Evidence

To move beyond assertion, we look at how leading global organizations leverage high-velocity, full-coverage AI architectures:

1. Klarna (Customer Service & Decisions)

• Context: Klarna deployed an OpenAI-powered AI assistant to handle customer service chats and initial resolution decisions.

• Data: The AI took over the work of 700 full-time agents, processing 2.3 million conversations in its first month.

• Outcome: It achieved equal customer satisfaction (CSAT) to humans but reduced errand resolution times from 77 minutes to under 2 minutes, saving Klarna an estimated $40 million in annual run-rate savings, proving that instant, automated resolution at scale offsets the need for massive front-end human triage.

2. Ping An Insurance (Instant Claims Processing)

• Context: Ping An implemented "Smart Fast Claim" for motor vehicle insurance.

• Data: The system processes over 90% of traffic accident claims automatically using image recognition and AI-driven decisioning.

• Outcome: Claims are settled in under 3 minutes (instant payout). By bypassing manual human reviews for the vast majority of cases, Ping An slashed operational overhead while maintaining fraud detection rates, using post-payment audits to catch any anomalies.

3. Lemonade Insurance (AI Jim)

• Context: Lemonade’s claims bot, "AI Jim," handles claims end-to-end without human intervention.

• Data: In a record-setting case, AI Jim reviewed, cross-referenced, approved, and paid out a theft claim in 3 seconds flat.

• Outcome: While some complex cases are routed, Lemonade’s philosophy of "instant-first" has driven their loss ratio down while scaling to millions of customers with minimal human claims staff, relying on post-facto algorithmic audits.

4. Amazon (Automated Inventory & Vendor Decisions)

• Context: Amazon’s "Hands off the Wheel" (HOTW) initiative shifted purchasing, pricing, and inventory forecasting decisions from human buyers to automated AI systems.

• Data: Millions of product SKUs are managed entirely by algorithms.

• Outcome: Despite initial edge-case errors (e.g., ordering too much of a niche product), the sheer speed, scale, and consistency of the system eliminated billions in operational overhead and outperformed the collective accuracy of human buyers over time.

5. Uber (Real-time Driver Document Verification)

• Context: Uber uses computer vision and NLP to approve or reject driver documents (licenses, insurance) globally.

• Data: Instant approvals allow drivers to get on the road immediately rather than waiting days for manual back-office verification

• Outcome: Uber accepts a tiny margin of OCR processing errors, mitigated by real-time random spot-checks and automated flags, saving millions in manual review centers globally.

6. Capital One (Credit Card Instant Approvals)

• Context: Capital One uses machine learning models to instantly approve or decline credit card applications, even for thin-file or subprime applicants.

• Data: By opting for instant decisions rather than putting complex files into "manual underwriting" queues (which can take 7–14 days), they capture high-value customers instantly.

Outcome: The lifetime value (LTV) of captured customers vastly outweighs the marginal credit risk of automated edge-case approvals.

7. Visa and Mastercard — full coverage at the largest scale that exists.

   Visa alone processed <cite index="18-1">over 233 billion transactions in 2024</cite>. Every one is approved or declined by an automated risk engine in real time — there is no human-review queue for "uncertain" purchases before checkout completes. Errors aren't prevented by slowing transactions down; they're caught afterward through a structured dispute/chargeback system, with <cite index="19-1">clear-cut cases like confirmed fraud routed through a fast, rules-based automated workflow and only genuinely complex disputes needing manual evidence review</cite>. That's full coverage plus tiered, automatic correction — not full coverage plus "good luck."

8. Gmail spam filtering — full coverage, self-correcting, no queue.

   Gmail evaluates <cite index="29-1">roughly 15 billion messages a day across 1.8 billion accounts</cite> and <cite index="28-1">blocks over 99.9% of spam, phishing and malware, with the RETVec update improving detection by 38% while cutting false positives by 19.4%</cite>. There's no selective escalation of "ambiguous" emails to a human moderator before delivery — every message gets an instant, fully automated verdict, and correction happens through a lightweight, always-available user action (mark as spam/not spam) that feeds back into the model.

9. American Express — the example Bex already raised, done right.

   Amex, like the other major networks, authorizes the overwhelming majority of transactions instantly through automated fraud models rather than routing uncertain purchases to a human before completing the sale. This is the correct shape for our scenario: instant coverage of everyone, with fraud investigation and reversal happening after the fact for the transactions that turn out to be wrong — not before, for a subset flagged as "hard."

10. UnitedHealth's nH Predict — a real cautionary tale, and the one View B would cite against me. I'll address it directly.

    UnitedHealth used an AI model to make post-acute care coverage decisions. A federal lawsuit alleges that <cite index="11-1">when denials were appealed, nine of ten were ultimately reversed</cite> — yet <cite index="10-1">fewer than 0.2% of patients ever filed an appeal</cite>. This is a genuinely damaging data point, and it does not support a naive version of View A. It proves that "a cheap appeals path" only works if people actually use it, and here they didn't — because the burden of initiating, documenting and pursuing an appeal fell entirely on a sick, unfamiliar, disempowered patient. The lesson isn't "abandon full coverage," it's "never make correction opt-in." Any full-coverage design that relies on the affected person to notice the error, understand it, and fight for a reversal will fail exactly this way. That directly shapes the framework below.

11. Robodebt (Australia) — the actual floor View A must never sink to.

    Australia's automated welfare debt-recovery system made <cite index="7-1">debt assessments against 453,000 people totaling $1.7 billion</cite>, built on an income-averaging method that <cite index="9-1">a Royal Commission found was applied in 76% of overpayment assessments despite being unsuited to people with irregular work hours</cite>, and it ran with <cite index="3-1">no human in the loop at all</cite>. The scheme was ruled unlawful, and <cite index="8-1">the government ultimately paid $112 million in compensation on top of refunding over $751 million in wrongly collected debts</cite>. This is what full coverage looks like when there is no correction mechanism, no explanation, and no monitoring for who's being harmed — it isn't an argument against full coverage, it's an argument against full coverage deployed carelessly. It tells us exactly what safeguards are non-negotiable.

12. The scenario's own numbers, read closely.

    Selective coverage isn't "safety" — it's swapping one error source for another at a steep price. Human reviewers are only 93% accurate on the hard cases, meaning even the "protected" tier still produces roughly 2,100 wrong decisions a month. The $6.5M buys a reduction from 9,000 to 3,850 errors — $1,260 per error avoided — while making 30,000 people wait three days regardless of whether their case was actually going to be decided correctly or not. That's a blunt, expensive instrument: everyone atypical pays the delay cost, whether or not they were one of the ~2,000 who would've been wronged anyway.

4. Deployable Solution Framework: "Active Feedback & Asynchronous Remediation" (AFAR)

To operationalize View A safely, we must build a system that manages the 9,000 errors without paying $6.5M upfront. We propose the AFAR Framework:

Phase 1: High-Velocity Execution Engine (The "Fast Lane")

• Action: Deploy the AI model to process 100,000 cases with 100% coverage.

• Mechanism: Every decision is accompanied by a "Micro-Explanation" (using SHAP/LIME) sent to the customer (e.g., "Your claim was declined because document X was missing..."). This empowers the customer to self-remediate instantly.

Phase 2: Post-Facto Anomaly Detection & Shadow Audit

• Action: Rather than reviewing 30,000 cases before a decision, run an asynchronous "Shadow Audit" on high-risk sectors of the 100,000 processed cases.

• Mechanism: An LLM-based auditor flags high-consequence decisions for post-payout or post-resolution review. Humans only review flagged anomalies (5,000 cases/month instead of 30,000).

Phase 3: Frictionless Digital Appeal Path

• Action: Provide an instant, one-click "Appeal" button for any rejected request.

  1. Mechanism: If a customer appeals, the system requests the missing specific data point. This appeal is routed to a highly specialized, small human-in-the-loop (HITL) team, drastically reducing the size of the manual review team.

     Budget logic: Instead of $6.5M for a standing pre-decision review team, allocate a smaller fund (illustratively $1.5–2M/year) to the automatic secondary-review and rapid-correction layer. If it catches even 50–65% of the 9,000 monthly errors within days — plausible, given Visa's model resolves most disputes in weeks and nH Predict shows the reversal rate on review is very high (up to 90%), the limiting factor is initiation, not review capacity — net wrong decisions could land around 3,150–4,500/month, in the same range as selective coverage's 3,850, at roughly a quarter of the cost, with zero people waiting three days for their first answer.

     What to measure (before claiming success)

     1. Net wrong decisions/month — after correction, not just at point of decision.

     2. % of flagged/sampled cases corrected within 72 hours.

     3. Error rate by case segment (atypical/complex vs. routine) — this is the number that would prove or disprove View B's fairness concern in production, and it must not be allowed to drift.

     4. % of corrections system-initiated vs. user-initiated — target >80% system-initiated, directly guarding against the nH Predict failure mode.

     5. Cost per net error avoided, benchmarked against selective coverage's $1,260.

     6. Complaint and regulatory-inquiry trend — an early warning if segment-level harm is building.

5. Measurement and KPI Dashboard

To ensure the AFAR framework delivers the promised value, we will track the following key metrics:

Metric Category Key Performance Indicator (KPI) Target Benchmark

Financial Net Cost Savings > $5.0M saved annually (calculating $6.5M minus appeal operations cost)

Financial Cost per Resolved Error < $150 per actual error corrected (vs. $1,262 in View B)

Operational Average Cycle Time (ACT) < 5 minutes for 95\% of all applicants

Customer Experience Appeals Conversion Rate < 10% of rejected applicants initiate an appeal

Model Health Continuous Learning Yield Model error rate decreases by > 2\% quarter-over- quarter via feedback loops

Conclusion

View B is a defensive strategy designed around the fear of failure, resulting in massive cost overhead and a poor user experience.

By contrast, View A, powered by the AFAR Framework, is an offensive, modern AI architecture. It accepts operational noise as an optimization problem, leverages speed as a competitive advantage, and uses smart, post-facto human intervention to keep costs low while driving continuous system improvement.

Full coverage wins on speed, cost, and consistency — that part of View A is already strong. The missing piece, which View B correctly senses, is that "instant and cheap" isn't automatically "safe" for the hard cases. The fix isn't to slow down 30% of people; it's to make correction automatic, fast, and system-driven rather than a queue people have to survive or an appeal they have to know to file.

View B is really making two claims, and only one of them survives scrutiny. The first — that errors concentrate in the hard cases and someone should be watching for that — is correct, and any responsible deployment has to build for it. The second — that the only way to catch those errors is to make 30,000 people wait three days for a human who is right just 93% of the time — does not hold up. Selective coverage doesn’t eliminate the hard-case problem; it relocates it, adds a second, non-trivial error source, and charges $6.5M a year, or $1,260 per error avoided, for the privilege.

Every real system examined here that operates at genuine scale — Visa’s 233 billion transactions, Gmail’s 15 billion messages a day, Amex’s fraud engine — has already made the same choice this scenario is asking us to make, and none of them do it by routing uncertain cases to a pre-decision human queue. They decide everything, instantly, and they catch what they got wrong afterward, fast and automatically. Where full coverage has gone badly wrong — Robodebt, nH Predict — the cause was never “there was no queue.” It was that correction was either absent entirely or left for the harmed person to discover and pursue alone. That is a design failure, not an indictment of full coverage itself, and it is precisely the failure the FC³ framework is built to close.

So the choice isn’t full coverage versus protecting the vulnerable. It’s whether you protect them before the decision, at $1,260 per error and a three-day tax on everyone atypical, or after it, at a fraction of the cost and zero wait for anyone. So back Bex’s position, strengthened this way: deploy full coverage, decide every case instantly, and spend the money not on a gate but on a fast, automatic, system-initiated correction layer that finds and fixes the errors before they can compound. That is the version of “let the AI decide everything” that actually earns the name — not because it never fails the hard cases, but because it is built to catch itself when it does.

I support View A — Full Coverage. Let the AI decide everything.

The numbers tell a story that View B's headline hides: selective coverage costs $6.5M per year for a human review team that is still 93% accurate on complex cases — meaning they get 7 out of every 100 hard decisions wrong anyway. You are not buying certainty for $6.5M. You are buying a 2.5 percentage point accuracy improvement on 30% of cases. That is $1,260 per wrong decision avoided — an expensive band-aid on a problem better solved at the root.

This is not an argument for selective coverage. It is an argument for deploying full coverage now, while redirecting human effort from permanent decision-making to active model improvement — so that the AI earns its 97.5% accuracy across all cases, not just the easy ones.

View Comparison — Selective Coverage vs Full Coverage + HITL
Criterion	View B — Selective Coverage	Our Approach — View A Full Coverage + HITL
Human role	Permanent decision-maker on complex cases	Temporary trainer to improve the AI
End state	Two-tier system forever	Full AI coverage, increasingly autonomous
Cost	$6.5M permanent annual line	Reduces over time as AI matures
Who bears the wait	30% of customers — permanently	Nobody — transitional phase only
What humans do	Review and decide	Retrain and improve input quality
Goal	Patch AI weakness	Eliminate AI weakness

Bex is right on efficiency, but misses the more important point: that $6.5M is better spent fixing the AI than patching it with human review.

Here is what selective coverage gets backwards — it treats human escalation as the destination. It should be the training ground. Every low-confidence case a human reviews is a labeled data point. Every correction is a model improvement signal. The right architecture is full AI coverage today, with humans redirected to two higher-value activities: model retraining on edge cases, and input quality improvement. Better indexed, cleaner inputs reaching the AI means fewer low-confidence cases next month. The 30,000 escalations shrink over time. The $6.5M cost disappears progressively rather than becoming a permanent operating line.

Two FinTech examples prove this is not theory — it is a proven trajectory.

In my current organisation, for few clients, we use BioCatch — a behavioral biometrics AI — for fraud detection. When deployed in 2025, the system flagged over 30 alert types requiring human review and decision. Full AI coverage at the detection layer, human review at the decision layer. As humans reviewed and decided, the model learned. Today, alert types have reduced by half — the AI now takes decisions autonomously on cases it has learned from, with humans retaining oversight only on genuinely novel patterns. The $6.5M question answers itself: the human review cost reduces as the AI matures.
Upstart, the AI-driven lending platform, took full coverage further. During 2024–2025, over 91% of Upstart-powered loans are fully automated with zero human involvement — including non-standard borrowers with thin credit files, the exact complex edge cases View B would escalate to humans. The AI approves 44% more borrowers than traditional models while driving more inclusive lending, with nearly 29% of loans going to low-to-moderate income communities. The accuracy did not collapse under full coverage — it improved continuously because humans were redeployed to train the model, not review its decisions one by one.

The appeals path handles consequential errors adequately. What no appeals path handles is a $6.5M annual cost that becomes structurally permanent, or a 30% customer segment stuck in the slow lane indefinitely — simply for being atypical.

Full coverage is not the end state — it is the starting point. Direct the savings toward making the AI worthy of that trust.

1. Suhail_J_CaJq

Position: View B (Selective coverage — the system must know what it doesn't know)

Specific Example: Cites "Health Insurance Prior Authorization" in general terms — routine vs. complex claims (rare conditions, unusual treatment plans, conflicting clinical notes) — but names no specific insurer, no dates, no figures, and no source.

Reasoning Quality: Competent. The five-point structure (harm concentration, value of abstention, human review being directionally correct, cost vs. harm, speed vs. correctness) is logically coherent and correctly restates the scenario's own numbers, but it never leaves the hypothetical and adds no external evidence.

✗ Not Approved — the position is clear, but the "example" is a generic industry description with no named company, timeline, or documented outcome, so it doesn't clear the specific-example bar.

---

2. rajan.arora2000

Position: View B (Selective coverage, without qualification)

Specific Example: An extraordinarily well-documented portfolio: Michigan's MiDAS unemployment system (Oct 2013–Aug 2015, ~400 staff cut, ~40,000 algorithm-only fraud determinations, 93% found non-fraudulent by the state Auditor General, Bauserman v. UIA $20M settlement approved Jan 2024); Australia's Robodebt (453,000 people, A$565M net loss per the 2023 Royal Commission); Cigna's PXDX as a negative control (300,000+ denials in two months, 1.2 seconds/case, Kisting-Leung v. Cigna); the UK Post Office Horizon scandal (£1,628M paid to 12,900+ claimants as of June 2026); and card-fraud authorization (Visa/Amex) used as a positive control for View A. Every case is cited with sources, dates, and figures.

Reasoning Quality: Exceptional. Corrects the scenario's own arithmetic error ($1,260 vs. the true $104.85 per error avoided), applies Chow's (1970) decision-theoretic reject-option formula, runs sensitivity/robustness checks in both directions, steelmans View A honestly, directly upgrades and rebuts Bex's own example, and closes with a falsifiable, numeric "abstention test" plus a pre-committed condition under which the author would reverse position.

✅ Approved — a rigorously argued View B position, built on multiple named, sourced, and dated real-world cases plus formal decision theory, that meets every element of the threshold at the highest level seen in this thread.

---

3. GoutamNamata

Position: View B (Selective coverage)

Specific Example: References "health insurance claims processing" as a strong example but does not name an insurer, cite a source, or provide any documented figures beyond restating the scenario's own numbers.

Reasoning Quality: Reasonable. The argument that the escalated 30,000 cases are where harm concentrates is sound and clearly stated, but the post does not go beyond paraphrasing the prompt's own data.

✗ Not Approved — position is clear, but the example is an unnamed, generic industry reference rather than a documented real-world case.

---

4. Ajay _Wadhwa_bs1h

Position: View B (Selective coverage)

Specific Example: Names Fannie Mae's Desktop Underwriter specifically, describing its actual process — instant approval for stable-income/clean-credit applicants, automatic routing of self-employed or non-traditional-income borrowers to manual review — and ties this to fair-lending exposure.

Reasoning Quality: Good. The "averaging hides where errors land" argument and the point that appeals disproportionately help people who already know how to navigate the system are sharp and original, even though no financial figures or dates are attached to the mortgage example.

✅ Approved — names a real, specific system with concrete process detail (not just a company name) and connects it coherently to the View B position.

---

5. Naijur Rahman

Position: View B (Selective coverage — the system must know what it doesn't know)

Specific Example: Derives the AI's hidden ~75.8% accuracy on hard cases from the scenario's own numbers, then supports the position with the Dutch Childcare Benefits scandal (26,000–35,000 families flagged, €20k–60k clawbacks, government fell Jan 2021), the Post Office Horizon scandal (£1.44B paid to 11,300+ claimants), Robodebt, and the FDA-authorized IDx-DR diagnostic system, plus academic citations (Madras et al. 2018 NeurIPS "Learning to Defer," a 2023 Nature Medicine CoDoC study).

Reasoning Quality: Exceptional. Directly rebuts Bex point-by-point, cites the machine-learning "long-tail problem" literature to explain why redirecting the $6.5M into R&D wouldn't close the gap, and reframes the fairness argument precisely.

✅ Approved — a clear position supported by multiple sourced, dated real-world scandals and academic literature, argued with unusual rigor.

---

6. anthony rebello

Position: View A (Full coverage — let the AI decide everything)

Specific Example: Visa Decision Manager (98.83% of transactions resolved automatically), Mastercard's 2025 fraud-prevention survey (80% of organizations eliminated unnecessary manual review; issuers/acquirers saved $5M+ over two years), Gmail (15 billion emails/day, 99.9%+ spam catch rate), and Lemonade Insurance (96% of claims intake automated, 55% fully resolved by AI as of year-end 2025), each with a named source.

Reasoning Quality: High quality. The toll-booth analogy is effective, the equity argument (that a review queue is "regressive" against atypical applicants) is a genuine original contribution, and the author honestly engages Meta's 2025–2026 moderation appeals failures as a cautionary design constraint rather than ignoring it.

✅ Approved — a clearly stated View A position backed by five named, sourced, large-scale deployments and reasoning that directly engages the strongest counter-evidence.

---

7. Adeniran_Ilesanmi_GYSH

Position: View B (Selective coverage — the system must know what it doesn't know)

Specific Example: Names Lemonade, Aviva, AXA, NHS 111, Mayo Clinic, Monzo, Revolut, Stripe, DWP, HMRC, and USCIS, but attaches no citations, dates, or verifiable figures to any of them — the stated percentages (e.g., "reduces claim disputes by 40–60%") appear to be asserted rather than sourced.

Reasoning Quality: Competent but ungrounded. The sensitivity tables and "risk-adjusted ROI" model are elaborate, but the cost distributions and probabilities (e.g., a 2% "catastrophic" event at $300,000) are invented for the exercise rather than drawn from documented data, so the quantitative rigor is more decorative than evidentiary.

✗ Not Approved — the real-world companies are name-dropped without documented process detail, timelines, or cited figures, so the example requirement is not met despite the length of the response.

---

8. Prateek _Harsh_dl5h

Position: View A ("I strongly support View A — Bex's position")

Specific Example: Klarna's OpenAI-powered assistant (replaced the work of 700 agents, 2.3 million conversations in month one, resolution time cut from 77 minutes to under 2 minutes, ~$40M annual run-rate savings), Ping An's "Smart Fast Claim" (90%+ of motor claims auto-processed, settled in under 3 minutes), and Visa (233 billion transactions in 2024, cited). The author also honestly addresses UnitedHealth's nH Predict controversy (90% of appealed denials reversed, but under 0.2% of patients ever appealed) and Robodebt as genuine failure modes rather than omitting them.

Reasoning Quality: Exceptional. Proposes a concrete alternative architecture (the "AFAR" framework) with its own KPI dashboard, directly derives the same 75.8% hard-case AI accuracy figure independently, and turns the strongest anti-View-A evidence into design requirements rather than dismissing it.

✅ Approved — a clearly argued View A position with multiple named, figure-rich examples and reasoning that engages the strongest counter-evidence directly.

---

9. Dinesh Selvarajan

Position: View A (Full coverage — let the AI decide everything, with humans redeployed to training rather than permanent review)

Specific Example: BioCatch behavioral-biometrics fraud detection, described as used in the author's own organization (deployed 2025, flagged 30+ alert types, alert types cut roughly in half over time as the model learned), and Upstart's AI lending platform (2024–2025, 91%+ of loans fully automated, approves 44% more borrowers, ~29% of loans to low-to-moderate income communities).

Reasoning Quality: Good. The core insight — that human review should be treated as a training signal rather than a permanent institution — is a genuinely different angle from the other View A responses, and the comparison table is clear, though the argument is less rigorously sourced than Prateek's or anthony's entries.

✅ Approved — a clear position supported by two named, reasonably specific real-world examples (one with direct organizational experience, one with quantified outcomes) and a coherent, if less exhaustively documented, argument.

---

🏆 Winner: rajan.arora2000

Among the six approved entries, rajan.arora2000 stands apart on all three criteria. On clarity of position, every other approved entry states a view and defends it, but rajan is the only one to formally derive a numeric threshold ("the Abstention Test," with four falsifiable conditions) and pre-commit to reversing his own position if the data changes — a level of intellectual honesty none of the others (including the excellent Naijur Rahman and Prateek_Harsh_dl5h) attempt. On reasoning quality, rajan is the only participant to catch and correct an actual arithmetic error in the scenario itself (the $1,260 figure conflating monthly and annual units), apply a named academic decision-theory framework (Chow, 1970) to compute break-even costs from first principles, and systematically steelman and then dismantle the opposing view point-by-point — going well beyond Naijur Rahman's strong but more narratively structured rebuttal and Prateek_Harsh_dl5h's excellent but comparatively less mathematically rigorous AFAR proposal. On specificity of examples, rajan's five-sector, three-jurisdiction evidence base (Michigan MiDAS, Robodebt, Cigna PXDX as a deliberate negative control, Post Office Horizon, and card-network fraud detection as a positive control) is unmatched in this thread for its use of matched natural experiments, disclosed confounds, and precise, dated financial figures — a level of self-critical, source-dense argumentation that no other entry, on either side of the debate, achieves.