Pratik Dilip Gawande

Position: View B — AI should not implement process changes autonomously. Confidence is not accountability. Challenging the “high-confidence = autonomy” argumentBex’s position assumes that once AI is accurate enough, it should be allowed to act independently. That sounds efficient — but it confuses prediction accuracy with decision accountability. AI can optimize for patterns. But process changes impact risk, compliance, and unintended consequences — areas where accountability cannot be delegated. A system can be 95% right and still cause 100% impact when it is wrong. The real risk is not wrong decisions — it is invisible decisionsWhen AI only recommends, decisions are visible, discussed, and owned. When AI implements autonomously, decisions become silent changes in the system. That is where risk multiplies: small changes compound over time, interdependencies are missed, and accountability becomes unclear when something breaks. Speed without visibility is not agility. It is uncontrolled drift. Example: US Payroll & Shared ServicesIn payroll operations, even a “high-confidence” AI-driven change — such as: modifying tax calculation logic, altering validation thresholds, or auto-adjusting pay components can have immediate financial and compliance impact. A seemingly beneficial optimization could: miscalculate taxes across thousands of employees, violate federal/state compliance rules, or trigger incorrect net pay at scale The cost is not theoretical: payroll errors can trigger $100K+ penalties, employee trust erosion, and legal exposure No payroll leader would allow autonomous implementation — not because AI is weak, but because the blast radius of a mistake is too high. What high-performing organisations actually doMature organisations do not block AI — they redefine its role: AI identifies and prioritizes improvements Humans validate and approve changes Systems automate execution after approval This keeps: speed in insight, control in decision, and accountability in outcome The leadership principleAutomation is powerful in execution. But change is a leadership decision, not a system action. The moment organisations allow AI to implement changes independently, they are not accelerating improvement — they are outsourcing responsibility. Final PositionAI should accelerate thinking, not replace ownership. Confidence thresholds are mathematical. Business impact is not. The question is not “Can AI act?” The question is “Who is accountable when it does?” That is why View B is not resistance to AI — it is the only model that preserves control, trust, and responsibility while still leveraging AI at scale.

Position: View A — Prioritize immediate resolution. Learning is critical, but stability is non-negotiable. Challenging the “learning-first” argument directlyView B assumes that deeper learning should take precedence because it prevents recurrence. That sounds strategically sound — but it ignores a fundamental reality: You cannot learn from a system that is still failing in real time. When customers are impacted, delays are not neutral — they are damage. Every additional minute spent analysing instead of stabilising compounds that damage. Learning creates future value. Resolution protects present trust. And in operations, present trust is always more fragile. The real mistake: treating this as a sequence, not a systemThis is not a choice between fixing fast or learning deep. AI has changed the equation. It enables parallel thinking: Immediate stabilization using AI-recommended fixes Simultaneous root cause capture using the same data trail The mistake is not prioritising resolution. The mistake is resolving without capturing learning signals. Example: US Payroll OperationsIn payroll, this trade-off is not theoretical. If a payroll run fails due to a calculation defect or interface breakdown: Immediate resolution restores processing and ensures employees are paid on time Delaying recovery for deeper analysis risks missing bank submission deadlines — a failure that can impact thousands of employees instantly The cost of delay is not just operational — it is reputational and contractual. A late payroll can trigger: employee dissatisfaction at scale manual wire costs SLA penalties exceeding $100K+ per incident In contrast, the same issue — if resolved quickly — can still be: logged automatically by AI, analysed post-run, and permanently fixed before the next cycle Leading payroll organisations do not pause payroll to investigate. They stabilize first, learn immediately after — without compromising delivery. What actually breaks systemsRecurring issues are not caused by fast resolution. They are caused by lack of disciplined follow-through after resolution. Blaming quick fixes for repeat failures is misdirected. The real failure is in governance — not in prioritisation. Final PositionIn any customer-impacting system, time to recovery defines trust. AI gives teams the ability to fix fast and learn deep — but the order still matters. Stability first. Learning immediately after. Fixing fast is not short-term thinking. Failing to fix fast is long-term damage. That is why View A is not just operationally correct — it is the only defensible choice in real-world, high-impact environments.

Position: View B — Keep the feature and fix selectively. Rolling back immediately is an overreaction to a localized problem. Challenging the “fairness” argument directlyView A assumes that a product failing for 8–10% of users is a reason to stop progress for 100%. That sounds equitable — but it is operationally flawed. All systems have edge cases. What matters is not eliminating imperfection instantly, but managing impact without collapsing value. Rolling back a feature that improves experience for 90%+ users does not protect trust — it erodes confidence in product direction. Users don’t just expect stability; they expect continuous improvement. The real risk is not the defect — it is the reactionAI has already done its job: it detected the issue early and precisely. The question is not “Is there a problem?” The question is “Is the problem systemic or segment-specific?” If the issue is limited to older devices or specific usage patterns, then a full rollback is not quality control — it is blunt-force decision making in a precision-enabled system. Example: Payroll & employee self-service platforms (US context)In payroll portals, new features like real-time pay previews or dynamic tax estimators often roll out with AI monitoring. Suppose a new feature improves experience for 90% of employees but causes performance issues for 8–10% users on older browsers. Rolling it back would: degrade experience for the majority, delay innovation cycles, and signal instability in product decisions. Instead, leading payroll platforms: isolate affected cohorts (device/browser-based), deploy targeted patches or feature toggles, and continue delivering value to the majority without disruption. Why? Because in payroll systems, continuity and predictability matter as much as accuracy. Frequent rollbacks create more uncertainty than controlled imperfections. What actually builds trustTrust is not built when nothing goes wrong. Trust is built when users see that: issues are identified quickly, impact is contained intelligently, and improvements continue without disruption. A rollback tells users: “We are not confident in our own system.” A targeted fix tells users: “We understand the problem — and we’re solving it without breaking what works.” Final PositionProgress should not be held hostage by edge cases. AI enables precision. Leadership must respond with precision. Rolling back for 10% sacrifices value for 90%. Fixing selectively protects both. That is why View B is not just efficient — it is the more mature, data-aligned, and trust-preserving decision.

Position: View B — The process should not be stopped immediately when AI predicts a potential defect. A risk-tiered response is more rational than automatic stoppage in probabilistic environments. AI predictions indicate likelihood, not certainty. When a system operates at 85–90% accuracy with a 12% false positive rate, stopping the process at every alert converts statistical risk into guaranteed operational disruption. In high-throughput environments, this leads to avoidable downtime, reduced trust in the system, and ultimately poorer overall performance. The correct control philosophy in predictive systems is not “stop at every signal,” but “respond proportionately to the confidence and impact of the risk.” Example from US payroll operationsThis principle is particularly evident in US payroll, where service continuity is as critical as accuracy. In a typical bi-weekly payroll: AI may generate 2–3 defect alerts per run Each investigation pause can take ~30 minutes Missing a bank submission window can cost $200,000+ in penalties, manual wires, and client escalation Most payroll defects are financially reversible through off-cycle corrections, usually costing $10,000–$30,000 If every AI alert forces a stop, the team repeatedly compresses its processing window and increases the probability of the one failure that is hardest to recover from — late salary deposits. Cost comparisonScenario Impact Type Estimated Annual Exposure Stop on every AI alert Guaranteed labour & schedule compression cost ~$126,000 Tiered response (some defects escape) Correctable payroll errors ~$140,000 One missed payroll deadline due to repeated stoppages Non-reversible service failure $200,000+ per incident The key insight is that payroll errors can be corrected; missed payroll timelines cannot. Therefore, a control strategy that repeatedly risks schedule failure to prevent correctable errors is not economically or operationally optimal. Conclusion AI should influence process decisions, but it should not automatically halt operations when it signals probabilistic risk. A tiered response — where only high-confidence, high-impact predictions trigger stoppage — preserves both quality and continuity. This approach maintains trust in AI systems, protects throughput, and aligns intervention with actual business risk rather than theoretical worst-case scenarios.

Position: View A — Organisations should act on AI-detected burnout signals early. Choosing not to act when risk is visible is not respect for privacy; it is organisational negligence. In most corporate processes, we do not wait for failure before responding to risk. In payroll, finance, and IT operations, we act on early warning indicators — variance reports, anomaly detection, or control breaches — precisely to prevent downstream damage. Yet when similar predictive signals relate to employee wellbeing, organisations suddenly become passive in the name of privacy. This inconsistency reveals the real dilemma: not ethics versus technology, but comfort versus accountability. In payroll and shared services environments, burnout is not abstract. It directly affects data accuracy, statutory compliance, and client trust. An exhausted payroll analyst is more likely to miss tax updates, mis-key compensation changes, or overlook exception reports. By the time the employee admits they are struggling, the organisation may already be dealing with incorrect salary credits, escalations, and reputational risk. Acting only after a human confession is operationally late and ethically reactive. AI systems that analyse patterns such as increased rework, late logins during critical processing windows, and declining communication tone do not create new surveillance — they synthesise signals that organisations already collect. The real risk is not that AI knows too much; it is that leadership might ignore what is already visible. However, early action must be designed as support, not enforcement. A responsible model is: AI flags a risk pattern. HR analytics validates the signal to avoid false alarms. The manager is prompted to conduct a routine check-in framed around workload, not performance. Any intervention focuses on workload balancing, schedule flexibility, or temporary support — never disciplinary review. In this design, AI does not make decisions. It ensures that human leaders do not overlook emerging distress. Critics argue that acting on AI signals can erode trust. In reality, trust erodes faster when employees feel that their struggle was visible in hindsight but ignored until it affected output. Most employees do not resent supportive check-ins; they resent being helped only after they have already burned out. The asymmetry of risk is also important. A false positive leads to an unnecessary but harmless conversation. A false negative can result in mental health decline, payroll failures, and preventable attrition. From a governance perspective, the cost of inaction is materially higher than the cost of cautious early engagement. Therefore, if an organisation has credible predictive capability and chooses not to act, it is not protecting employees — it is protecting itself from difficult conversations. Responsible organisations should use AI to detect burnout early and intervene through transparent, human, and non-punitive support mechanisms.