rohan modak

AI Self audit for Claims Process Optimization I see AI audit as evolution of six sigma thinking, where process becomes capable of detecting, measuring and its own variation. Build a system which is powerful enough to learn but transparent enough to be questioned In Healthcare BPO we frequently encounter AI agents that continuously review claims data and recommend routing logic for claims adjudication process They recommend if claim should be auto adjudicated or flagged for manual review to SME. Even though these agents learn from historical data patterns, it also makes them prone to bias, drift or increasing become in accurate if underlying data changes. To prevent this, I would design a self-loop for AI in three layers 1. Automated AI and drift checks- Periodically, AI should compare with a benchmark dataset- - a sample which has been previously audited by trained SME. If AI ‘s decision pattern starts to diverge beyond the defined threshold, it would generate alert (an internal “Bias alert”) 2. Shadow model validation: A shadow model will run alongside in parallel, generating recommendations that are not acted upon but logged in the database for comparison. This would act like a mirror check – it would highlight performance degradation or data quality issues before they impact real workflows 3. Explain ability for human oversight: Every decision that AI makes should be explainable in simple language- thus ensuring that SME can trace the logic. These explanations would also go to monthly dashboard for Quality & compliance team to review To keep the system accountable and aligned with organization boundaries, I would define below controls: a. Human in loop control- no recommendation should override compliance and policy review without SME review b. Governance of retraining cycles: Only after validation and sign of by Data Governance team and SME, will an AI be allowed to retrain and that too from approved data sources c. Audit traceability: Every AI decision, data source, model version, etc. should be logged in as part of audit trail as a part of accountability

AI with Human Compass- Balancing the Speed and Quality through Context aware intelligence In the world of healthcare BPO – process improvement and automation, one reoccurring conflict is between speed and quality in claims adjudication. Both goals are desirable for the business- faster turnaround improves cost efficiency and client satisfaction while quality of claim processing protects trust and compliance. Yet when we use AI systems for claims auto adjudication, these goals ca easily pull in opposite direction For e.g.: AI might adjudicate 80% if the claims in few seconds, but rushing through complex notes (involving multiple providers or new payer edits) could result in wrong denials. Conversely, being too cautious and routing all the claims for manual intervention negates the very efficiency AI was meant to deliver That’s why we need to implement below principle in AI design: AI with human compass- A framework where AI doesn’t make binary choices but instead used context aware and logic which goals to prioritize. Below is how it will work: · For low risk, rule-based claims, AI will prioritize speed. driving costs and time savings · For high-risk ambiguous claims, AI will escalate for human intervention, ensuring compliance and accuracy This layered decision logic is not static, but risk adaptive. AI continuously learns from past interventions updating its confidence scores and thresholds. AI should be guided by human values and judgment frameworks, not just data driven optimization. AI can be trained to recognize “risk zones” where human oversight is essential, just like copilot alerting the pilot when conditions are uncertain. Human oversight plays a guiding role in this ecosystem, not as a supervisor but a moral and contextual governor. Human examiners will define trade-off zones, review exceptions, and retrain the AI where Payer logic and policy nuances arise Thus, AI with Human Partnership creates a partnership where: · AI drives scalability · Human examiners anchor judgement, empathy, trust This intelligent, self-improving balance between performance and prudence- An AI that doesn’t just process faster but decides smarter

Symbiotic Intelligence Loop: Where Ai and Human examiners help each other to correct their blind spots In healthcare BPO setup, we see AI and human examiners make different but complementary mistakes, for e.g. in pend resolution and claim adjudication processes. I think future lies in creating Symbiotic Intelligence Loop, where continuous learning occurs from each other Mistakes that AI makes and where human examiners can help it improve · AI may flag a claim as “Provider mismatch” because it may not know that new provider ID was added last week. Human examiner will look at the context, correct the classification and add a note. AI will learn this pattern and stop flagging false pends from next time ( false positive) · AI may auto deny claims that looks same as past denials, even if the policy rules have changed. Human examiner will override such cases and feed updated rule in AI. AI will realign with the new payer logic ( false positive) · Free text notes including comments, one-word instructions, or human annotations (e.g.: “confirmed via email “are many times interpreted incorrect by AI. Human examiner gives structured feedback. AI learns to tag these patterns as valid justifications ( false negative) · AI struggles interpreting exception scenarios (for e.g.: duplicate claims linked to two different provider NPI). Human examiner in this case will provide valuable training inputs which will help AI to broaden its exception logic ( false negative) Thus, with human examiners as contextual partners, AI can become more adaptive, context aware and trustworthy Mistakes that Human examiners make where AI can help improve · Human examiner may miss trends across hundreds of claims that get processed every day for single facility. AI can help detect data patterns, do better analysis- create dashboards and correlation charts helping human examiners arrive at exact root cause · Different human examiners may pend different claims differently. AI can help standardize decision rules, flag outliers for review making the process consistent · Human examiners have bandwidth issues and limited attention span which causes dip in accuracy in case of sudden volume spikes. AI can pre screen the claims, auto process straightforward logics so that examiners can only focus on claims steps needing true human judgement · Human examiners rely on periodic audits to identify process gaps for improvement. AI can give real-time feedback and predictive alerts helping examiners to correct issues proactively This way, with AI as analytical partner, examiners can start making faster, data driven, more accurate decisions In essence, this “Symbiotic Intelligence Loop” is not about humans training machines or machine replacing humans but both evolving together to create a faster, adaptive and smarter claims ecosystem

Let’s answer this question taking example of healthcare claims adjudication process. HealthCare BPO Associate receives the claims from the Payer where decision has to be made to accept or reject the claim In Claims adjudication process, AI can easily handle 90% of work – most of which is repetitive. Tasks include scan the claim form, basic validation, duplicate checks, COB checks, verifying CPT/ ICD codes, checking if any exception code has been triggered, and flagging any rule violations using configuration logic defined by Payer However, the final decision to release the claim (approve or deny) should always rest with human examiner, especially in case if complex claims, where medical necessity is questionable and patient circumstances don’t always fit the rulebook Such cases demand empathy and ethical judgement -reading between the lines and understanding doctors’ intent. The decision should always reflect balancing compliance with compassion. Since algorithms are trained purely on data patterns, such nuances cannot be expected to be executed by AI accurately A practical AI Precision Assistant – Human collaboration model can be designed in this example where AI can do the heavy lifting 1. Claim intake and basic validation can be done by AI agent using OCR+NLP models to extract and validate member, provider and service fields 2. Rule based checks and duplicate checks which involved applying payer policies, checking if claims fall between designated coverage limits, and assessing historical claims data – all these can be handled by AI agent 3. Predictive scoring: AI agent can rank the claims by likelihood of approval/denial and can assign confidence score to it 4. If confidence score assigned by AI model is <85% or if patient context is not clear, then these claims will get routed to human examiner who will evaluate intent, context and fairness before making the final call This clear demarcation allows AI to work as Precision assistant while human examiner will retain the ethical and strategic accountability. This clear demarcation between AI Agent and Human examiner is required and human should make final decision because: · We are not just processing claims, but impacting person’s health and financial well being · Only human examiner can recognize when a rule should not override compassion · Only human can bring in unique blend of empathy, sound judgement and ethical awareness

I will explain with provider credentialing example in healthcare domain. In my domain, process SLA (namely TAT and FTQ%), are single most important metrics that we cannot afford miss. Here, we face two major challenges a. Sudden Volume Surges: due to new launches, seasonal hiring by big provider group b. Policy shifts: on account of Payer specific primary source verification changes, NCQA reverification windows We have to be very wary of our SLA, and cannot risk penalties so advance scenario planning is critical My solution design would comprise of creating Credentialing Digital Twin - a safe sandbox to ask “what-if” and see operations react. Below are the features of this tool 1. Demand model and mix driver: Basis past submissions, go live calendars; create time series by payer, state, specialty. Then add a classifier to simulate denial mix shifts 2. Policy as a code engine: Encode each NCQA/payer rule as logic (for e.g.: FL licenses require PRN verification step + Address proof). We can further flip a toggle to model new requirements to see extra verification steps and expected deficiencies 3. NLP Deficiency predictor: A small NLP model will scan attestation text along with any attachments to predict missing/invalid items (DEA, Board cert, etc.) 4. Capacity and Skill routing: Models skill mix, learning curves and shift patterns 5. Monte Carlo Simulation – will give outputs (Distributions for TAT, FTQ%; and FTE gap) that Operations can action on 6. Generate Panel Readiness Risk Score This approach is one step beyond basic forecasting, as 1. We are simulating flow of work as policies, skills, deficiencies interact 2. Using Policy as code feature, we can quick toggle changes proposed by payer/ NCQA which will prevent tribal debates 3. This will prevent back and forth outreach as we now have capability of front load fixes on day 0 4. Backlog trajectory can be defined and notified to leadership beforehand 5. FTE delta and recommendations to redeploy or cross train can be notified to leadership So, AI is not replacing judgement, but rather gives clear actionable plan to tackle the upcoming volume surges or policy shifts. This can turn into huge value addition for Healthcare Credentialing Shop as it can prevent costly escalations.

There are many examples we can find in Healthcare BPO domain where same process often looks different across regions and clients. If we look at Provider Credentialing process, we prominently see below challenges: Cultural habits: Teams across various geographical regions rely on implicit/intuitive knowledge for eg: “They always pend claims for XYZ group” Regulatory and payer rules: Decisions and actions taken are region specific - based on state regulations, plan specific edits, different network rules Local platforms and codes: CRM, Code lists, escalation paths vary according to regions and geographies If we insist everyone follows same standard framework globally, we risk slowing down progress and increase rework. On the other hand, if we give leeway to everyone to follow their own localized processes, it will cause confusion which may lead to quality and reporting issues Our design approach for AI is to create Smart Credentialing Assistant comprising of Core Global Spine+ Local adapters = standardize the decision model along with logics globally while allowing AI to apply region specific policies real-time For e g: we apply same framework to provider credentialing process – which involves verification and enrolment of providers and facilities Global core Spine- (standardized globally) a. Common taxonomy: Credentialing types, adverse findings categories, decision outcomes categories (approve, deny, pend, conditional), pend reasons remain same globally b. Measurable CTQs implemented universally: First time right (%), Rework Rate (%), PSV Completeness % c. Notes templates to be standardized globally: verified by, description, timestamp, result, adverse flags, next action Local adapters (flexible region wise/payer wise) a. State rules: Licence types, renewal cadence, state specific policies, etc b. Payer Network specific: attestation clauses, plan participation criteria, CDS requirements, etc c. Local SLA calendar: state specific appeal windows, regional holidays, state specific blackout periods, etc d. Language variants: Accepted proof types, notarization needs, state specific EN/ES templates AI design would include Policy as a code engine (global) with region specific credentialing policy packs (Plugin) with below prominent features 1. Retrieval augmented provider credentialing: we will augment the capabilities of chat gpt by adding information retrieval step incorporating proprietary enterprise content for answer formulation. This would allow us to fully constrain the LLM to enterprise content a. If we input provider NPI, state, specialty, payer, LLM will retrieve exact micro-SOP and accepted sources (state board URL, phone script) to prepare a customized verification checklist b. LLM will scan global knowledge base (National Practitioner Database (NPTB), Office of inspector general (OIG) and add local flavour (state specific policies for California Medical Board c. ChatGPT will then purpose next action depending on the specific scenarios for e.g.: – “Verify the malpractice coverage period> 12 months for specific payer and draft notes using standardized template 2. Guardrails and explainability: Every AI suggestion will have clear logic explained with action recommended summary. If local SME disagrees there will be an option to override decision with reason. These use cases will be used to further improve the model 3. Human in loop: For high-risk scenarios, LLM will ask 1-2 targeted questions and route to Clinical Credentialing Committee queue Smart Credentialing assistant will provide Credentialing associates with below User Interface: · Smart checklist – customized to specific scenario/ context · On click verifications – respective state boards, auto screenshot capture, · AI drafted notes · Risk bar- Green/Amber/Red for fast decision making · Explainability tray- to display fired rules, sources, etc · Override capture- with reason codes · Audit trail · Dashboards with relevant metrics

In my healthcare BPO claims support , one of the biggest challenges is helping new associates is during new joiner training build confidence in understanding process and handling complex claim scenarios. Traditional training is not effective mainly because of limited bandwidth of trainers, generic learning modules, and delayed feedback, when trainees usually get after errors have already occurred. In addition to this, there is always added pressure to learn the process fast and start performing. AI can make a real difference — by acting as a Supportive coach. For example, an AI agent could listen to how an associate interacts with provider calls and then gives supportive tips to trainee like, “Pause for 2 seconds before responding will help to calm down the agitated customer.” Or at the end of the week, the AI could provide a summary: “You’ve improved your accuracy on validating the claim intake forms by 15%, now let’s focus next on speeding up benefit configuration checks. Here is a link to few short practice exercises which will help you to get up to speed". This way, AI helps encourage the trainee and also points out the next area of improvement. Training thus becomes continuous, personalized and aligned with each associate’s real work On the backend, AI can aggregate trainee performance data and share with managers — capturing how individuals respond to guidance, the pace of their improvement, and the specific skills they’re measuring and improving. AI can then translate this into simple dashboards for managers, showing each employee’s learning curve alongside benchmarks from top performers. The value here is twofold: managers get clear visibility into progress without sifting through raw data, and they can step in with targeted coaching or corrective action only where it’s needed. This shifts management from reactive oversight to proactive, data-driven support, ensuring every employee moves forward with confidence and at the right pace. The key is making sure AI feels less like a taskmaster and more like a supportive friend who wants you to succeed. To achieve this, I’d design it with empathy at the core: Guiding, not penalizing: Feedback is framed as gentle suggestions or learning opportunities, never as penalties. Giving control Employees stay in control — they can accept, snooze, or revisit prompts when they feel ready, just like a friend reminding you at the right time. Coaching with empathy: Progress is shown through benchmarks, not in comparison with fellow trainees, so employees see how they’re improving in a safe, encouraging way rather than feeling compared or judged. This way, AI becomes a companion in growth — always patient, supportive, and focused on helping employees build confidence, not fear. This, the balance is about empowerment over oversight: AI should reduce fear of mistakes and build employee confidence, not replace human judgment or make them feel watched. Done right, it can create a culture where every associate feels like they have a coach in their corner — always available, always patient, and always focused on their growth.

When I think about Business Transformation in U.S. healthcare domain — the first instinct is to use AI for process optimization: faster auto-adjudication, fewer pend reasons, reduced AHT, cleaner QA. But as I think deeper, the bigger opportunity is using AI as a spark for innovation, not just efficiency. “As a team, we can use AI not only as a tool for efficiency but as a co-innovator — one that can spark fresh ideas we may not see on our own. In these ideation sessions, we’ll explore how AI can inspire new service models, smarter triage approaches, and innovative ways to elevate the provider and member experience.” For e.g.: In claims triaging, we route claims based on product, different US state specific rules, or high dollar thresholds. However, AI could instead scan historical closures, denial patterns, and even provider behavior trends to proactively suggest new triage categories we never thought of earlier. For e.g.: today we route claims based on age of claim (e.g., >30 days old = priority queue). However, AI could scan historical pend and touchpoint data to find patterns like “Low-dollar outpatient lab claims from new providers often pend multiple times because of missing NPI or taxonomy mismatches.” Or “even though they’re low dollar, they consume a lot of rework effort and delay claims closure rates”. Here the impact is that what was previously ignored as “low priority” (low-dollar claims) is reframed as high process impact because of rework. AI surfaces that hidden cost and recommends a smarter triage approach. Another example is better customer experience: In provider support, agents spend hours explaining rejections to providers. What if AI generated short explainer videos or interactive claim-walkthroughs for common denial codes? That’s not just efficiency; it reimagines the customer experience. Instead of sending a PDF, what if a provider clicks a link and sees a 90-second AI video: “Here’s why this claim was denied, here are 3 corrective steps, and here’s how to avoid it next time.” In summary, I see AI as a co-innovator in healthcare domain: helping us not just work smarter but imagine service models and customer experiences we’ve never attempted. Our role is to keep its creativity grounded in compliance, operations, and value.

I work in a healthcare BPO where we have separate teams handle who everything from claims adjudication to enrolment to provider support. Honestly, scheduling has always been one of our biggest pain points. Some days we’re drowning in work, other times half the team is sitting idle. Managers spend hours reshuffling rosters, and associates often feel stuck with unfair shifts. When I imagine AI stepping in, I see a big change. Instead of static schedules, AI analysing historical data and interpreting incoming work patterns — for e.g.: knowing that Medicare claims spike after weekends, or enrolment volumes shoot up during open enrolment. It could automatically shift resources, balance workloads, and even suggest cross-training so it will help team members grow their skills. From my side as an employee, I’d feel more respected if the system recognized my preferences — like not always putting me on late nights, or letting me swap shifts without making my manager chase approvals. That kind of fairness builds trust. For the business, it means claims don’t pile up, SLAs aren’t at risk, and patients/providers aren’t waiting endlessly for resolutions. At the end of the day, AI scheduling isn’t about robots replacing us — it’s about making sure the right person is on the right task at the right time, while still letting us have a good work life balance I wouldn’t want AI Agent to just shuffle names on a roster based on volume spikes. For me, the real value would come when AI looks at human, operational, and compliance layers together. Yes, it should definitely factor in basics like forecasted claim inflows, call volumes, and shift hours. But beyond that, I’d want it to consider: Skill depth vs. task criticality – who is faster at processing denials, who is certified for provider credentialing, and who is still learning. Matching skill with work type can raise accuracy and reduce rework. Employee well-being – things like not giving someone three-night shifts in a row, respecting time-off requests, or balancing high-intensity queues with lighter work. Burnout is real in BPO. Fairness & Preference: - Many of my team members have personal constraints — parents needing evening shifts, or learners preferring night hours. AI should balance fairness: not always giving the “easy shift” to the same people but distributing opportunities equitably. Real-time Dynamics- If claim volumes spike unexpectedly due to a provider update or a new enrollment cycle, AI should instantly recommend reallocation — not tomorrow, but now. Creatively, I see AI not just as a scheduler but as a “workforce well-being coach.” It could predict burnout risk from overtime trends, recommend micro-breaks during high cognitive load tasks, and even gamify flexibility by awarding credits for accommodating tough shifts. That way, scheduling becomes a lever for engagement and growth — not just an operational necessity.

Imagine in a healthcare BPO environment, we suddenly see a spike in pending claims and leadership needs to decide fast — do we pull people from another process which will put strain on exiting workflows, or do we push for overtime, which will eventually hit project bottom-line? Each decision will have pros and cons. Pulling people from other processes means putting strain on existing teams and approving Normally, that mean lots of manual efforts involving hours of pulling reports, debating in meetings, and still not having the full picture. This is where I’d see an AI agent as a co-pilot – someone who will do the hard work and compile the necessary information from various places. It could instantly scan the claims backlog, existing workforce availability, and compliance rules, then frame the decision for me. For example, it might say: ‘Option A — move 20 people from Membership tower to claims tower, closure rate improves by 4% but average handle time rises 2%. Option B — authorize overtime, closure improves 3% with no impact on handle time.’ Now instead of me digging for data, I’m looking at clear trade-offs. But I wouldn’t want the AI making the call. I’d build in checks — it has to show me its data sources, flag any blind spots, and tie its recommendations back to our bigger goals like compliance and customer satisfaction. And I’d always want it to offer multiple options, not just one, because the judgment is still mine to make. This way, I ensure the final decision is not just data-driven, but also in line with our company’s culture, values, long-term strategy, and the real-world factors that AI can’t fully capture.” So, in my mind, AI doesn’t replace leadership — it accelerates the analysis, but the responsibility to weigh culture, strategy, and people still rests with me.”

Below is how I will manage versions of AI flows and prompts in a claims processing scenario, where things are constantly evolving based on feedback from claim examiners, auditors, and compliance. 1. Keep Track of Changes While building claims-processing AI assistant, the prompt that guided the “claims eligibility check” step worked… but only for the first few weeks. Then, business rules changed, compliance flagged some outputs, and examiners started giving us feedback. Instead of editing the prompt and hoping for the best, I store every single version of my flows and prompts in a company GIT repository Each branch is new iteration — for example, feature-improve-prior-auth-check. I clearly document why I made the change: When I deploy a new version, I tag it in GIT and log that version ID in our monitoring dashboard, so when a claim examiner says, “The bot did not process a specific scenario,” I can instantly see which version they were using. 2. Documenting the Story Behind the Change Clearly document story behind the change in order to delineate why I made that particular change v2.1.2 — 2025-08-15 Change: Updated “denial reason explanation” prompt to include ICD-10 lookup when code not in local cache. Why: Several claim examiners escalated cases because the bot said “code not found,” even though it existed in the database. Expected Impact: Reduce “code not found” errors by 20%. This makes it easy for me to tell the story of the bot’s improvement over time 3. Testing Before I Roll Out I never just push changes live. In claims processing, one wrong rule application can delay thousands of claims. Below are few things I follow Shadow Testing: I run the old and new prompts side-by-side on 100 recent real claims (with PHI data masked). Regression Suite: I maintain a set of tricky test cases — like coordination-of-benefits disputes or secondary insurance retro adjustments — to make sure the new version doesn’t break things that used to work. SME Review: I share sample outputs with our senior claim SME for human- in loop- scoring. They tell me if the new explanation is actually clearer or just longer. 4. Metrics tracking and feedback from team After Deployment Once the new version goes live (usually to 10% of examiners first), I: Track auto-adjudication accuracy — if it dips, I know something’s off. Collect feedback tied to the exact version. Categorize any errors: prompt misunderstanding, missing data, or wrong business logic. This way, I don’t just hear “the bot is processing incorrectly” — I know why. 5. Protecting Against New Problems I’ve learned the hard way: never delete a working version. I keep the last stable prompt ready so if my experiment tanks, I can roll back in minutes. In claims processing world , the cost of a bad AI update is delayed payments, or regulatory fines or angry providers - un term seriously impact customer satisfaction By treating flows and prompts like living assets with a documented history, I never lose track of why something changed, and I can always prove whether the change actually helped. It’s not just version control — it’s trust control.

This is very important in a highly regulated and data sensitive domain like healthcare. I would include below features in an AI agent to ensure focus, efficiency and answer with precision 1. Prompting: Clear Role and context setting early in the conversation to avoid tangents- this will ensure the agent anchors to specific role. Example of such Prompt Framing could be : “Hey there! I’m your virtual assistant here to help you escalate claims issues quickly and accurately. I’ll ask a few questions to get to the bottom of the problem and guide you to the right solution or team. Let’s get started.” 2. Flow Constraints: Implement Decision Trees and Guardrails to Stop long-winded answers or unnecessary diversions. Agent should Create step-by-step flows based on known scenarios (for instance, if there’s a denial reason, check if documentation is missing, then escalate). Also, agent should offer limited options instead of open-ended questions where clarity is required. For Example, agent can ask: “Is the claim denial related to one of these and wait for user response?” · Authorization issue · Eligibility · Coding error · Not sure This makes sure to keeps both the user and AI focused on a clear path and makes it easier to map backend logic and escalation rules 3. Checkpoints: Agent should confirm understanding before moving forward to avoid misunderstandings and backtracking For eg,“Just to confirm — the claim was denied for CO-109 (missing documents), and you’d like to escalate this because you did submit supporting files. Is that correct?” AI agent should use yes/no confirmations along with summary feedback. This will ensure agent is aligned with the user and automatically cuts down on misdirected escalations or wrong team routing 4. Clarifying Questions: Agent should only Ask clarifying questions when Necessary and keep them specific Good AI question would be “Do you have the reference number or denial code?” instead of “Can you explain everything for the particular claim” 5. AI Agent should mention audit trail and explain the next steps. This is to ensure user knows why a step is necessary or who’s going to take care of it.. For eg: Agent should respond like” “This needs to go to the Membership Correction team because the provider ID seems to be mismatched. They typically resolve it within 48 hours.”

Designing prompts is an important and critical step in getting most out of AI large language-based tools. For example, healthcare BPO claims process, let’s consider a scenario where we want to LLM model to review the claim which was denied. We are expecting the LLM model/ AI agent to detail the reason for claim denial, suggesting the next appropriate action and AI agent to execute the action (route the claim to concerned team if requires) Initial Prompt design: ““Explain this claim denial and next action to be performed”. AI model may give a generic explanation like: “Claim is denied with CO-22 denial code meaning the claim is not covered by the current payer. You should escalate it.” While this response is correct, it is not clear exact action that agent has to take which may lead to misrouting or unnecessary escalation, and this does not inspire trust in the LLM reliability Improved Prompt Design: “The claim has been denied with code CO-109. Explain in 1-2 sentences what this denial means, identify if this is related to member eligibility or provider mismatch, and suggest the most appropriate internal team to route the claim to basis historical resolution logic.” LLM output basis improved prompt: The AI gives a more accurate, trustworthy, and actionable response: “Denial code CO-22 indicates another health insurance payer may be primarily responsible for covering the service. The current payer believes they are not the primary insurance, and payment responsibility may lie with another insurer. Based on historical data, this should be routed to the Coordination of Benefits (COB) / Member Eligibility Verification team to verify COB. Next actions to be performed are: a) Review member’s COB file in the system. b) Check for existence of other insurance coverage (via client portals, IVR, or eligibility tools). c) Update COB order if incorrect. d) Coordinate with member or client for COB attestation if required. e) Route the claim back to processing once correct COB details are confirmed Now this is a detailed LLM output which AI agent can action with precision and high accuracy. This way, even in a highly regulated environment such as healthcare BPO, a small change in the prompt wording can transform a AI output from ambiguous to precise

Examples of different slices of cheese meaning different defence layers (checkpoints, controls) include member eligibility verification, claims intake checks, auto adjudication rules engine, QC checks Holes in system include- outdated feed form leads to incorrect member eligibility status, OCR incorrectly read intake scanned forms , outdated policy logic in auto adjudication rules in engine, incorrect sampling in audits By visualizing healthcare claims processing workflow through the Swiss Cheese Model, we can gain clarity on: Where the defences exist, Where they are weak, And how errors escape intense scrutiny of reviewers Integrating Business Excellence principles ensures those “holes” are proactively identified and closed, creating a more reliable, efficient, and error-proof claims process.

We go ahead with process improvement when we have a stable but inefficient process, where we want to use lean/six sigma techniques to improve these processes we use lean six sigma when we can identify gaps using traditional tools such as RCA, SIPOC, HLPM,, FMEA, FPY, Pareto, etc which can yield us incremental process optimization We reimagine processes with AI when 1. Budget and client tech approval for AI implementation are available 2. We have exhausted gains from lean six sigma initiatives and traditional improvement techniques 3. Decision making steps which cannot be done by RPA ( pattern recognition/prediction etc.) 3.We identify potential to automate judgement based cognitive work which can be done by AI