Walk a validated claims process step by step and assign each a treatment — eliminate, rules automation, AI-assist, AI-agent, or keep human — with rationale that survives challenge. Three traps are planted: two steps where layering AI onto a broken handoff is the seductive wrong answer, and one "automation candidate" that is actually a control.
Divya Nair's validated motor-claims map from Lab 04 made her the obvious choice when Meridian General Insurance's COO asked the question every operations leader is asking this year: "Where should AI go in claims?" The wrong way to answer is to list steps that could take an AI. The right way — the discipline this lab drills — is a per-step treatment scan where "no AI here" and "delete this step entirely" are first-class answers, followed by a value-feasibility-risk ranking that tells the COO what to do first and what to refuse.
Twelve steps, walkthrough-validated (Lab 04), now enriched with volumes and error rates from the ops team.
CLAIMS PROCESS — OPERATING DATA (9,000 motor claims/month)
S1 FNOL capture (app/call/garage) 12 min | 6% capture errors
S2 Policy & cover validation 70% auto today; 30% manual, 4 min
S3 Registration in ClaimsCore 5 min | manual re-key of the FNOL
form into ClaimsCore; 3% key errors
S4 Document completeness check 8 min | 34% of claims incomplete
S5 Customer doc chase (pending queue) 3 chase attempts; claims wait
6.2 days average in this queue
S6 SIU fraud triage 4% of claims | 25 min | SIU team
S7 Assessor desk review 60% of claims | 15 min | 15% kicked
back to intake (rework loop)
S8 Surveyor assignment 3 min | rule-based: postcode + load
S9 Estimate review 30 min | 20% of garage estimates
are handwritten/scanned
S10 Payment approval 6 min review | team lead approves
every payout > INR 1,00,000;
queue adds ~1.5 days
S11 Payment execution batch, nightly, fully systematic
S12 Claim-status communications 11,000 inbound "where is my
claim" contacts/month, 7 min each
Fix the treatment vocabulary first. Five treatments, strict definitions in your sheet header: eliminate (the step exists only because of a broken structure — remove the cause, delete the step), rules automation (deterministic logic, no model needed), AI-assist (model drafts/extracts/flags, human decides), AI-agent (model acts, human supervises by exception), keep human (judgment, accountability, or control — AI does not belong in the decision). Order matters: eliminate beats automate — automating a step that should not exist is polishing waste.
Scan all twelve steps yourself before consulting AI. For each: treatment + two-line rationale + the operating number that justifies it. Force yourself to ask, at every step: "Why does this step exist at all?" — at least two steps in this process exist purely because something upstream or sideways is broken. Thirty minutes; this sheet is the graded artifact.
Run the AI scan as a second opinion. Prompt: "For each step in this claims process, assign one treatment: eliminate / rules automation / AI-assist / AI-agent / keep human. Rules: prefer eliminate where the step compensates for a structural problem; never assign AI-agent to a step that exists only to move work between systems or queues; flag any step that functions as a financial or fraud control. Justify each with the operating data." Compare against your scan. Every disagreement gets adjudicated in a Challenge column — data decides, not confidence.
Hunt the layering anti-pattern explicitly. The most common AI-adoption failure is bolting a model onto a process shaped by pre-AI constraints — an agent that writes faster emails into a queue that should not exist, an agent that re-keys between systems that should be integrated. Re-examine every step where you or the AI wrote "AI-assist" or "AI-agent" and ask: is the step itself the defect? Expect to reverse at least two verdicts.
Hunt the control masquerading as a bottleneck. Somewhere in this process, a step that adds delay and looks ripe for automation is actually there to stop money or fraud going out the door. Automating a control is not efficiency — it is removing a safeguard and calling it throughput. Find it, mark it, and write what would be a legitimate improvement for it (threshold tuning, risk-based sampling) versus what is off the table (removing the human sign-off wholesale).
Build the value × feasibility × risk matrix. For every step whose treatment involves change: score value 1–5 (minutes × volume × error impact — compute from the data, show the arithmetic), feasibility 1–5 (data availability, rule clarity, model maturity — handwritten inputs cost feasibility points), risk 1–5 inverted (customer harm, money movement, compliance exposure). Rank by value × feasibility × risk-inverse. Ask the AI to check your arithmetic; verify its check.
Write the backlog and the refusals. The AI-opportunity backlog: top 3–5 ranked items, each with treatment, expected impact grounded in the operating data, and first validation step. Then the section most scans omit — refused items: the layering traps and the control, each with one line on why the obvious AI answer is wrong. The refusals are what make the backlog credible.
Draft the redesign-scope recommendation. Three sentences to the COO: which part of the process should be redesigned rather than augmented (name the steps), which single AI-assist opportunity to pilot first and why it wins on the matrix, and which step must keep its human regardless of pressure. This paragraph is the lab's executive deliverable.
The step that most wants an AI agent bolted onto it is usually the step that should not survive the redesign — automate nothing you have not first tried to eliminate.
A complete lab has all of the following:
The opportunity scan's power is its refusals — eliminate before you automate, keep controls human no matter how bottleneck-shaped they look, and never let an agent fossilise a handoff that integration should have killed. The backlog is what you build; the refusals are why anyone should trust it.