Redesign the claims process around an agent — not by bolting AI onto the old map, but with an autonomy tier on every step, exception queues and escalation SLAs as first-class branches, a change-impact summary, and a measurement plan with baseline and kill criteria. It must survive an SME walkthrough and a finance review.
Everything in the track converges here. Divya's validated as-is map (Lab 04), the opportunity scan with its refusals (Lab 10), the honest business case (Lab 11), the data-readiness prerequisites (Lab 12), the eval harness (Lab 13), and the AI-requirements patterns (Lab 14) — Meridian General Insurance's steering committee has approved designing the agent-era claims intake process. The deliverable is a blueprint: the process as it should run when an agent does real work inside it, with human authority engineered in rather than promised.
AS-IS (validated, condensed) - 9,000 motor claims/month
S1 FNOL capture -> S2 policy validation -> S3 registration ->
S4 doc completeness check -> [34% incomplete -> S5 chase queue,
6.2-day avg] / [4% fraud signals -> S6 SIU triage] ->
S7 desk review (15% kickback loop) -> S8 surveyor assignment ->
S9 estimate review -> S10 payment approval (>1L human) -> S11 pay
-> S12 status comms (11,000 "where is my claim" contacts/mo)
APPROVED TREATMENTS (Lab 10) & CONSTRAINTS
T1 S3 eliminated via FNOL-ClaimsCore integration
T2 S4/S5 redesigned: dynamic doc checklist at FNOL; AI extraction
and completeness checking at point of capture
T3 S10 stays a human control above threshold (finance mandate)
T4 SIU referral stays human-decided; AI may flag, never disposition
T5 Customer-facing AI-drafted comms: allowed with Art.50 disclosure,
human review until eval evidence supports relaxing (Lab 14)
BASELINE (Lab 11, measured) - for the measurement plan
B1 S4 median 8.0 min/claim B2 incomplete-on-arrival 34%
B3 chase wait 6.2 days B4 downstream extraction errors 3.1%
B5 status contacts 11,000/mo B6 intake staff cost INR 480/hr
OVERSIGHT STANDARD (steering-committee minute, verbatim)
O1 "Every agent-performed step carries a declared autonomy tier:
SUGGEST (agent proposes, human acts), APPROVE-THEN-ACT (agent
prepares, human approves, agent executes), ACT-AND-LOG (agent
executes, logged, sampled audit). Tier assignments require a
written rationale and a demotion trigger."
Assign an autonomy tier to every to-be step before drawing anything. Build the tier table: Step | Agent role | Tier (SUGGEST / APPROVE-THEN-ACT / ACT-AND-LOG / NONE-human) | Rationale | Demotion trigger. Work from the treatments: extraction and completeness checks can start ACT-AND-LOG (low harm, reversible, eval-covered); customer-facing drafts start SUGGEST per T5; payment release above threshold is APPROVE-THEN-ACT at most, honoring T3; SIU disposition is NONE per T4. Every ACT-AND-LOG rationale must name why errors are reversible and how the sampled audit catches them. Demotion triggers are pre-written: "IF weekly eval accuracy < X or sampled-audit error rate > Y, step drops one tier automatically."
Draw the to-be map with tiers on the diagram. Mermaid or draw.io: one lane or color per tier, tier stamped in each node label (e.g., Doc extraction [ACT-AND-LOG]). The map must show the redesign, not the old map annotated: S3 gone (T1), completeness checking moved into FNOL (T2), the chase queue shrunk to the residual incomplete cases. Prompt the AI for a first Mermaid draft from your tier table — then correct it with Lab 04 discipline; regeneration reintroduces deleted steps, and this time you know to look.
Make exception queues first-class branches, not footnotes. Draw each as a named queue with an owner and an entry condition on the arrow: low-confidence extraction queue (intake, threshold from the eval harness), suspected-injection/anomalous-input queue (security + intake — the Lab 14 lesson applied to claims documents), SIU referral queue (SIU, T4), agent-outage fallback (all steps revert to manual routing — the process must be drawable with the agent switched off). If a branch only exists in prose, it does not exist; it must be a drawn path with somewhere to go.
Attach escalation SLAs to every queue. Per queue: maximum dwell time, escalation target, and the breach action — e.g., low-confidence queue: 4 working hours to human verdict, breach pages the intake lead; SIU triage: 48h per the Lab 04 walkthrough; outage fallback: manual capacity plan invoked at 2 hours. An exception queue without an SLA is where the old 6.2-day wait will quietly reincarnate — the SLA is the requirement that stops the new process from growing the old process's tumor.
Write the change-impact summary — whose work changes how. A table per affected role: Role | Tasks removed | Tasks added | New skills needed | Trust barrier | Verification duty. Intake staff: less typing and checking, more exception adjudication and audit sampling — a harder job, not a smaller one; say so. Team leads: new kill-switch authority and sampling duties. Assessors: receive structured extractions — their trust barrier is "why should I believe the agent's field values," answered by visible confidence scores and the audit trail. SIU: unchanged authority, new AI-flag input to triage. This page is what makes the blueprint adoptable instead of merely correct.
Build the measurement plan on the captured baseline. For each redesign claim, a metric pair: baseline (B-line) → target with date. Minimum set: incomplete-on-arrival rate (B2 → target), residual chase dwell (B3 →), extraction error rate (B4 →, measured by the Lab 13 harness weekly), status contacts (B5 →, expected to fall as delays fall), S4-equivalent handling time (B1 →), and cost per claim (derived, B6). State the measurement mechanism and cadence for each — a target without a mechanism is a hope.
Write the kill criteria and the demotion ladder. Blueprint-level kill criteria (Lab 11 discipline, now for the rollout): e.g., extraction accuracy below the eval floor for two consecutive weeks → all extraction steps demote to SUGGEST; any agent-caused wrongful payment → immediate APPROVE-THEN-ACT freeze pending review; customer-complaint rate about AI comms above threshold → drafts return to full human review. Plus the standing kill-switch: who can pull it, how fast, and what the process looks like the minute after (your step-3 outage path answers this — cross-reference it).
Run the two survival reviews. First, the SME walkthrough (Lab 04 script style) against your to-be map: trace a clean claim, an incomplete claim, a fraud-flagged claim, an injection-suspect document, and an agent-outage hour — every token must have a path, an owner, and an SLA. Second, the finance review: give the AI your measurement plan and one-pager with "You are a skeptical CFO analyst: which target lacks a baseline, which benefit lacks a mechanism, which cost is missing?" — adjudicate and patch. Then assemble the final deliverable: the blueprint (map + tier table + queues/SLAs) and the adoption-and-measurement one-pager (change impact + metrics + kill criteria).
Autonomy is not a property the agent has — it is a permission the process grants, per step, in writing, with a pre-agreed way to take it back.
A complete lab has all of the following:
An agent-era process is designed in permissions, not features — a tier on every step, a queue for every exception, an SLA on every queue, a human who can see, correct, and stop, and a measurement plan that can kill the whole thing. Draw that, and both the SME and the CFO will sign.