Turn a messy claims-intake narrative into a BPMN draft with AI, then run a structured SME walkthrough against the real process. The draft merges two exception paths, invents a step nobody performs, and misses a rework loop — the walkthrough script is designed to surface all three before the map gets a validation stamp.
Divya Nair is a systems analyst at Meridian General Insurance, a fictional mid-size insurer, assigned to document motor-claims intake (first notice of loss through to surveyor assignment) before any improvement work starts. The only documentation is an email from Arun, the claims operations SME, written in a hurry between escalations. Divya's plan: let AI produce the first BPMN draft from Arun's narrative, then validate it with a disciplined walkthrough — because a process map that has not survived an SME walkthrough is a diagram, not a fact.
Arun's email, verbatim. Messy on purpose — this is what process discovery actually looks like.
From: Arun (Claims Ops) — "how motor intake works, roughly"
Customer reports a claim by app, call centre, or garage partner.
Whoever takes it captures policy number, incident date, description,
photos if they have them. The intake team checks the policy is active
and the cover type matches the loss — maybe 8% fail this and get
a rejection letter. Then someone registers the claim in ClaimsCore
and it gets a claim number, customer gets an SMS.
If documents are missing (police report for theft, driving licence,
FIR for third-party) we park it in a pending queue and chase the
customer. That is easily a third of claims. Separately — and please
keep this separate — if intake spots fraud signals (late reporting,
photo metadata mismatch, repeat claimant) it goes to the SIU referral
queue. SIU cases must NOT sit in the normal pending queue, compliance
was very clear after last year's audit.
Once docs are complete, claims under 50k with clean signals go
straight-through to surveyor assignment. Everything else gets a desk
review by an assessor first. Assessors kick back maybe 15% of
registrations to intake for re-capture — wrong cover code, bad
incident date, that kind of thing — and intake fixes and resubmits.
Then a surveyor is assigned by postcode and workload, customer gets
the surveyor's name by SMS. That is where my team's world ends.
A typical AI-generated draft from that narrative, in Mermaid. Treat it as your colleague's homework to check — the defects in it are the lab.
flowchart TD
A[Claim reported: app / call / garage] --> B[Capture claim details]
B --> C{Policy active and cover valid?}
C -- No --> D[Send rejection letter] --> Z1[End]
C -- Yes --> E[Register claim in ClaimsCore, SMS customer]
E --> F{Issues found?}
F -- Missing docs or fraud signals --> G[Exception queue:
chase customer / investigate]
G --> H[Issue resolved]
H --> I{Claim < 50k and clean?}
F -- No issues --> I
I -- Yes --> J[Assign surveyor by postcode and workload]
I -- No --> K[Desk review by assessor]
K --> L[Team lead quality check]
L --> J
J --> M[SMS surveyor details to customer] --> Z2[End]
Read the narrative and mark it up — no AI yet. Highlight every actor, every decision, every queue, every percentage, and every sentence where Arun is emphatic. Emphatic sentences are load-bearing: SMEs raise their voice exactly where the last analyst got it wrong. Ten minutes.
Generate your own draft too. Prompt: "Convert this process narrative into BPMN-style Mermaid flowchart. Rules: one box per activity with the actor named, explicit gateways for every decision with percentages from the text, separate paths for separate queues, show any loop the text implies. Do not add steps the text does not state." Compare your AI's draft with the one above — differences between two drafts are free clues about where the narrative is ambiguous.
Build the walkthrough script. A walkthrough is not "does this look right?" — SMEs politely wave through diagrams they have not truly read. Write 8–10 closed questions that force the SME to walk specific tokens through the map, for example: "A theft claim arrives with no FIR — trace its exact path." · "Intake spots a metadata mismatch on the photos — which queue, and can it ever sit with the doc-chase claims?" · "An assessor finds a wrong cover code — what happens to the claim next, step by step?" · "Who performs each box on this map — can you name the team for every one?"
Execute the walkthrough against the SME answer sheet. The answer sheet below stands in for Arun. For each script question, first answer from the draft map, then check the sheet. Every mismatch goes in a correction log: Question | Draft says | SME says | Defect type (merged / invented / missing / wrong actor).
SME ANSWER SHEET (the real process — consult only during step 4+)
Q-paths Missing-docs claims wait in the PENDING queue (intake owns,
3 chase attempts, 15-day lapse letter). Fraud-signal claims
go to the SIU REFERRAL queue (SIU owns, 48h triage SLA).
The two queues have different owners, SLAs, and outcomes;
a claim can be in both states — SIU takes precedence.
Q-review Desk review outcome: approve for assignment, OR kick back
to intake for re-capture (about 15%). Intake corrects and
resubmits to desk review — this loop runs until clean.
Q-lead There is no team-lead quality check anywhere in intake.
Assessors' desk review is the only review step.
Q-stp Under-50k straight-through applies only after documents
are complete. No claim skips the completeness check.
Q-end Process ends at surveyor-assignment SMS. Payment, estimate
approval, settlement are downstream — out of scope.
Diagnose the three defect species. Sort every correction-log entry into: merged paths (two things the SME keeps separate, drawn as one), invented step (a box no one performs — AIs add "review" and "approval" steps because most processes have them), missing loop (a rework cycle flattened into a straight line — AIs prefer forward motion). If any species has zero entries, walk the map again with that species as the explicit target.
Correct the map. Redraw: split the merged queues into two labelled paths with their owners and SLAs, delete the invented step, add the rework loop with its ~15% rate on the arrow. Ask the AI to regenerate from your correction log, then verify the output box by box — regeneration can silently reintroduce a defect you already killed.
Re-run the walkthrough on the corrected map. Every script question must now be answerable from the map alone, matching the answer sheet. When all pass, stamp it: add a title-block note — "Validated against SME walkthrough, [date], scope: FNOL to surveyor assignment." An unstamped map should look naked to you from now on.
Write the walkthrough report. One page: number of questions asked, defects found by species, the correction log, and the one process fact you would never have learned from the narrative alone (there is at least one — the dual-state SIU precedence rule).
Ask the SME to trace specific claims through the map, never to approve the map. Approval invites politeness; a token that has nowhere to go invites the truth.
A complete lab has all of the following:
AI turns narrative into diagrams in seconds, and it will merge what must stay separate, invent what usually exists, and straighten what actually loops. The structured walkthrough — tokens traced, answers logged, corrections stamped — is what turns a plausible diagram into a validated map.