A draft business case for the claims document-AI pilot is already circulating — built on an unmeasured baseline, with three cost lines missing. Critique it first, then rebuild it honestly: measured baseline, benefit ranges with confidence levels, the full cost stack including tokens and evals, time-to-value, and kill criteria agreed before a rupee is spent.
The COO liked Divya's opportunity scan from Lab 10 and asked for a business case for its top item: AI-assisted document extraction and completeness checking at claims intake (step S4). A strategy consultant got there first — a glossy two-pager is already in the COO's inbox claiming 4.2x ROI. Divya's job is not to compete on gloss. It is to do the thing that makes AI business cases survivable: anchor every benefit to a measured baseline, price every cost including the unglamorous ones, and pre-agree what failure looks like.
From the ops team's time study and the Lab 10 operating data. This is what "measured" means — sourced, dated, reproducible.
BASELINE EXTRACT — Claims intake document handling (June, time study n=412 claims)
M1 Volume: 9,000 motor claims/month
M2 S4 completeness check: 8.0 min/claim median (measured), all claims
M3 Incomplete on arrival: 34% of claims (3,060/month)
M4 Doc-chase cost per incomplete claim: 22 min staff time across
3 attempts (measured on the chase subset)
M5 Chase-queue delay: 6.2 days average (per claim in M3 subset)
M6 Extraction errors found downstream: 3.1% of claims, 18 min rework each
M7 Fully loaded intake staff cost: INR 480/hour
M8 Current software cost of S4: none (manual step)
M9 NOT MEASURED: what share of total intake time is "document work";
customer NPS impact of chase delays; assessor time saved downstream
Critique target. Read it the way a CFO's analyst will read yours.
DRAFT BUSINESS CASE v1.0 — "Intelligent Document Processing for Claims"
D1 "Intake agents spend 60% of their time on document handling.
IDP will cut this by half, freeing 27 FTEs."
D2 Benefits: INR 3.9 Cr/year staff efficiency + "significant NPS
uplift" + "fraud leakage reduction of up to 2%"
D3 Costs: IDP platform licence INR 36L/year + one-time integration
INR 25L. Total year-1 cost: INR 61L.
D4 ROI: 4.2x in year one. Payback: 5 months.
D5 Recommendation: enterprise-wide rollout across motor, health,
and property lines in Q1.
Audit the draft against the baseline extract — line by line. Build a critique table: Draft claim | Baseline support? | Verdict (SUPPORTED / UNMEASURED / OMISSION / OVERREACH). Check D1's "60%" against M9 before anything else. Then list every cost category you would expect an AI system to incur across year one and mark which ones D3 contains. Expect one unmeasured foundation, at least three cost omissions, and one scope overreach.
Rebuild the benefit model from measured lines only. Three benefit streams the baseline can support: S4 handling time (M1 × M2 — model what an assist can realistically shave, not "half" by assertion), chase reduction (M3 × M4, driven by catching incompleteness at capture), and downstream rework (M6). Compute each in staff-hours/month, convert via M7. Show every formula in the sheet — auditable means a reviewer can trace any rupee to a cell to a baseline line.
Express benefits as ranges with confidence levels. For each stream, three cases with stated logic: conservative (e.g., assist saves 25% of S4 time; pilot-evidence confidence HIGH once measured), expected (40%), stretch (60% — confidence LOW, requires the model to handle the hard tail). Anything from M9 — NPS, assessor time — enters as unquantified qualitative benefit, listed but excluded from ROI. Ask the AI to sanity-check your ranges: "What assumptions must hold for a document-extraction assist to save 40% of an 8-minute review task? What typically prevents it?"
Build the full cost stack. Columns: one-time / year-1 recurring / ongoing. Lines the honest stack must include: platform or API licences; inference/token costs (compute from volume: 9,000 claims × pages × tokens per page at current provider list prices — have the AI estimate tokens per claim document set, then price it at the published per-million-token rates of two providers and take the dearer for planning); integration build; evaluation and monitoring (golden-set construction, monthly eval runs, drift review — price the analyst and reviewer hours via M7); change management (training 40 intake staff, floor support in weeks 1–6, SOP rewrites); model/prompt maintenance. Every line priced or marked with a priced assumption.
Compute honest ROI and time-to-value. Expected-case benefits over the full cost stack, year one and year two shown separately (year two drops one-time costs but keeps evals, licences, tokens, maintenance). Time-to-value: months until cumulative benefit exceeds cumulative cost, on the expected case — with the pilot ramp (limited volume in months 1–3) modelled, not assumed away. Compare your number to D4's "5 months" and note what drives the difference.
Write the kill criteria — before anyone falls in love. Three to five pre-agreed conditions that end the pilot, each measurable and dated: e.g., extraction accuracy below X% on the eval set at week 8; S4 median time reduction below Y% at 90 days; chase-queue entry rate not moving by Z; token cost per claim exceeding budget by 50%. Include who reviews and the decision they are empowered to make. A case without kill criteria is a commitment, not an experiment.
Scope the recommendation to what the evidence supports. D5 wants enterprise-wide rollout. Your baseline covers motor intake only (M1–M8). Write the recommendation as: pilot on motor intake, success gates from step 6, expansion decision only after gates pass — and name what would have to be measured before health/property lines could be cased.
Assemble the two-pager. Page 1: problem (baseline numbers), proposal, expected-case ROI with the range table, time-to-value, kill criteria. Page 2: full cost stack, benefits model summary with formulas, assumptions register, qualitative benefits, scope limits. Then the closing paragraph: three sentences on why this case can be trusted — every number traces to M-lines or a priced assumption, failure is defined, scope matches evidence.
The most dangerous line in an AI business case is a benefit computed from a baseline nobody measured — the ROI can be arithmetically perfect and still be fiction all the way down.
A complete lab has all of the following:
An honest AI business case is mostly subtraction — unmeasured baselines out of the benefits, missing lines into the costs, scope shrunk to the evidence — plus one addition no draft ever includes on its own: kill criteria agreed while everyone can still afford to be wrong.