Business case — modelled, not measured
Audience: the founder, before agreeing to the read-only pilot. Purpose: what the desk could be worth, as a model with every assumption on the table and adjustable, and what it would cost. Nothing in this document has been measured. The pilot's fifth step (roadmap) replaces the modelled column with a measured one; until then no figure here should be quoted without the word modelled.
1. What is being valued
Three things, in the order a finance lead would rank them:
- Money that does not leave wrongly. A duplicate paid, a payment to a changed bank account, a bar charge on the award that the sponsor disallows at audit.
- Hours in the monthly close that are tie-out, review and evidence-gathering rather than judgement.
- The cost of not being able to answer — audit findings, sponsor questions, a founder decision made on a stale number.
Only the second is modelled below in hours; the first is modelled as exposure, not as "fraud caught", because nothing has been caught.
2. Assumptions — adjust these first
| # | Assumption | Value used | Source | Adjust it if |
|---|---|---|---|---|
| A1 | Invoices per month | 26 (313 over twelve months in the seed) | synthetic ledger | the accounting export says otherwise |
| A2 | Bank lines per month | 31 (371 / 12) | synthetic ledger | — |
| A3 | Finance lead's fully loaded hour | $95 | placeholder, to be replaced by finance | — |
| A4 | Founders-office associate's fully loaded hour | $55 | placeholder | — |
| A5 | Hours per close today, by task | see §3, the finance lead's own estimate | interview, PLANNED — placeholder until then | the pilot's measurement differs |
| A6 | Share of each task the desk carries | see §3 | the demo's screens, judged by what they compute | the finance lead disagrees on any row |
| A7 | Duplicate exposure | 0.5 % of invoice value paid twice per year, half recovered | industry pattern, unverified here | the first labelled month finds a different rate |
| A8 | Bank-change fraud exposure | one attempt per two years at the median rush invoice ($18,500 in the seed) | pattern, unverified | — |
| A9 | Award disallowance exposure | 2 unallowable lines of 31 in the seed ($3,336) | synthetic ledger | the sponsor's audit history |
| A10 | Build cost to pilot | 6 engineer-weeks (adapters for four exports, a labelled month, pilot instrumentation) | estimate | — |
| A11 | Run cost | Hosting inside the existing Wiki, no model, no licence — $0 incremental | fact for the demo; the pilot's exports may carry a cost | — |
3. Hours per close — modelled
| Task | Today (A5, finance lead's estimate) | With the desk (A6) | Modelled saving | What the desk actually does for the task |
|---|---|---|---|---|
| Bank tie-out | 14 h | 6 h | 8 h | Three passes propose; a person decides suggestions and works exceptions |
| Supplier-bill review | 10 h | 5 h | 5 h | Findings on the row with evidence; the reviewer reads the drawer, not the mailbox |
| Award compliance | 8 h | 4 h | 4 h | Every line judged on read; reclass drafts prepared |
| Accrual drafting | 6 h | 3 h | 3 h | Received-not-invoiced drafts from approved POs |
| Audit preparation | 8 h | 3 h | 5 h | Hash-chained log, provenance on every number |
| Total | 46 h | 21 h | 25 h |
Modelled value of the hours: 25 h × a blended $75 (A3, A4) × 12 = $22,500 a year, modelled. This is the number most likely to be wrong, in either direction, and it is the one the pilot measures first.
4. Exposure — modelled, not caught
| Exposure | Modelled annual amount | Basis |
|---|---|---|
| Duplicates paid and not recovered | ≈ $12,000 | A7 on $4.9M of ap.manual spend in the seed |
| A payment to unverified bank details | ≈ $9,250 a year, expected value | A8 |
| Award disallowance at audit | ≈ $3,300 | A9, the seed's two unallowable lines |
| Total modelled exposure | ≈ $24,500 | none of it observed |
The desk does not "save" this; it makes each of the three a refusal or a finding before the money moves, which is a different claim. The honest statement is: on the synthetic ledger the rules find every planted instance of these three; whether the real ledger contains any is unknown.
5. Cost
| Item | Modelled |
|---|---|
| Build to pilot (A10) | 6 engineer-weeks |
| Run (A11) | $0 incremental |
| Finance lead's time to label one month | ≈ 6 h, once |
| Founders-office associate's time in the pilot | inside the supplier-bill review hours above |
6. Sensitivity — the three assumptions that decide it
- A5 / A6 (hours). If the finance lead's real close is 30 h rather than 46, the modelled saving halves. Measure before quoting.
- A7 (duplicate rate). At 0.1 % the duplicate exposure is ≈ $2,400; at 1 % it is ≈ $24,000. The first labelled month gives the real rate with an n.
- A10 (build). Adapters for real exports are where estimates go wrong; the two adapters that exist took the shape of their sources, and the four planned ones will too.
7. What would make this case wrong
A close that is mostly judgement rather than tie-out (the hours model collapses); an accounting system that already refuses duplicates and unverified bank changes at entry (the exposure model collapses); or a pilot in which the finance lead does not open the queue (the whole case collapses, and the roadmap says so).