Capstone: Design the workflow
Pick a real finance workflow from your own map, design it with controls as first-class citizens, and write the design doc a controller would approve.
The capstone ships one AI-assisted finance workflow, end to end, with the full evidence and control apparatus — the whole course in one artifact chain. Strongly preferred: a workflow from your Module 1 opportunity map, on your real (sanitized) documents and data. The Alder fallback: build out the carrier rate-audit workflow (extraction reference → per-invoice check → exception lanes → disposition log → recovery tracking) using the course's data.
Choosing well
- Document- or matching-heavy beats judgment-heavy — the capstone should showcase extraction/checking/flagging with human gates, not attempt to automate an estimate or an accrual judgment (proposing one of those is an automatic redesign note from review).
- Material enough to matter, bounded enough to finish: one document type or one detection domain, a few weeks of volume. 'AP invoices for our top 10 vendors' beats 'all of AP'.
- Control-impact positive or neutral — revisit your map's fourth column; the strongest capstones strengthen a control (coverage 5%→100% stories). If yours threatens one, the design doc must show the redesign that resolves it.
- Data policy cleared in writing before any real document touches any tool: classes checked, masking plan stated, approved tools named. The clearance paragraph goes in the design doc; capstones without it don't proceed to build — same rule your workplace should have.
The design doc (2 pages, six sections — a controller is the audience)
- 1The workflow and the money: what it does today, hours and error profile, the materiality line ('worst realistic misstatement'), and the payback estimate with the capture-rate honesty from the strategy course if you've taken it (rough ranges otherwise).
- 2The design: the step diagram — extraction/check/match steps marked AI or deterministic, every gate with its named approver role, every lane with its destination. The bright lines visibly honored (no posting path, no self-approval).
- 3The contracts: each AI step's prompt sketch with its escapes (UNREADABLE/CONFLICT/confidence) and its injection line; the test-set plan (12+ cases: clean, messy, edge, one adversarial, one duplicate).
- 4The control rows: the 5-8 matrix rows this workflow needs, in the Module 5 format, evidence locations named. Written now, before build — controls designed in cost columns; controls added after cost quarters.
- 5The rollout: parallel-run period (AI alongside the current process, outputs compared — finance's natural A/B), the review-capacity plan, thresholds starting conservative with their evidence-to-loosen criteria.
- 6Success criteria: three checkable sentences for the soak ('≥95% of invoices auto-extract with clean checks; 100% of exceptions dispositioned within 2 days; zero unreviewed items reach payment'), plus the baseline measurement you'll take before switching anything on.
Show the doc to a finance colleague — ideally whoever owns the control environment (controller, senior accountant, your manager wearing that hat). Their objections are the capstone working: every 'but what about the case where…' becomes a test case or a matrix row. Reviewers from this course will go straight for your SoD grid and your evidence columns; a workplace controller will go for the degradation plan. Both are right.