Sheet 05 · PIL-005
Invoice coding assist
AP staff hand-code roughly 2,400 invoices a month, and coding errors surface late in close and require rework across two teams.
Record summary
- Stage
- Paused
- Risk pathway
- Elevated review
- Data classification
- Confidential
- Participants
- 6
Current status
Paused since 2026-07-24 pending vendor documentation. Resume criteria are written and the vendor has a September 30 checkpoint.
- Next decision
- Resume if documentation is executed by 2026-09-30; otherwise recommend retiring the vendor evaluation
- Where it is decided
- AI Governance Committee, 2026-09-15
The record
- Expected outcome
- Coding suggestions accurate enough to cut rework, with every posting traceable to a person who approved it.
- Business owner
- Sofia Brand, Controller
- Program coordinator
- AI Program Manager — coordinates and records; does not approve
- Participating teams
- Finance & Procurement
- Tool
- Ledgerly AI — Ledgerly Financial Software TL-05
- Originating use case
- Accounts payable invoice coding assistance UC-006
- Related decisions
- DEC-2026-06
- Start date
- 06 Jul 2026
- Decision date
- 30 Sep 2026 (in 28 days)
- Cost to date
- $7,100 — licensing, vendor fees, and third-party support. Participant time is not included and is the larger cost.
Measures
Defined before the pilot started. That is what makes a result a finding rather than an opinion formed afterwards.
| Measure | Target | Current | Evidence |
|---|---|---|---|
| Coding error rate against a manual audit sample | At or below the 4.1% manual baseline | Not started | Pending validation |
| Rework hours in month-end close | Reduce from 34 to under 20 | Baseline captured | Pending validation |
What the pilot has surfaced
Open questions and lessons sit here rather than in a closing report, because they are what the next pilot needs and the next pilot starts before this one ends.
Open questions
- Which subprocessors touch invoice data, and in what jurisdictions?
- If the tool posts an incorrect entry, what is the correction path and who is accountable for it?
Lessons learned
- Pausing early cost two weeks. Discovering the subprocessor gap after go-live would have cost far more.
- Writing resume criteria at the moment of pause kept this from quietly becoming a cancellation.
Guardrails and dependencies
Guardrails
- Suggestion-only during the pilot — no automated posting until the review completes
- 100% human approval of every entry rather than the sampled review originally proposed
- Test environment only until the data-processing agreement is executed
Dependencies
- Vendor data-processing agreement naming subprocessors (owner: Risk, Legal & Privacy)
- Reversal and correction procedure for posted entries (owner: Finance)