This issue only includes materials that can be decomposed into inputs, machine boundaries, and human sign-off. Benchmark tests and portfolio repositories are marked as such and are not production conclusions from your ledger.
Top Implementable Items Today (3 items)
1. Close Working Papers: Model produces only summary-table draft; the same package requires human review before posting
- Scenario: Month-end close. Provide a set of fictional-company working papers and require the model to locate figures, apply company rules, and deliver one results table. Common pitfalls: incorrect account coding, cross-period invoices, omitted commission accruals.
- Actionable this week: Do not connect to the production general ledger. Use last month’s already-closed copy (details, contracts, commission schedules) and instruct the model only to locate numbers, apply current cut-off rules, and populate the summary table. Posting, accruals, and adjustments are all disabled.
- Review control: Preparation ≠ Controller. Any differences identified by the model must trace back to source cells or file pages. Twelve licensed accountants (average 5.4 years’ experience, six with Big Four backgrounds) averaged ~37 % under no-colleague conditions; frontier models achieved full marks on the same simplified task set. 5/12 participants viewed the questions as unlike daily work. Scoring was performed by model judge with ~98 % human spot-check agreement. Do not translate “model is faster and cheaper” into “close can run unattended.”
- Deliverables: Summary-table draft, source-list of figures, human-review checklist. Zero entries posted to the ledger.
- Sources: Mercor · Accountant对照实验｜2026-10-01｜CFO Brew coverage｜2026-10-09｜Nature: research/media, fictional working papers, not client close.
2. Accounts Payable: Model reads invoices only; three-way match is code-enforced; unsafe auto-approvals must equal zero
- Scenario: Pre-payment AP. Inputs are invoices, POs, and goods receipt (warehouse-provided sample ledger).
- Actionable this week: Extract items suitable for model handling; matching, release, and rejection must use hard-coded rules. Sample logic: three-way match →
auto_approve; unit price > PO by 10 % or above auto-approval threshold → human review; no PO, quantity mismatch, total imbalance, supplier mismatch, duplicate invoice → reject. - Review control: AP owner reviews only holds; payment approval remains separate. Sole acceptance criterion:
unsafe_auto_approvals = 0(invoices that should have been held or rejected must never auto-release). CI failure halts the pipeline. Warehouse sample is self-described as fully offline, 4-star, not production ERP. - Deliverables: Decision table (pay/hold/reject + reason), human-review queue, evaluation report. Zero payments posted to the ledger.
- Source: tahasiddiquii / invoice-ap-agent｜Repository page shows creation 2026-06-30, last code push 2026-07-01｜Nature: repo / portfolio.
3. Board Commentary: Numbers come from the table; model writes only MD&A draft
- Scenario: Monthly management reporting / board pack. Make reads Google Sheet
actuals(entity, currency, period_yyyy_mm, metric, amount) → fixed JSON → model writes commentary → writes to Notion run log → Slack notification. - Actionable this week: Use only one already-closed entity table. Exchange rates and amounts must not be altered by the model. Commentary is excluded from the board pack and external distribution by default.
- Review control: FP&A validates numbers and cut-off; Controller spot-checks material variances back to source tables. 2-star learning project; code last updated 2025-12-18.
- Deliverables: Structured actuals JSON, MD&A draft, run log, human sign-off.
- Source: marjaanah-stack / finance-board-pack-ai-automation｜Repository page shows creation 2025-11-24, last code push 2025-12-18｜Nature: repo / template, not a live company board process.
Accounting / Close / Controls
- Close working-paper draft only: see Top Implementable Item 1. Mercor对照实验
- AP safety gate: see Top Implementable Item 2. invoice-ap-agent
Closed loop: closed-period copy → rules first generate pass/intercept → human reviews exceptions only → posting or payment only after sign-off. Independent revenue-recognition automation connectable to the general ledger: Data unavailable.
FP&A / Planning / Reporting
See Top Implementable Item 3. Board-pack template
Validated actuals input → model writes commentary only → human corrects facts and tone. Independent budget-preparation or approved-forecast production cases: Data unavailable.
Treasury / Cash / Risk
Data unavailable. No new production cases this period for cash forecasting, bank-reconciliation, or DSO automation that can connect to bank or treasury ledgers.
Tax / Compliance / Audit
Data unavailable. No new AI production cases or practical methods in tax research, SOX/internal control, or audit-evidence management published within the last 365 days.
CFO / Leader Team-Building Experience
Independent, named corporate finance leaders discussing AI fluency, ownership splits, or review/control practices: Data unavailable. LinkedIn data unavailable / authentication failed; single social posts or vendor GTM materials are not substituted.
Open Source / AI Engineering References
Use the public 10-question close world first; test whether “the same working-paper set produces consistent results across multiple runs” before discussing production rollout
- Reusable architecture: Fictional-company world (ledger export + tables + PDFs) → agent executes task → independent scorer. Public set contains 1 world, 10 questions; full 160-question set is private. Authors note: best model achieves ~2.6 % fully correct across 8 runs on the complete set; ~70 % of failures are reasoning errors, not file-location failures—differences previously identified are subsequently dropped from the voucher.
- Pilot: Run only the public 10 questions; run each question at least 3 times. Any single run producing entries or totals inconsistent with the working papers disqualifies the workflow from production posting.
- Note: Evaluation set, not a close system; does not include tax, audit, consolidation, or multi-currency.
- Source: Hugging Face · APEX-Accounting Harbor public set｜Corresponding research page 2026-07-31｜Nature: evaluation data.
AP safety gate and board-commentary draft: see Top Implementable Items 2 and 3; not repeated here.
Pending Verification Leads
- Social-media claims of building multiple agents (“financial analysis / FP&A / finance operations / …”) with Teamily lack independent code, data flows, and human-review gates and are not treated as cases.
- No usable LinkedIn posts this period.
Small Experiments for This Week
-
Close Working-Paper Draft (Owner: Accounting; Controller sign-off) Last month’s closed single entity: details + one error-prone schedule (cross-period invoice or commission). Model populates only the summary table and lists sources. Spot-check 10 figures back to source files. Mismatch rate >5 % → prompt discarded. Zero entries posted to the ledger.
-
AP Unsafe-Auto-Approval Gate (Owner: AP; Internal Control co-sign) Use warehouse sample or 20 desensitized invoices (including no-PO, duplicate, price variance). Auto-release limited to three-way match only. Acceptance: zero invoices that should have been intercepted are auto-released. Zero payments posted to the ledger.
-
MD&A Draft Timing (Owner: FP&A) Fix one validated P&L. Draft once by hand, once with model; record drafting + review minutes for each. Any number altered by the model is rolled back. Proceed to a second entity only if total time decreases and no material new factual errors are introduced.
-
Same Package Run Three Times (Owner: Finance Systems) Select one close task package and run the identical prompt three times. Any inconsistency in entries or totals → that workflow is excluded from posting-candidate status this week.