This issue only includes new materials that can be broken down into inputs, machine boundaries, and human sign-off. Close / bank reconciliation repetitive templates and vendor lists do not enter the main body.
Today’s Most Worthwhile Implementations (2 items)
1. Month-end reconciliation: First copy your current working papers; then write “what a wrong entry looks like” into shutdown rules
- Scenario: Accounting month-end close. A Controller with approximately 21 years of bookkeeping experience claims to have used Claude / Cowork to connect to NetSuite, developing approximately 50 close skills covering balance sheet reconciliations, month-end accruals, CEO month-end draft, sales YoY, budget vs. actual, 13-week cash forecast, etc. The hard point is not the model, but teaching it what constitutes an error on your books.
- Actionable steps: This week, select only one high-frequency account (the example in the text is customer deposits: receipts recorded as liabilities, cleared upon delivery). In the first phase, do not redo the format: first open the prior month’s reviewed reconciliation, then query the current period general ledger; the GL query must first reproduce the prior month balance, then roll forward to the current period; still produce your current four-page working papers; unreconciled items must be traceable to sales orders and expected/actual/cancelled shipment dates; differences with the GL must be formula-linked to 0.00, no hardcoded numbers allowed. If it does not reconcile, stop; do not “recalculate history to force balance”.
- Review controls: Preparation ≠ approval; the robot only prepares. Journal entry skills first output a preview: draft entry + supporting documents + duplicate check + debit/credit balance, human reviews before posting. Routine accruals may only be considered for auto-posting after checks pass and working papers are finalized; sensitive entries such as payroll must be pre-approved by the Controller. Commentary drafts can self-check against your scoring table; changes to the scoring table require human sign-off. Self-reported savings of approximately 20–30 hours/month; no independent verification, not to be treated as ROI.
- Deliverables: Reconciliation package with the same structure as the prior month, list of unexplained differences that trigger shutdown, journal entry preview (zero entries on books), human review and sign-off section.
- Source: Books & Bots · Teaching the Model What a Wrong Entry Looks Like｜Date unclear (long-form public post visible this period)｜Nature: operator self-report, not a third-party audited production case.
2. AR sub-ledger vs. GL: First exact invoice number match, then fuzzy match, treat all remaining as exceptions, do not fabricate stories
- Scenario: Month-end AR. Prompt words publicly shared by the Anthropic finance team that can be copied: inputs are two tables — open invoice details and GL AR aging.
- Actionable steps: This week, use only a copy of the prior month closed books, do not connect to production posting. The machine performs only three layers: ① exact match by invoice number; ② remaining items fuzzy match by vendor + date ±3 days + amount ±1%; ③ items still unmatched go to exceptions and generate suggested journal entry drafts. Output fixed four blocks: summary, matching table, exception table, suggested entries. Prohibited: changing amounts, changing master data, auto-posting. Can be extended to department/account MoM variance commentary: only list amounts and percentages, leave reasons blank for others to fill, do not fabricate business stories.
- Review controls: Controller only reviews exceptions and suggested entries. The execution sequence in the text is: data extraction → processing → output in format → human review → then distribution or posting; step 4 cannot be skipped. Skills should be written by the accountants performing the reconciliation, not by IT on their behalf. Demonstration/transcript materials, not to be treated as your reconciliation conclusions.
- Deliverables: Matching table, exception table, suggested journal entry drafts, human review checklist. Zero entries on books.
- Source: CFO Connect · Anthropic Finance Workflows and Copyable Prompts｜Date unclear (text references 2026-06 workshop)｜Nature: community transcription of internal finance team processes + prompts, not a customer implementation white paper.
Accounting / Close / Controls
See Today’s Most Worthwhile Implementations item 1. Close Skills / Shutdown Rules
Closed loop: Prior month reviewed working papers → machine rolls forward in original format and reconciles to 0.00 → unexplained differences trigger shutdown → Controller signs preview before posting. AR three-layer matching see item 2, not repeated. Independent, GL-connectable revenue recognition automation: Data unavailable.
FP&A / Planning / Reporting
See Today’s Most Worthwhile Implementations item 1 (budget vs. actual, CEO month-end draft) and item 2 (MoM variances list only numbers, reasons to be supplemented). Close Skills Long Post｜Copyable Prompts
Input closed actuals vs. budget → model only writes commentary draft and self-scores → human supplements one-time items and signs off. Independent, approved forecast or budget negotiation production cases: Data unavailable.
Treasury / Cash / Risk
See Today’s Most Worthwhile Implementations item 1. 13-Week Cash Independent Check Skill
What is reusable is not “automatic forecasting”, but: if the forecast model’s own diagnostic items score the numbers it itself produced, bad forecasts get washed into PASS. In one run 9 out of 10 items passed, yet cash inflows were significantly higher than prior year actual receipts; upon breakdown roughly half the checks could never fail. Approach: build a separate non-shared code audit skill that only reviews the forecast; each diagnostic must be able to both flag red and pass on historical snapshots before going live. Independent, bank/funds ledger connectable new production cases: Data unavailable.
Tax / Compliance / Audit
Data unavailable. No new AI implementation cases or practical methods for tax research, SOX/internal controls, or audit evidence management within the last 365 days were identified this period.
CFO / Leader Team Building Experience
See Today’s Most Worthwhile Implementations item 2. How Anthropic Finance Skill Library Is Divided
Tim Ross (Head of Finance AI Products) organizational stance: the best preparation writes procedures in plain language, versions them, and the whole team runs the same playbook; skills are written by accountants/FP&A, not outsourced to IT; posting, reporting to leadership, and external documents must carry human sign-off. Lisa To side handles systems and transformation. Do not treat “150 skills” as the target; first write one month-end procedure into a version-controllable desktop manual. Named enterprise finance leads from a different source: Data unavailable. LinkedIn data unavailable / authentication not passed; do not substitute with single social media posts.
Open Source / AI Engineering Takeaways
Data unavailable. Visible repositories this period are mostly slogan-style READMEs or reconciliation templates duplicative of recent days, with no new information on fields or human review gates; not recommended as templates. Reusable engineering points are already covered in items 1 and 2: formula reconciliation, shutdown rules, isolation between check skill and the object being checked.
Pending Verification Leads
- Construction company finance lead claims using AI/RPA for batch payments, Copilot agent, Xero automation, and promises follow-up report on “what works, what doesn’t, real numbers”; currently only self-introduction, no field tables or human review logs, not treated as a case. Related post
- No new cross-verifiable startup headcount reduction signals this period. LinkedIn data unavailable / authentication not passed.
This Week’s Small Experiments
-
Duplicate an existing reconciliation worksheet (Owner: Accounting; Controller sign-off) Select one liability or intercompany account from last month closed. Machine may only roll forward last month’s worksheet, must not redo format. Gate: GL query can reproduce prior month balance; difference rows are formulas not hardcoded; sample 10 unreconciled items can trace back to orders. Zero entries on books.
-
Unexplained differences must trigger shutdown (Owner: Accounting) Hard-code in the prompt: if differences have no explanation the reconciliation is incomplete; output list then stop. Gate: if sentences such as “reasonable estimate / seasonal” without source appear, the prompt is invalidated.
-
AR three-layer matching (Owner: AR; Controller reviews exceptions) Prior month open invoices + GL aging. Run only exact → fuzzy → exception. Gate: exception table must not contain items that could have been exactly matched; suggested entries do not post to books.
-
Forecast health check must not self-congratulate (Owner: Treasury or FP&A) If a 13-week cash or collection forecast already exists, write a separate non-shared-code checklist using prior year same-week actual receipts as benchmark. Gate: at least one check must be able to fail on a “clearly overstated” historical snapshot. If you cannot write such a check, do not mark the forecast health check as PASS this week.