This issue only includes materials that can be decomposed into inputs, machine boundaries, and human sign-off. Named finance team/CFO self-descriptions and controlled demo repositories are clearly marked and do not represent production conclusions from your books. Supplier rankings and course promotions were not adopted.
Today’s Most Actionable Implementations (3 items)
1. AR Sub-Ledger vs. General Ledger: Write Matching Rules First, Then Have the Skill Generate Exception Tables (Anthropic Finance Team)
- Scenario: Accounts receivable reconciliation. Anthropic Finance (Finance AI Product Lead Tim Ross, Finance Systems and Transformation Lead Lisa To) demonstrated at the June 2026 workshop: accountants write the skills themselves, not IT on their behalf. The AR reconciliation demo used synthetic tables, not audited production close data.
- Actionable Steps: Do not connect to ERP this week. Simply take two Excel copies from the prior month closed: invoice details (invoice-level), GL AR aging. Hard-code in the skill: match by invoice number, exception classification, variance bridge, exception formatting. Their on-site synthetic run criteria can serve as a checklist: GL has invoice but not in sub-ledger, sub-ledger has invoice but not in GL, amount mismatch. Prohibit auto-posting.
- Review Controls: Preparer ≠ Controller. The skill must include its own reconciliation layer: rows where source cannot be traced, amounts do not match, or fields are missing must be flagged by the system itself rather than waiting for manual review. They explicitly state: human review cannot be eliminated; full auditability is still under development. Run in parallel against closed packages; discard entire erroneous segments.
- Deliverables: Exception table (exception type / amount / source file row) + variance bridge + Controller sign-off. Zero entries on the books.
- Source: CFO Connect · Anthropic Finance Team Claude Skills | Source page indicates June 2026 workshop; publication date not disclosed | Nature: operator workshop transcript, AR demo uses synthetic data.
2. Variances: First Map the Quote-to-Cash Process, Then Let the Model Flag Exceptions; 302 Certification Remains Human (Former Nasdaq CFO)
- Scenario: Operating variances after PE ownership and multi-entity consolidation. Brad Wolfe describes himself as a three-time Nasdaq CFO now embedded in $20M–$500M PE portfolio companies. Self-report, not client attestation.
- Actionable Steps: Do not deploy a “live model” this week. First draw one revenue-to-cash process (opportunity → contract → invoice → collection) and the corresponding systems. Then use one cost center, closed month actual vs. budget. The model only flags rows exceeding your current thresholds and can point to customer / cost center / transaction type. Humans judge the same day: noise / trend / control issue requiring escalation. Board narrative is edited from the first draft only; do not build tables from scratch.
- Review Controls: Every source system has a named data owner; model exceptions have an escalation path; capital allocation, revenue recognition judgments, board risk disclosures, and any numbers requiring certification are never finalized by AI. His original logic: if a variance draft enters the board narrative without data quality checks, model validation, and output certification, it is equivalent to signing an unverifiable 302. Preparer ≠ final approver.
- Deliverables: Process map + exception list (noise/trend/control) + human-edited variance narrative. Unsigned items do not enter the management package.
- Source: The CFO Club · Brad Wolfe | Page indicates 2026-09-14 | Nature: named former public-company CFO interview transcript, self-report.
3. Monthly Reporting: Daily Extracts → Threshold Exceptions → Human Business Context → Board Package (Consultant Self-Report)
- Scenario: Month-end management reporting. Andrew Lokenauth describes himself as a former bank FP&A / strategic finance professional now serving as part-time CFO/COO for $1M–$5M companies. The tool chain reflects his consulting engagements, not independent client case studies.
- Actionable Steps: This week work on one closed departmental P&L only. Fixed sequence: accounting / CRM / payroll first land in one clean table (he uses daily-refresh pipelines); flag variances against your thresholds and draft plain-language explanations; finance lead adds business context line by line before any formatting. Do not generate a full board package on the first pass.
- Review Controls: He provided positive and negative examples used directly as controls: ① One Series B company’s AR aging pattern flagged customer concentration risk approximately 6 weeks early and proved useful; ② Another client with high invoice volatility saw forecasts miss by approximately 15%–20%, yet the output appeared stable until the CFO caught it at the board meeting. Looks complete ≠ reliable. When data is dirty, definitions change, or revenue mix shifts, the first draft must never be treated as a forecast. He notes that roughly 60% of project time in his engagements is spent cleaning data. Perform data inventory before discussing models.
- Deliverables: Variance exception table + human-added context + time comparison (baseline draft vs. machine first draft). Unsigned items are not distributed externally.
- Source: The CFO Club · Andrew Lokenauth | Page indicates 2026-09-28 | Nature: part-time CFO / consultant self-report, includes vendor tool names used, not an independent evaluation.
Accounting / Close / Controls
- Accounts receivable sub-ledger vs. general ledger: see item 1 above. Anthropic AR Reconciliation Skill
- Invoice-to-PO matching: rules first, model only explains, cannot post. See open-source item. controlled-finance-agent-demo
Closed loop: two closed-period tables → rule-based matching / exception interception → human reviews exceptions only → sign-off before any posting consideration. Independent, GL-connectable revenue recognition automation: data unavailable.
FP&A / Planning / Reporting
- Trace variances to customer / cost center before writing narrative: see item 2 above. Brad Wolfe
- Threshold exceptions require human-added business context: see item 3 above. Andrew Lokenauth
Item 1 can be extended to “decompose revenue movements by source / sales segment, output leadership slides” — Anthropic mentions BigQuery connectivity; this week still use closed P&L copies only, do not connect directly to the warehouse. Independent budget preparation or approved forecast production cases: data unavailable.
Treasury / Cash / Risk
See item 3 above. AR aging patterns for customer concentration / liquidity early warning
This week’s boundary: only flag Top customer concentration and aging migration on closed AR aging; treasury reviews whether collections escalation is warranted; do not change collection accounts or issue payment instructions. Independent, bank-connectable 13-week cash forecast production cases: data unavailable.
Tax / Compliance / Audit
Data unavailable. No new AI implementation cases or practical methods for tax research, SOX / internal control, or audit evidence management within the last 365 days were identified this period.
CFO / Leader Team-Building Experience
- See item 1 above. Anthropic Finance Skills Library
- See item 2 above. Brad Wolfe
Directly usable organizational actions (all self-reported):
- The best preparers codify steps into skills: inputs, calculation rules, how missing data is flagged, output format; place in version control so new hires use the same playbook from day one rather than each person maintaining separate prompts.
- Accountants write reconciliation skills; FP&A writes analysis skills; IT does not write business rules. The person judging correctness must understand the accounting, not necessarily the code.
- New skills first run in parallel on closed historical data (community reports: some teams ran parallel for two to three months before retiring manual processes).
- Outputs that may enter board or covenant discussions (variances, cash flow) must identify both data owner and signatory; the model may draft, but certification responsibility does not transfer.
- Before requesting headcount or tools, ask: once this process is decomposed, is an intermediate human layer still required? If owners and prohibition items cannot be written down, do not add headcount or deploy tools.
Open Source / AI Engineering Reference
- Accounts payable: code enforces controls, model only explains, only humans can post
- Reusable: synthetic lab scenario, vendor invoices matched against approved POs. Deterministic checks (duplicates, approvals, vendor, currency, amount, cost center) produce conclusions first; model may optionally be called to write explanations. Control results are passed as immutable context to the model. Authoritative order is hard-coded: approved permission > secondary evidence > pending conflict > UNKNOWN. Hash-chained audit log. Test coverage: intercepted invoices cannot be approved, missing permissions, currency conflicts, human review gate, log tampering.
- Do not: connect to production AP, allow model to modify POs / supplement evidence / post. 0 stars, last push 2026-08-11, small repository, intended as architectural example.
- Source: controlled-finance-agent-demo | Source page shows creation / code push 2026-08-11 | Nature: repository / controlled demo, not recommended as production AP system.
Small Experiments for This Week
- AR Exception Table (Owner: Accounting; Controller sign-off): Prior-month closed invoice listing vs. GL AR aging, match on invoice number only. Produce three tables: GL only / invoice only / amount difference. Sample 10 items back to source files. If mismatch rate >5%, discard the rule set. Zero entries on the books.
- Same-Day Variance Judgment (Owner: FP&A; Finance lead final approval): One cost center, closed month, use current materiality thresholds. Model outputs only rows exceeding thresholds that can be traced to account / customer. Human labels noise / trend / control. Unsigned items do not enter the weekly report.
- Controls Before Model (Owner: AP; Internal control co-sign): 20 historical invoice copies (including no PO, duplicates, currency mismatches). First apply rules to produce pass / intercept; model is permitted only to write “why intercepted.” Any change to collection accounts is frozen. Zero payments on the books.
- Data Inventory Before Tools (Owner: Finance lead): List naming differences and missing months for the same customer / account across accounting, CRM, and payroll systems. If you cannot write “source system, field, owner,” do not launch any new model this week.