For finance teams: First remove “compilation, collections preparation, month-end close working papers” from human hands; machines only perform first-layer monitoring/drafting. Budgeting, credit approval, posting, and external narrative remain signed off by humans. This issue’s two items are named operators’ self-descriptions, not audited transformation reports.
Today’s Most Actionable (2 items)
1. Month-end close for a 40-person company: Claude Code agent reduces 8–10 hours to under 30 minutes (self-report)
- Scenario: Fuelfinance founder and CFO Alyona Mysko, using internal finance for month-end close, not an external close product brochure.
- Actions: Install Claude first, enable code execution/file creation/Skills/memory; build “About me + work rules + output format” context; first project only on “month-end close”: chart of accounts, prior month TB, close checklist, fixed output template. She also has a 13-week cash agent (claims 2–3 hours saved weekly) and MCP feeding finance/operational data for management self-service what-if, not covered separately this issue.
- Review controls: She notes Claude cannot handle cleaning, consolidation, permissions; complex large tables prone to generic drafts. Capex and budget final approval must be human. Month-end package reviewed by controller against TB before close.
- Deliverables: Month-end working papers, close checklist check records, variance table vs TB; no direct posting.
- Source: Alyona Mysko / The CFO Club | Page dated 2026-07-27 | Nature: Supplier founder self-report / internal usage, not client attestation.
2. Health-tech CFO: Invoice, aging, and expense anomaly monitoring first; collections workflow still under construction
- Scenario: Eden CFO Fernando Blumenkron. U.S. parent with multiple Latin American entities (Mexico/Brazil/Colombia, etc.), SaaS + usage + implementation + distributor model, lean finance team covering FP&A, accounting, invoicing, collections, cash application, and tax.
- Actions: Over the past year, shifted “pulling data from multiple systems, cleaning tables, checking invoice status, reviewing aging, classifying expenses, manual follow-up” to: monitor data → flag exceptions → classify transactions → rank collections priority → provide first-layer analysis to humans. Collections direction is a single application showing bills, aging, email/Slack, receipts, and customer notes; model suggests next step (reminder/ escalation/ payment plan/ legal) and drafts messages. He notes the collections setup is still under construction, with early results not yet rolled out.
- Review controls: Sensitive actions require human approval. Budget approval, capital allocation, credit approval, payment plans, legal escalation, investor communications, and runway trade-offs are not delegated to the model. Runs on Claude Enterprise. He emphasizes: the hard part is clean data, clear owners, and documented processes—not the model itself.
- Deliverables: Overdue account priority list, invoicing gap list, expense classification pending review, collections suggestions (not sent until approved).
- Source: Fernando Blumenkron / The CFO Club | Page dated 2026-08-04 | Nature: Operator interview with visible full text, not an audited transformation.
Accounting / Close / Controls
- See Today’s Most Actionable item 1. Minimum pilot: single entity, parallel run with prior month already closed; output only into working papers. Independent, new revenue recognition / journal automation connectable to the general ledger: Data unavailable.
FP&A / Planning / Reporting
- Scenario: Former P&G / Unilever FP&A, participated in Squarespace listing narrative; Christian Wattig now provides FP&A training.
- Actions: Model only writes analysis / model first drafts. He observes: while some expect 95% modeling replacement, significant time is actually spent reviewing output, eroding timeliness. Time-series: feed historical data to the model to write Python that runs ARIMA in Excel; variance analysis uses driver trees to explain “what moved”; budget review used to flag risks, not to edit numbers directly.
- Review controls: FP&A owner reconciles against approved budget and actuals; every driver-tree line must trace back to account / cost center. First drafts do not enter the board package.
- Deliverables: Variance driver trees, forecast code / worksheets, budget risk list.
- Source: Christian Wattig / The CFO Club | Page dated 2026-07-11 | Nature: Former corporate FP&A leader methodology share, not an audited transformation of any single company.
- See Today’s Most Actionable item 1; extendable to what-if and board materials; management self-service queries still require permissions and final approval.
Treasury / Cash / Risk
- See Today’s Most Actionable item 2: overdue identification, collections action suggestions, sent only after human approval. Extendable to cash monitoring; he did not provide the 13-week model fields.
- See Today’s Most Actionable item 1: 13-week cash agent is self-reported hours saved; table structure not disclosed. Independent, new cash forecasting production case connectable to bank / treasury ledger: 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
- Scenario: Mitch Petracca, Managing Partner at Forward Firm (remote CPA / consulting, ~50 QoE projects per year), later founded Finsider.ai.
- Actions: First write “better” numbers (cycle time, coverage rate, hours per project, close days), then select tools. Tackle only one process at a time that has a standard answer, can be timed, and can be accepted; complete before moving to the next segment. He acknowledges he did not follow this approach and paid the price in trust. Layering: confirmation layer (classification, reconciliation, full-ledger exceptions, KPIs, aging, data-package drafts) to machines; conclusion layer (per-project QoE conclusions, tax sign-off, etc.) to humans. Operational examples: post-meeting notes, journal classification (unclassified items cleared within 48 hours max), health-score QBO file before QoE work.
- Review controls: If an indicator cannot be named and moved, there is no project yet. Do not automate reports no one reads.
- Deliverables: One-page pilot card (metrics, owner, acceptance criteria) + single-process run log.
- Source: Mitch Petracca / The CFO Club | Page dated 2026-08-17 | Nature: Operator interview; opening summary states Forward Firm reduced project duration ~40% and expanded sampling to full-ledger scan, but the body does not detail the calculation method—do not treat as audited result.
- See Today’s Most Actionable items 1 and 2: human review boundaries (budget / capital / credit / legal) and “data & owner first, then model.”
Open Source / AI Engineering Reference
- Scenario: Third-party payment reconciliation (order / gateway settlement / bank receipt), suitable for gateway reconciliation pilot, not general-ledger close.
- Actions: Deterministic three-way matching, calculate differences using agreed MDR, GST benchmark, bank UTR; exception classification (MDR overcharge, GST mismatch, gateway hold, bank non-receipt / short payment, missing gateway record); then produce balanced journal draft (example: DR 1140 Gateway Receivable / CR 5120 Fees / CR 2210 Input VAT). Three CSVs can be uploaded. XGBoost uses only order-timepoint features for risk scoring; settlement results are excluded from features to avoid leakage.
- Review controls: 50,000 synthetic demo records, 0 stars. Drafts must not be posted. Audit trail is append-only ledger.
- Deliverables: Exception list, journal draft, dispute JSON; human review before entry into gateway / general ledger.
- Source: Ledgermind-AI | Created 2026-08-26, code pushed 2026-08-29 | Nature: repo / 0-star synthetic demo.
- Scenario: FP&A portfolio appearing on LinkedIn; verifiable repo: clean → rolling-mean forecast → actual vs budget variance → cost center → Streamlit.
- Actions: First run variance table locally with own actual / budget CSVs; model only writes insights after numbers are calculated in the table.
- Review controls: 0 stars; no ERP connection or posting observed. Treat as architecture reference only; do not connect to production.
- Source: fp-a-finance-analytics-automation | Created 2026-04-08, pushed 2026-09-23 | Nature: repo / portfolio. Corresponding LinkedIn post is discovery seed only, not a case.
Pending Verification Leads
- LinkedIn contains only abstracts, not cross-verified: Gabriel Cavasino (snippet only, not a case). No usable LinkedIn full text this period.
Small Experiments This Week
- Month-end parallel run: Take single-entity prior-month closed TB + close checklist, build Claude project per item 1; output working papers; controller reconciles to TB; log variances in review log. Stop if no improvement in cycle time or reconciliation error rate.
- Collections suggestions without auto-send: Export aging + invoice status with only three columns: “overdue, suggested action, draft”; collections owner selects before any message is sent. Compare weekly collections recovered vs false-positive counts.
- Three-way CSV reconciliation: One file each for orders, gateway, bank; match on amount + date window; produce only exception table and journal draft—do not post. Manually match 20 items; discard rules if mismatch >5%.
- Write metrics before touching tools: Per the Mitch item, select one process with a standard answer (e.g., expense classification of uncleared items), define “48-hour clearance” or “working papers complete by close day-1”; do not start if metrics cannot be written.
- Variance first draft: Calculate P&L actual vs budget inside the table; model only writes driver tree; FP&A owner traces every line back to account. Do not take to operating review if it cannot enter the package.