This issue only includes materials that can be decomposed into inputs, human review, and outputs. Template repositories and personal experiments are not treated as audited working papers. LinkedIn contains only summaries without cross-verification, LinkedIn data unavailable / authentication not passed.
Today’s Most Worthwhile Implementations (3 items)
1. Bank/Clearing Reconciliation: Five-Layer Matching, Prohibit Forced Matching
- Scenario: Internal mirror ledger versus bank/clearing statements. Teaching Skill, not a live client ledger.
- Actions: Do not connect to production online banking this week. Desensitize 20–30 transactions and run in original order: amount + value date + reference number must match exactly for auto-clear → truncated reference numbers, leading zeros, and other fuzzy matches must enter today’s exception table → reference matches but date within ±2 business days marked as timing difference → multi-transaction to single must receive human confirmation before clearing → still unmatched items only output investigation hypotheses and are marked unmatched. Unidentified credits must enter suspense within 24 hours per the original record and must not remain in the mirror account.
- Review Controls: Treasury/AR must review all “partial sums” and unmatched items. Forced matching without written justification is prohibited. The USD 5, aging days, and large-amount notification thresholds in the template are bank-specific; replace with your policy before use. Posting and changing collection accounts are closed.
- Outputs: Matched table (by layer/method), unmatched items + hypotheses, today’s exceptions, human accept/return records.
- Source: GitHub: bank-reconciliation Skill (open-source template / non-production sample; source page shows approx. 2026-03 submission, repository has updates since)
2. First Distinguish “Requires Signature” from “Only Requires Attention”, Then Decide Whether the Agent Can Act
- Scenario: Multi-agent orchestration presented as a “work operating system” that was later shut down. Independent laboratory notes, not a finance site.
- Actions: List existing spreadsheets, mailboxes, and cloud-drive assistants in a single table. Ask only two questions for each: does this require my authorization (posting, external data release, master data changes), or only my attention (re-running failed exports)? Authorization items default to write permissions disabled; attention items may auto-retry, but the completed state must be reconciled back to the source system. “Closed period / deployed” statements in Slack/webhooks or prior conversations are never accepted as accounting evidence.
- Review Controls: Controller review: whether anyone treated notifications as general ledger entries; whether any agent released data or changed numbers without signature. Tasks marked closed but with no movement in the close package are treated as idle and the process is stopped.
- Outputs: One-page permission matrix (which ledgers can be read / whether posting or external release is allowed / who must sign / which system is the system of record).
- Source: Danie S Joel: I Built an AI Operations System and Then Turned It Off (operator self-report, not a finance case; published 2026-09-07)
3. Monthly Commentary: Model Produces Only the First Draft; Numbers Presented to Leadership Must Be Human-Reviewed
- Scenario: FP&A monthly report / P&L commentary. Consultant methodology note, not a client working paper.
- Actions: Select only one table that is rewritten repeatedly. Inputs limited to approved actuals versus budget/forecast, plus your materiality thresholds (amount or percentage). Model only: lists differences exceeding thresholds and drafts explanations in the prescribed format. Analysts supplement with contract details, one-off items, and next steps the model does not know. Do not activate forecasting modules or allow the model to revise forecasts this week.
- Review Controls: FP&A owner reviews all numbers to be presented. Differences that do not reconcile may not be smoothed over in commentary. Internal claims that another team reduced commentary time from three days to four hours are not accepted as benchmark without working papers.
- Outputs: Variance list + commentary draft + human sign-off.
- Source: Prime AI Solutions: How to Use AI in FP&A (consultant methodology note with course promotion; page shows published 2026-02-17, updated 2026-07-07)
Accounting / Close / Controls
See Today’s Most Worthwhile Implementations items 1 and 2. bank-reconciliation Skill｜Danie S Joel
Extensible: unidentified credits into suspense, prohibition of unjustified forced matching. Posting rights remain in the existing ERP this week. New close working papers containing account fields: data unavailable.
FP&A / Planning / Reporting
See Today’s Most Worthwhile Implementations item 3. Prime AI Solutions
- Inputs: Approved actuals, single set of budget/forecast.
- AI Processing: Rank differences above threshold, produce first-draft commentary.
- Human Review: FP&A owner determines whether items are one-off or structural; any numbers inconsistent with working papers cause the entire package to be returned.
- Outputs: Variance table + commentary draft + human sign-off.
Independent on-site cases with approved reports and human-review logs: data unavailable.
Treasury / Cash / Risk
See Today’s Most Worthwhile Implementations item 1. bank-reconciliation Skill
Extensible aging: notify account owner in 3–5 days, escalate to Controller after >15 days. These are template days; replace with your treasury policy. Cash forecasting and position scheduling new cases: data unavailable. Item 1 performs matching only; automatic payments and changes to collection accounts are not permitted.
Tax / Compliance / Audit
Data unavailable. No new AI implementation cases or practical methodologies for tax research, SOX/internal control, or audit evidence management within the last 365 days were identified this period.
CFO / Leader Team-Building Experience
Data unavailable. No new enterprise finance leader shares with clear owners, reviewers, and quality metrics were available this period.
Open Source / AI Engineering References
See Today’s Most Worthwhile Implementations item 1. bank-reconciliation Skill
Reusable architecture: deterministic matching → timing differences → many-to-one requires human confirmation → unresolved items enter exceptions with hypotheses. Low-star teaching Skill to be used as a gate template, not as an accuracy benchmark. Do not connect to posting.
Additional: Anthropic’s Finance plugin decomposes period-end close into commands: accrual/deferral/depreciation entry drafts, cash & AR reconciliation, P&L fluctuation analysis, SOX sampling working papers. Requires ERP/data-warehouse MCP connection or pasted files; the repository states it does not provide financial, tax, or audit opinions and that outputs must be reviewed by qualified personnel before entering reports or working papers. Do not connect production keys this week. GitHub: knowledge-work-plugins/finance (official methodology template; source page shows commits as recent as approx. 2026-09-08)
Pending Verification Leads
- Numeric MCP appears only as a LinkedIn summary with no independent full text or client operational records; LinkedIn data unavailable / authentication not passed, not treated as fact.
- Claims that operating profit margin rose from 5% to 70% after acquiring an accounting firm, with daily operations handled by agents and humans shifted to advisory roles. No account-level detail, working papers, or independent cross-verification available; cannot be used as a substitute conclusion. Minerva post (approx. 2026-09; single-source social media / pending verification)
Small Experiments for This Week
- Reconciliation on sample only, no posting: Follow item 1. Desensitize 20–30 bank/clearing transactions versus internal ledger. After rule-based matching, Treasury/AR reviews all unmatched items and “many-to-one” cases. Stop if multiple legitimate combinations are marked matched. Deliverable: matched/unmatched table. Owner: Treasury or AR.
- Split assistant permissions by authorization vs. attention: Follow item 2. List all currently used spreadsheets, mailboxes, and cloud-drive assistants. Disable write permissions this week for any item that can change numbers or release data without human confirmation. Completed state must reconcile to source system; chat logs are not authoritative. Deliverable: one-page permission matrix. Owner: Controller.
- Draft commentary for only one monthly report table: Follow item 3. Select one management table from the most recently closed month. Model produces threshold-exceeding variances and first draft; FP&A owner reviews all numbers to be presented. Stop if inconsistencies are written into commentary. Deliverable: working paper + human sign-off. Owner: FP&A.