This issue only includes materials that can be broken down to inputs, machine boundaries, and human sign-off. Control design documents are explicitly marked as such and are not production conclusions for your ledger. Appointments, financing, and compute marketing materials were not adopted.
Today’s Most Worthwhile Implementations (2 items)
1. Monthly variances: First break changes down to vouchers; if decomposition is incomplete, mark “unexplained” and do not fabricate reasons
- Scenario: After close, FP&A / accounting prepares flux commentary. The model is most prone to generating fluent commentary even when evidence is insufficient; reviewers may assume analysis has been performed and inadvertently carry fabricated reasons into the management package.
- Actionable steps: This week work on only one account and one comparison period. Inputs = closed account balances + traceable entries / invoices / accrual listings. The model may only output: variance amount, whether it exceeds your materiality threshold, each driver with evidence number, and unexplained residual. Classification is limited to: volume / price / mix / timing / one-time / unknown. Writing “why it rose” is prohibited. Scope changes (reclassifications, chart-of-accounts adjustments) are recorded separately in the comparison-basis notes.
- Review control: Every evidence number must trace to the source record with matching amounts; driver amounts + unexplained = total variance; percentile differences must not be smoothed. Unexplained amounts exceeding the policy threshold (to be defined internally, e.g., 20 %) must be investigated by humans and must not enter the commentary. FP&A writes reasons; Controller samples three items back to vouchers. Preparer and reviewer must not be the same person.
- Deliverables: Variance decomposition table (account / variance / driver / evidence number / unexplained / human explanation required) + human-written reason column. Model output must not serve as commentary text.
- Source: finance-ai-controls (variance decomposition / control design / audit trail) | Source page shows created and pushed 2026-09-15 | Nature: repo / control design document, 0 stars; author states it is a starting point validated in audit scenarios, not an assurance opinion and not a runnable ledger.
2. Accounts payable invoices: Model extracts fields only; authorization, master data, and payments are never delegated to the model
- Scenario: AP converts email / uploaded invoices into an auditable queue. The objective is a shorter exception queue with evidence, not an automatic payment robot.
- Actionable steps: This week select one entity, one currency, and one cohort of legacy vendors. Process five already-paid invoices through a shadow flow and add one dirty invoice (duplicate, PO mismatch, or vendor bank account changed versus master data). The model performs only: retain original, distinguish invoice from statement / receipt, extract vendor / invoice number / date / currency / total / tax / PO / bank account / line items. Validation must use deterministic rules that return named failure reasons (duplicate, PO mismatch, bank account change, arithmetic imbalance, low extraction confidence); do not synthesize a risk score. Payment release is excluded from this week’s pilot.
- Review control: Extraction identity, approval identity, vendor master data, posting, and payment must remain segregated; the same automated account may not both modify master data and release funds. Bank account changes are frozen and verified by an independent person. New vendors, cross-border tax items, fuzzy scans, credit notes, and non-standard contracts route to human review first. Prompts are not financial controls; approval limits and segregation of duties must be enforced in the ledger and permission system.
- Deliverables: Extraction comparison table (face value vs extracted fields vs rule result vs human accept / reject) + exception reason list. Posting is permitted only for invoices where “all rules pass and amount is below the pre-approved threshold”; this week all items are still recommended to stop at human sign-off.
- Source: AgentFlow Atlas “AI Invoice Processing Workflow and Controls” | Corresponding repository source page shows created and pushed 2026-07-21; page publish date not disclosed | Nature: planning manual / educational material, not a production AP system.
Accounting / Close / Controls
See Today’s Most Worthwhile Implementations item 2. Invoice extraction and authorization separated
Closed loop: original → extract fields → deterministic rules supply named reasons → human signs exceptions → only then consider posting. Can be extended to three-way match on duplicates / PO, but do not create a separate process and do not enable auto-payment. Independent revenue recognition new case: Data unavailable.
FP&A / Planning / Reporting
See Today’s Most Worthwhile Implementations item 1. Variance decomposition must carry evidence and unexplained residual
Closed loop: closed balances + source entries → machine only decomposes “what moved” → arithmetic and evidence verification → human writes reasons → unsigned items do not enter commentary. Independent budget preparation and board-package production cases: Data unavailable.
Treasury / Cash / Risk
Data unavailable. This issue contains no verifiable financial-team operational cases on cash forecasting, bank-statement matching, or DSO / collection risk. Payment-gateway demo repositories overlap with recent three-way matching topics and were not opened separately.
Tax / Compliance / Audit
Data unavailable. This issue identified no new AI implementation cases or practical methods for tax research, SOX / internal control, or audit-evidence management within the last 365 days. Item 1 could be extended to “whether a single run can be restored after the model is decommissioned”; this issue does not open that topic separately and does not treat the control design document as a live SOX control.
CFO / Leader Team-Building Experience
Data unavailable. This issue contains no verifiable financial-leader sharing on team division of labor, AI capability building, owners, review mechanisms, or quality metrics. Compute marketing materials and board appointments were not adopted.
Open Source / AI Engineering Lessons
See Today’s Most Worthwhile Implementations item 1. finance-ai-controls
What is reusable is not an agent framework but three hard constraints: model only proposes, deterministic rules dispose; logs must allow reconstruction of the original question text of that run, model version number (no aliases), retrieval context hash, raw output, and whether humans overrode failed checks. Before go-live, randomly sample ten historical runs and answer solely from logs: “input / what was asked / what was returned / who approved”. Missing fields constitute gaps. The repository contains no runtime code and should not be treated as a production controller.
Pending Verification Leads
- Organizational alternative metrics, not independently audited: Deel CEO states internal tool Akai comprises approximately 8,000 agents, approximately 600 person-equivalents, revenue per employee rose from approximately $130k to approximately $215k (company release; recent days already contained payment-reconciliation self-description; this issue does not treat it as new best practice)
- LinkedIn contains only summary, not cross-verified: Roman Murtazin states FP&A agents should first pull P&L / balance sheet
- Additional operator leads: LinkedIn data unavailable / authentication not passed; not recorded as facts.
Small Experiments This Week
- Unexplained residual for one account (Owner: FP&A; Controller samples 3 items)
Follow item 1. Output:
flux_decomp_YYYYMMDD.xlsx. Passing criteria: drivers without evidence numbers are invalid; residual must not equal zero unless entries are exhausted; reason column must be human-written. - Five clean invoices + one dirty invoice (Owner: AP; master-data changes verified by independent person) Follow item 2. Deliberately alter the bank account on one vendor. Output: face value vs extraction vs named rule reason vs human accept / reject. Dirty invoices must not recommend payment.
- Reconstruct 10 runs from logs only (Owner: internal control / accounting) For any existing “model drafts, human reviews” process, sample ten runs using the fields in item 1. If the original question text is missing, model alias is not pinned, or human overrides of failed checks cannot be recorded as query fields, this process must not expand scope this week.