This issue only includes materials that can be decomposed into inputs, actions, review, and outputs. LinkedIn data unavailable / authentication not passed and is not treated as a factual source.
Today’s Most Actionable Implementations (3 items)
1. Variance Explanation: Scan GL memos first, then external events, and end with a single “because” sentence
- Scenario: Expense variances in monthly reports / board packs. This is a public tutorial; the demo company is a fictional retail ledger, not a production FP&A reconciliation from any specific company.
- Actions this week: Take only one closed month and one material expense line (tutorial example: Maintenance & Repairs +40% MoM). Input = GL export for that month (including memos) + comparison month. Layer 1: Have the model semantically classify memos only (e.g., HVAC, utilities, carpentry), summarize by region, and highlight the primary driver; the human reads this section first. Layer 2: Only then allow lookup of external events (weather, shutdowns, policy), and they must align with the expense occurrence date. Synthesize a single “because” sentence suitable for the board; do not write “expenses rose because more repairs were made.”
- Review controls: FP&A owner validates the numbers; business owner validates external causes. If memos do not match dates or external events do not match the occurrence week, the entire sentence is invalidated. Source amounts must not be altered to “align the narrative.” This week’s output is not the final board version.
- Outputs: Thematic classification table for the account, primary driver, external event cross-reference, one “because” sentence, human edit log.
- Source: YouTube | Simon Sez IT, Excel Two-Layer Variance Analysis (tutorial/demo; 2026-07-29)
2. Budget Import: AI only reads layout; totals must reconcile via code; unmapped accounts go to a human
- Scenario: Budgets remain in Excel while actuals sit in the ledger; monthly manual paste creates BvA. This is vendor product documentation, not an independent client case study.
- Actions this week: Do not change the ledger this week. Take an approved annual/quarterly budget file (multiple tabs or a single annual column are acceptable). Pilot rules copy these three steps: the model only identifies headers, periods, and subtotal rows; amounts are read via deterministic code and reconciled to the workbook’s own subtotals/totals; rows are mapped to the chart of accounts (including account numbers); unmapped rows are flagged for human review—guessing is prohibited. Files that do not balance must not enter “current version.” Versions are appended only, never overwritten; current BvA references only the latest approved version.
- Review controls: FP&A signs off on mapping and version naming. Imported figures must trace back to source cells. Dimensional budgets (department/location), in-system budget amount changes, and scenario/approval workflows: the vendor states these are not provided—do not assume they exist.
- Outputs: Mapping exception list, reconciliation results (import total vs. workbook total), named budget version, current-period BvA (internal use only).
- Source: Digits | Budgets documentation (vendor material; 2026-11-10)
3. Payments / Vendor Master Data: Before an Agent can change a payee, it must prove “who did it, under which authority, and who approved it”
- Scenario: Payment integrity. Kyriba CIO/CISO position: for finance, an altered payment path is more damaging than a one-time breach. When an Agent changes a payee or proposes a disbursement, identity boundaries encompass the Agent, token, connector, and human approval chain. This is a control framework, not a post-incident review from any treasury team.
- Actions this week: Select one category of sensitive action (change supplier bank account or submit payment proposal). Conduct a tabletop exercise simulating “an auditor asks about an unexplained payment”: can the team immediately produce identity, permissions, approval, and business purpose? If the answer cannot be produced on the spot, stop automation—do not deploy the Agent. Dual control, least privilege, and sensitive actions must carry named individuals.
- Review controls: CISO owns security execution; CFO / Treasury owner owns cash exposure. When control evidence is insufficient, the person authorized to stop the process must be named. Agents may not self-propose limits, add payees, or change policy.
- Outputs: Evidence-field checklist for the action category, named stop-process authority, tabletop exercise record. Zero payments on the books.
- Source: Kyriba | Kevin Bailey, Payment Integrity (vendor control framework; release date not disclosed; cites 2026 Risk Radar)
Accounting / Close / Controls
Revenue Contracts: Checklist against policy; rules may rerun only after human sign-off (vendor material)
- Scenario: Complex contracts, usage-based billing, mid-term price changes. The vendor breaks the workflow into: contract terms mapped to ASC 606 policy checklist → human sign-off → deterministic engine reruns rules → source system / engine / GL reconciliation with explanations → draft variance commentary. Self-described as “no posting unless a human signs.” Labor-savings figures are vendor-provided and not used as a preparation basis.
- Actions this week: Prohibit posting this week. Sample one archived contract and build a checklist against the company’s revenue policy (performance obligations, variable consideration, modifications). The model outputs only: hit / gap / judgment required. Only after human sign-off is rule-engine ingestion discussed.
- Review controls: Revenue accounting owner signs off on policy hits. Every conclusion must open to the contract clause and the rule. Missing prices, anomalous allocations, or system-GL mismatches are treated as exceptions.
- Outputs: Contract checklist, exceptions, human sign-off. Zero entries on the books.
- Source: RightRev AI product page (vendor; date not stated)
Bank-statement line-by-line reconciliation prior to import: see Open Source section. New production cases for standalone month-end reconciliation packages: Data unavailable.
FP&A / Planning / Reporting
See Today’s Most Actionable Implementations item 1 (GL memo → external event → single “because” sentence) and item 2 (original budget file import; unbalanced files do not become current version).
Rolling forecast / full-company board-pack new production cases: Data unavailable. Can be extended to Wellhub’s multi-model comparison (see Leader section); do not replace point forecasts with “AI says one number” this week.
Treasury / Cash / Risk
See Today’s Most Actionable Implementations item 3.
Gorgias treats collections / billing as a growth asset: first map manual control points (sales manually completes contract → manual billing entry → manual collections), then discuss tooling. This is system redesign, not a payment Agent. New automated disbursement cases: Data unavailable. See Leader section; no separate item opened.
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 past 365 days were identified this period.
CFO / Leader Team-Building Experience
1. Wellhub: When forecasts are inaccurate, run parallel models; reject black boxes
- Speaker: Bruno Annicq, Wellhub (formerly Gympass) CFO; finance team of approximately 100. Interview dated 2026-01-15.
- Approach: Marketplace business point-forecast error reduced from roughly 10% to approximately 2%. He emphasizes the improvement is not model-driven alone: the team also strengthened. The method is to engage two vendors to build separate models, now running four in parallel, observing where they converge and diverge, then applying financial judgment. When hiring, require candidates to deliver maintainable Python models and train internal staff rather than delivering results only. Models that cannot be explained are used to flag anomalous dimensions (including fraud), not to make business decisions.
- Organization: EMPOWER hard rule: each item has only one accountable owner; OKRs are company priorities while finance additionally tracks cash. Expense approval should be pushed to teams—when the CFO is still approving routine expenses, delegation has failed.
- Source: YouTube | Run the Numbers / Bruno Annicq (operator interview; 2026-01-15)
2. Gorgias: O2C first requires manual touchpoint disassembly; internal Q&A tools must trace back to source data
- Speaker: Kunal Agarwal, Gorgias’ first CFO. Interview dated 2026-03-26.
- Approach: Upon joining, first took ownership of O2C and removed “sales manually completes contract → another person manually enters billing → manual collections.” At approximately $25M ARR he asks: how many manual input points exist, and where will the process break at 10× scale? Unstable product catalog causes repeated billing implementation failures. LLM costs are reviewed daily; simple tickets use cheaper models, complex tickets use expensive models. Pricing is outcome-based (“charge per resolved ticket”); unresolved tickets are not billed to the customer—costs still occur, so resolution rate must be tracked.
- Internal tool Cortex (built by data/analytics team as decision layer): CRM, finance, board materials, and Slack feed into the same layer. Example: for a customer cohort with churn in the last six months, the tool reads call transcripts and usage then generates operational recommendations. Self-reported adoption by approximately 70% of employees. External numbers still require human narrative adjustment.
- Source: YouTube | Run the Numbers / Kunal Agarwal (operator interview; 2026-03-26)
Named finance leaders with headcount and finance ROI comparisons on new career pages: Data unavailable. LinkedIn data unavailable / authentication not passed.
Open Source / AI Engineering Reference
Bank-statement CSV: Run rolling balance first, then discuss ledger import
- Reusable architecture (reference script, not a client production system): After PDF / converter produces CSV, reconcile opening balance + line-by-line activity against the printed bank statement closing balance. Stop at the row where the rolling balance breaks. Difference exactly twice a line amount → likely sign reversal; difference exactly one transaction amount → likely missing line (common at page breaks); difference divisible by 9 → likely transposed digits. Exit code 1 can be wired into batch processing to block bad files.
- Suitable pilot: One closed month, desensitized bank-statement CSV. Do not import or reconcile this week.
- Note: Low star count; vendor-adjacent repository. Sample data is fictional 2026-09 activity. Reconciliation success does not equal “classification correct.”
- Source: GitHub | statement-balance-check (repo; page shows update circa 2026-11-04)
Leads to Validate
- LinkedIn data unavailable / authentication not passed. No new, cross-verifiable startup finance headcount-reduction career pages.
- Close Agent and billing Agent videos are mostly titles or product narration only and are not elevated to cases.
Small Experiments This Week
-
Two-layer “because” for a single expense line (Owner: FP&A; business signs external cause)
Input = prior-month closed GL (with memos) + comparison month. First classify memos; human reads before external events are checked. Gate: events must match occurrence week; mismatch invalidates the sentence. -
Budget original file only does reconciliation and mapping (Owner: FP&A)
Input = approved budget Excel. Model only flags layout; code reconciles subtotals; unmapped accounts go to exceptions. Gate: until exceptions are cleared, the file may not be treated as “current budget version.” -
Tabletop exercise for changing supplier bank details (Owner: Treasury; CISO accompanies)
Do not connect to production payments. Question: can identity, permissions, approval, and business purpose be produced on the spot? Gate: if any item is missing, keep the action category manual. -
One contract only builds policy checklist (Owner: Revenue Accounting)
Input = archived contract + company revenue policy. Output only hits / gaps. Gate: human sign-off; zero entries on the books. -
One bank CSV first runs rolling balance (Owner: Accounting)
Input = desensitized statement. Gate: rolling balance stops at the break row; file may not be imported until it passes.