This issue retains only materials that can clearly document inputs, posting gates, and signatories. LinkedIn data unavailable/authentication not passed; no cross-verifiable startup financial close alternative cases. Both public repositories below are sandbox/synthetic data, not a company’s on-site working papers.
Today’s Most Worthwhile Implementations (2 items)
1. AP Posting: Pass Five Gates First, Amounts Above Threshold Require Human Review, Prohibit Silent Duplicate Posting
- Scenario: Small and medium teams / accounting handles supplier bills themselves. Invoices come in from OCR or shared folders and need to be posted to QuickBooks payables. The goal is not to hire another full-time AP clerk, but to codify “whether posting is allowed” as verifiable rules.
- Actions: Do not implement full automation yet. Take 5 already-posted bills and run them through the counterparty’s five gates: whether debit/credit positions align; whether the account is in the active chart of accounts; whether the period is still open; whether the submitter has permission for that account; whether an audit log entry for review has been written before posting. Amounts below your own materiality threshold may be posted automatically; amounts above the threshold only trigger approval, without writing to the general ledger. If the same invoice number appears again, must return the original bill number; do not post a second entry.
- Review Controls: Any gate fails: do not write to QuickBooks, log the failure reason. After approver clicks approve, log changes from “Pending” to “Posted”, recording who clicked. These are the five paths the ACA author ran successfully in the QBO sandbox (small-amount phone bills auto-post, large legal advisor fees via Slack, duplicate submission idempotency, account no permission, account does not exist), not a production client ledger.
- Deliverables: Five-gate checklist; pass/fail reason for each bill; approval screenshot or signature; duplicate submission records that were blocked from posting.
- Source: GitHub: tabii-dev/accounting-automation-portfolio (ACA author’s n8n composition repo / sandbox evidence; source page shows latest push 2026-08-31) The same repo can also be extended to: bank CSV → first match invoice/bill, if no match then classify by chart of accounts → Slack only provides suggestions, human still confirms on Xero reconciliation page; and splitting Stripe/Shopify payouts into revenue/refund/fees before posting. Do not run three parallel workstreams this week.
2. Tax Working Papers: Model Only Performs Reconciliation Steps, Tax Rates and Filings Still Reviewed by Humans Against Current Regulations
- Scenario: Tax / accounting prepares BAS or VAT control account working papers. Input is ledger export, not asking the model to “search tax rules” online.
- Actions: Pull three exports from a closed filing period: Activity Statement, trial balance at period end, GST/VAT control account detail. Give the model only this instruction: prepare working papers by cash/accrual and filing cycle, reconcile output tax minus input tax to the control account’s current period movement, list exceptions, do not force balance. Human reviews the reconciliation, then decides whether to proceed to filing process.
- Review Controls: The repo explicitly states: the skill encodes steps and exceptions, not tax rate tables; tax rates, thresholds, deadlines must be checked against current regulations. Do not file, do not substitute judgment. Do not put client data into public repo; use demo or anonymized files first. Low star count; treat as step checklist, not a production tax filing system.
- Deliverables: One reconciliation working paper (control account movement vs. return net amount); exception list; tax owner sign-off section.
- Source: GitHub: ryanduguid/australian-accounting-skills (Australian practicing accountant’s Claude/Codex skill package / open-source workflow; source page shows latest push 2026-09-01; explicitly states Not lodgment)
Accounting / Close / Controls
See item 1 under Today’s Most Worthwhile Implementations. Do not implement “smart month-end close” this week. First incorporate the five posting gates into the existing payables process: posting authority remains with the ERP; model or n8n at most drafts, cannot close the books itself.
FP&A / Planning / Reporting
Data unavailable. No new budget, forecast, or variance analysis workflows this period; do not treat generic “use Claude for dashboards” as a pilotable method.
Treasury / Cash / Risk
See the extensible portion of item 1 under Today’s Most Worthwhile Implementations. Payment channel payouts should first be split into reconcilable entries; bank reconciliation still requires human confirmation. Do not combine payout splitting and bank reconciliation into a single automatic posting this week.
Tax / Compliance / Audit
See item 2 under Today’s Most Worthwhile Implementations. The only reusable pattern: export → reconcile → list exceptions → human sign-off; filing authority remains closed.
CFO / Leader Team Building Experience
Data unavailable. No new finance leader public shares this period that clearly document division of responsibilities, review mechanisms, and measurement metrics. Do not pad with compensation surveys or infrastructure press releases.
Open Source / AI Engineering References
1. Multi-channel Receipt Reconciliation: Put Matching Rules into Testable Code, n8n Only Handles Pull and Notifications
- Scenario: Stripe / PayPal / Square / ACH coexist with bank deposits, fees, and refunds; accounting needs to produce unreconciled items rather than treating “match score” as the amount.
- Actions: Architecture worth copying: channel pull → standardize → five-round matching (exact/reference number → date and fee tolerance → refund association → split posting → below threshold go to exception). n8n handles scheduling and alerts; amount rules in standalone Python with tests. QuickBooks posting must have manual approval; missing debit/credit or unbalanced directly rejected.
- Review Controls: Author explicitly states this is a reference implementation, not a verified client deployment; synthetic benchmark is 6/6 expected matches, 0 false matches, 7 unresolved items — cannot be used as evidence of time savings. Unreconciled items must have an owner; prohibit writing match scores into journal entry amounts.
- Deliverables: Matched/exception table; unbalanced entry rejection log; manual approval records.
- Source: GitHub: Etherlabs-dev/multi-processor-reconciliation (open-source reference architecture / synthetic data; source page shows created 2026-02, latest push 2026-08-09)
Leads Requiring Verification
- A weekly report claims the community has seen “month-end close that can run to completion autonomously”; no chart of accounts, matching rules, or signatory visible — cannot be used as a close template. X: TheRundownAI (around 2026-09; single-source social media / to be verified)
Small Experiments for This Week
- Run Five Payables Bills Through Gates, Do Not Post: Select 5 bills already posted last month. Accounting checks off against the five gates; any above your materiality threshold must include approver name. Controller reviews failure reasons. Deliverable: one-page gate checklist. No one may modify the ERP.
- One Bank Statement Produces Suggestions Only: Export last month’s bank statement for one cash account. Classify only into: matched invoices / suggested accounts / requires investigation. Accountant samples 10 “requires investigation” items; if misclassifications exceed 3, revise rules first. Deliverable: three-column unreconciled items table. Reconciliation confirmation remains clicked inside the existing ledger.
- One Tax Control Account Performs Reconciliation Only: Pull control account detail + return net amount for the most recent filed period. Model only lists reconciliation differences and exceptions; must not provide filing conclusions. Tax owner reviews 5 items. If they do not match, discard. Deliverable: reconciliation working paper + sign-off.