Access review
Reviewers, due dates, segregation of duties, and the evidence pack.
The friend that helps you close. It writes the workpaper. It does not post, pay, or approve go-live.
Reviewers, due dates, segregation of duties, and the evidence pack.
Reason codes, evidence, and the queue. A refund stays a person's click.
Trial-load tie-out, duplicate masters, and the fix list for the next load.
Bank reconciliation, three-way match, expenses, and 1099 readiness.
Setup is in English. The checks read files from the United States, Germany, France, Spain, Brazil, Italy, Japan, China, and Korea, including the encodings those systems write. The finding sentences stay in English. This is not a statement about any country's law.
Matches the statement to the ledger, including CAMT, BAI2, OFX, and MT940.
Matches orders, receipts, and invoices, including partials.
Reads receipt text and ranks what a person should review.
Masked tax ids and who looks ready for a 1099.
Classifies a card dispute and lists the evidence. It does not refund.
Ties a trial load by company and currency before go-live.
Finds likely duplicate vendors and items for a person to merge.
Plans the count and shows the variance.
Compares budget with actual and packs the evidence.
Checks a percentage against the locked number.
Reviewers, due dates, and segregation of duties.
Reads tool descriptions and lockfiles. It does not deploy.
Tracks expiry from the inventory you export.
Maps the files you have to your own control list.
Compares exported state with the approved baseline.
Reads a contract against your playbook. It does not sign.
Applies your retention rules. It does not delete.
Compares a site capture with your vendor list.
Reads lockfiles and flags a weak license for a person.
The role checklist and the offboarding mirror.
Ages the ball-in-court list from the log you export.
Unused seats against the next renewal you can still cancel.
Estimates a conversion from the scripts you point at.
Checks a changelog against the version you name.
Grades recorded judgments against your rubric.
Reads an exported automation for secrets. It does not run it.
Holds a draft that misses a window. It does not send.
Bad addresses, role mailboxes, and withdrawn consent.
Reads DMARC and a saved header.
Contrast, focus, and labels from the static page.
Each one does one job, on your computer, from a file you export. There is no login to your bank, payer, carrier, or property system.
Measured in the final pass, 19 Nov 2026: 7 of the 30 checks
On fictional test data, 19 Nov 2026, for the 7 agents behind the four engagements (bank reconciliation, three-way match, expenses, access review, refunds and chargebacks, migration tie-out, duplicate vendors) of 30. The clean sets were written by the same author as the rules. The 1099 readiness agent is not measured: a clean vendor needs a well-formed taxpayer id, and the test files use invalid placeholders on purpose. The other 23 agents are checked by their eval cases only (596 expected markers, all found). This pass also fixed two bank-file reading bugs found on public files, two file-reading bugs found by the eight-language check, and one place where full tax ids reached the AI model. The test files and the checks were written by the same team, so these numbers show the rules do what they say. How often they are right on your own files is measured at your setup.
git clone https://github.com/fpavia111-creator/fairbooks-agents
cd fairbooks-agents
npm install
node packages/fairbooks-bank-reconciliation/bin/fairbooks-bank-reconciliation.js demo
Open source, MIT licence, on GitHub: fpavia111-creator/fairbooks-agents. Needs Node.js 20 or later. The demo runs on made-up files and needs no network and no setup.
| Source | What ran and what it found | Licence |
|---|---|---|
| PostFinance sample files for business customers (camt.053) 2026-11-19 | fairbooks-bank-reconciliation read all 4 statements; lines, totals and balances equal a separate re-count. In 2 of them the booked closing balance leaves out a 1,537.00 credit value-dated the next day; the agent now says so instead of only "does not tie". | Published by PostFinance as sample files for software tests (no licence stated) |
| ofxparse test files (OFX bank and card statements) 2026-11-19 | All 8 statements read; line counts, totals and ledger balances equal a separate re-count. No wrong results. | MIT |
| bai2 test files (Citi, NatWest and SVB layouts) 2026-11-19 | All 4 files read and equal a re-count. The SVB file holds two days for one account and did not tie because the reader took the second day's opening balance; fixed, it ties. The Citi sample's own balances do not add up, which the agent reports. | MIT |
| SEC EDGAR XBRL company concept data (Apple, Microsoft, Walmart, Alphabet) 2026-11-19 | fairbooks-erp-migration compared each period's first-filed value with its last-filed value: 448 periods, 26 reported differently later (for example Apple's fiscal 2008 net income, 4,834 then 6,119 million). The agent and a re-count list the same periods in all five series. | U.S. government data, public domain |
Public, lawful files only; no private data was used. The files are not copied into the repository: a script there downloads them again.
| Package | What a person reviews and decides |
|---|---|
| Access review | Each system owner decides keep or revoke for every flagged account; revoking access is a one-click request a named person approves, and the agents never change access themselves. |
| Dispute packs | A person reviews the reason code, the evidence list and the response draft before anything is submitted; a refund is always a person's click. |
| Migration readiness | The controller and the implementation lead decide whether untied totals, missing records and duplicate masters block go-live; go-live approval is never given by an agent. |
| Month-end | The accountant reviews matches, reconciling items, match exceptions, expense recommendations and the 1099 list, and posts any adjusting entry in the books; nothing is posted, paid or filed by the agents. |
| System | Report or export | Status |
|---|---|---|
| Okta | Users API JSON (GET /api/v1/users) | Not yet tested on a real export |
| Microsoft Entra ID | Users > Download users (CSV) | Not yet tested on a real export |
| Microsoft 365 admin center | Active users > Export users (CSV) | Not yet tested on a real export |
| Microsoft 365 usage reports | Active users report detail (CSV) | Not yet tested on a real export |
| Google Workspace Admin | Users > Download users (CSV) | Not yet tested on a real export |
| Slack | Manage members > Export member list (CSV) | Not yet tested on a real export |
| GitHub | Organization People > Export (CSV or JSON) | Not yet tested on a real export |
| GitHub Copilot | Seat assignments API JSON (GET /orgs/{org}/copilot/billing/seats) | Not yet tested on a real export |
| AWS IAM | Credential report (CSV) | Not yet tested on a real export |
| Atlassian Administration | Directory > Users > Export users (CSV) | Not yet tested on a real export |
| BambooHR | Employee report (CSV or Excel) | Not yet tested on a real export |
| NetSuite | Vendor saved search (default column labels) | Not yet tested on a real export |
| NetSuite | Transaction saved search (default column labels) | Not yet tested on a real export |
| NetSuite | Item saved search with location quantities | Not yet tested on a real export |
| QuickBooks Online | Vendor Contact List report (CSV or Excel) | Not yet tested on a real export |
| QuickBooks Online | Transaction List by Vendor report (CSV or Excel) | Not yet tested on a real export |
| QuickBooks Online | Transaction Detail by Account / General Ledger report | Not yet tested on a real export |
| Xero | Contacts > Export (CSV) | Not yet tested on a real export |
| Xero | Bills > Export (CSV) | Not yet tested on a real export |
| Xero | Account Transactions report (Excel or CSV) | Not yet tested on a real export |
| Xero | Bank statement CSV (Xero import layout) | Not yet tested on a real export |
| Stripe | Payments > Export (CSV) | Not yet tested on a real export |
| Shopify | Orders > Export (CSV) | Not yet tested on a real export |
| Shopify | Products > Inventory > Export (CSV) | Not yet tested on a real export |
| HubSpot | Contacts > Export (CSV or Excel) | Not yet tested on a real export |
| Salesforce | Contacts or Leads report export | Not yet tested on a real export |
| Salesforce | Data Loader / Data Export of Contact (API field names) | Not yet tested on a real export |
| Mailchimp | Audience > Export audience (CSV) | Not yet tested on a real export |
| Procore | RFIs > Export > CSV | Not yet tested on a real export |
| Ramp | Transactions > Export (CSV) | Not yet tested on a real export |
| Expensify | Default CSV export of expenses | Not yet tested on a real export |
| Chase | Checking or savings account activity (CSV) | Not yet tested on a real export |
| Chase | Credit card activity (CSV) | Not yet tested on a real export |
| Bank of America | Checking account activity (CSV, with its summary block) | Not yet tested on a real export |
| Wells Fargo | Account activity download (CSV with no header row) | Not yet tested on a real export |
| Capital One | 360 account transactions (CSV) | Not yet tested on a real export |
| Mercury | Transactions export (CSV) | Not yet tested on a real export |
| AWS Certificate Manager | aws acm list-certificates JSON | Not yet tested on a real export |
| crt.sh | Certificate search JSON (output=json) | Not yet tested on a real export |
| ISO 20022 bank statement | camt.053 (PostFinance published samples, 2009, 2013 and 2019 versions) | Tested on a public sample |
| OFX bank and card statements | OFX 1.x and 2.x statement download | Tested on a public sample |
| BAI2 cash management file | BAI2 prior-day file (Citi, NatWest and SVB layouts) | Tested on a public sample |
| SWIFT MT940 statement | MT940 customer statement | Not yet tested on a real export |
| SEC EDGAR XBRL | Company concept data, converted to the migration compare columns by scripts/realdata.mjs | Tested on a public sample |
No layout has yet been tested on a customer's own export. Every setup starts by confirming the layout on your files, and any column can be mapped by hand.
When an assistant calls a tool, its model sees the result: amounts, dates, ids, finding codes and text from your files, including people's names and e-mail addresses (Fairbooks has no general masking, unlike the newer Fair families). Masked: taxpayer ids (last four digits), secrets found in configuration and automation exports, card numbers in receipt text, and the person in an erasure request. The report and draft files written on your computer hold the full values. Results go to whatever model your assistant uses; for staff, customer and vendor data, use a model covered by your own agreements or a local model.
Rules that change are settings or dated files, and a result names what it used: the 1099 threshold by tax year (600 before 2026, 2,000 in 2026, later years marked for confirmation), taxpayer id patterns, the cookie and tracker vendor list (version 2026.11.0), the MCP package advisory snapshot (2026-11-15), license text fingerprints and certificate strength rules. Expense, refund, tolerance, retention and access policies are your own JSON files, checked against a schema. You edit those files yourself, or support updates the bundled lists when a rule or a vendor changes.
We send the list of reports by name. You export them from the systems you already use; nothing is connected.
Over a shared screen: the checks are installed, your columns are matched, and your own rules and dates are set.
The checks run on your files. We go through the worklist with you, and every flagged row is a person's call.
You run them whenever you like. Support keeps the dated rules and the export layouts current.
A fixed scope and a fixed price, agreed before any work starts. Your files stay on your computer the whole time.
Written by the check itself on made-up files. Every person, business and number is invented.
Access review evidence All samples