Accounting firms can license de-identified ledgers, reconciliation history and close procedures to AI labs, which are training models to do bookkeeping and month-end close. Most of that data belongs to clients, though, so professional confidentiality rules and, for tax return information, IRC section 7216 usually require each client's specific consent before any of it is used.
This is general information, not legal or tax advice. Talk to a lawyer who knows professional responsibility rules before you license client data.
What labs want from accounting firms
A firm with many clients over many years holds something no public source has: real books, kept by people, with every correction recorded. The data labs ask for falls into three groups.
- Ledgers and journals. Real charts of accounts, categorizations and adjusting entries across many clients and years.
- Reconciliation history. How mismatches were found and fixed. The audit trail, such as the Xero audit log, shows each decision in order, and labs value it most.
- Close checklists and procedures. How the firm actually runs month-end, step by step, including what gets checked and in what order.
Reconciliation history can become episodes: a mismatch, the action a bookkeeper took and the outcome. Episodes can be graded, which makes them more useful to a lab than a static ledger. The accounting industry page shows a sample record before and after scrubbing.
Why client consent comes first
An accounting firm usually holds client data in a position of trust, and three sets of rules apply to it.
IRC section 7216 (tax return information)
US federal law makes it a crime for a tax return preparer to knowingly or recklessly disclose tax return information, or use it for anything other than preparing returns, unless an exception applies. The penalty under 26 U.S.C. 7216 is a fine of up to $1,000, up to one year in prison, or both. A separate civil penalty under section 6713 is $250 for each unauthorized disclosure or use, up to $10,000 a year.
The regulations allow disclosure with the taxpayer's consent, but Treasury Regulation 301.7216-3 requires that consent to be in writing and signed before the disclosure or use, with specific content and format rules. An engagement letter clause or a line in a privacy policy is unlikely to meet them. The IRS section 7216 information center collects the guidance.
In practice, tax return information is the hardest category to license, and many firms will simply leave it out.
AICPA confidential client information rule
For CPAs in public practice, rule 1.700.001 of the AICPA Code of Professional Conduct says a member shall not disclose confidential client information without the client's specific consent. It covers all confidential client information, so bookkeeping and advisory work fall under it alongside tax.
ICAEW confidentiality (UK)
ICAEW members must keep information gained through professional work confidential. Its helpsheet on disclosing confidential information covers when disclosure is allowed, and client authorization is the usual route. UK GDPR also applies wherever the records contain personal data, such as sole traders' accounts or payroll.
Does de-identification remove the need for consent?
Not reliably. Professional rules focus on whether you disclose a client's information, and scrubbing doesn't change whose information it was. Whether de-identified data still counts as "confidential client information" is a question for your own lawyer, and the cautious answer is to get consent.
De-identification still matters a great deal, for two reasons. It protects the clients whose consent you obtain, and it is what privacy law looks at: data that is truly anonymous falls outside the GDPR under Recital 26. Removing names is not enough to reach that standard. The de-identification vs anonymization guide explains why.
How firms get consent
Firms that license client data usually ask for it explicitly, client by client. A workable approach:
- Pick the clients and years you would include. Leave out tax return information unless you are prepared to meet the 7216 consent rules for it.
- Write a short, specific consent: what data, for what use (training AI models), which kinds of buyers, for how long, and that the data is de-identified before it leaves.
- Collect signed consent before any data from that client is included.
- Keep records of every consent, and honor withdrawals for future batches. Be plain that data already licensed can't be pulled out of a trained model.
Clients who decline are excluded. A dataset built from fewer clients with clean consent is worth more to a lab than a bigger one with doubtful rights, because labs check rights in diligence. Our calculator reflects this: answering "not sure" on rights lowers the estimate.
How accounting data gets scrubbed
Scrubbing runs on your machine, and raw data never leaves it through our SDK. For ledgers and invoices it:
| Field | What happens |
|---|---|
| Client and supplier names | Replaced with consistent pseudonyms such as ORG_7F3A |
| Emails, phone numbers, addresses | Removed |
| Bank and account numbers | Removed |
| Amounts | Rounded into ranges, so £14,382.17 becomes £14k-15k |
| Dates | Generalized to the month |
| Invoice numbers | Replaced, because a number plus an amount can identify a transaction |
| Free-text notes | Dropped unless reviewed |
Unusual amounts plus dates are a known re-identification risk in financial data. After scrubbing, we test whether any record can be singled out from a combination of fields and remove the ones that can. You review the scrub report before anything goes to a buyer.
What accounting data sells for
There is no public price list for accounting ledgers licensed for AI training, and we won't quote one we can't source. Our calculator gives an indicative range based on the systems you use, years of records, team size and how clear your rights are. The range is a starting point that we confirm on a call. What AI labs pay for accounting data collects the price signals that do exist.
The recurring part
Books never stop. After the first license sells the history, the SDK can run monthly, pull only new and changed entries since the last run, scrub them on your machine and upload a new batch. Every license with a refresh term pays per accepted batch, and the SDK needs no work from you once the schedule is set. A growing client base means bigger batches each month. Recurring data revenue explains the model.
See what your ledgers could earn
Run the valuation calculator with "Accounting (Xero, QuickBooks)" selected. It needs no access to your systems, and we check your engagement letters and consents on the call. For the full process, see selling data to AI companies.
Frequently asked questions
Can an accounting firm sell client data at all?
Only with the right consents. Tax return information falls under IRC 7216, which requires specific written consent signed in advance. CPAs in public practice need specific client consent under AICPA rule 1.700.001, and ICAEW members have similar duties.
Does removing client names make consent unnecessary?
You shouldn't assume so. Professional confidentiality rules look at whose information it is, and a lawyer should decide whether de-identified data still counts as confidential client information. Getting consent is the safe course.
What about the firm's own data?
Your own procedures, close checklists and internal workflows belong to the firm, so the client consent issue doesn't arise for them, though employee data inside them still needs scrubbing. Procedures describe how work is done, and labs value them.
What if a client withdraws consent later?
Their data is left out of future batches. Data already delivered under a license can't be removed from a model that has been trained on it, so the consent form should say so plainly.