How records are anonymized and reviewed before AI use, and what to ask before you share any

Your clients trusted you with their numbers, their paperwork and sometimes their worries. So when a buyer says it removes personal information from business records before AI developers work with them, the right reaction for an accounting firm is a polite “how, exactly?”. This article explains what anonymization and review mean in practice, where they can fall short, and what to ask.

We are not lawyers, and this is not legal or tax advice. Whether your firm may share any material is a question for your own advisor and your professional body.

What anonymization means

Anonymization, often called de-identification, means taking identifying information out of a record so that it can no longer be linked to a particular person or client. The US National Institute of Standards and Technology describes it as removing identifying information from a dataset so that individual data cannot be linked with specific individuals, and adds that it can be done by people, by software or by both.

In business records, the usual targets are:

• Names of clients, contacts and staff
• Email addresses, phone numbers and street addresses
• Account, invoice, case and tax identification numbers
• Company names and other details that point to one client
• Free-text mentions inside emails, notes and documents

What it looks like

On our home page we show a made-up example. A line such as a thank-you email to a named client, with a phone number and an invoice number, becomes the same sentence with placeholders such as [CLIENT CONTACT], [PHONE] and [INVOICE NUMBER]. The structure and the meaning of the message stay. The identifying details go. The example is illustrative text, not a real client record.

The usual stages

Buyers describe their own process, so ask for theirs. In general, preparing records for AI use tends to involve these stages:

1. Scoping at your firm. You decide which kinds of material could be included, and which are off-limits. This is the strongest protection, because material that is never included is never at risk.
2. Automated detection. Software looks for names, numbers, addresses and similar patterns, and masks or replaces them.
3. Human review. People check samples or flagged items, because software misses things.
4. Quality checks. Someone confirms that the result is still useful and that identifying details do not remain.
5. Controlled use. The buyer’s agreement sets out who may see the material, how it is stored and what it may be used for.

Whether a given buyer does all of this, and how, is something to ask and to have in writing.

Where anonymization can fall short

Honest answers include the limits.

Removing a name is not always enough. The UK Information Commissioner’s Office explains that a person can be identifiable from many factors, not just a name, and that information is only anonymous if no means reasonably likely to be used could identify someone. Data that has had direct identifiers removed but could be linked back to a person may still count as personal data.

Free text is hard. Emails and notes describe situations in ordinary language. A unique story, an unusual job title or a small town can point to a single client without a name appearing.

Software makes mistakes. NIST notes that de-identified data can be manually reviewed or audited to check whether identifying information remains. That check exists for a reason, since researchers have shown that some de-identified data can sometimes be re-identified.

None of this means anonymization is pointless. It means you should treat “we remove personal information” as a claim to examine, not a guarantee.

What to ask a buyer

Ask for the answers in writing, and keep them:

1. What exactly is removed or masked, and who decides?
2. Is it done by software, by people or both? How is the result checked?
3. Does anyone see the original before it is cleaned, and who?
4. How do you test whether someone could be re-identified?
5. What happens if something is missed, and how quickly will we be told?
6. Where is the material stored, who can see it, and for how long?
7. What may the buyer do with it, and what may it not do?
8. Can the arrangement end, and what happens to the material afterwards?

If a buyer cannot explain its process in plain words, treat that as an answer.

What to check at your own firm

Before you ask a buyer anything, check what binds your firm:

• Engagement letters. What do they promise clients about confidentiality and third parties?
• Professional rules. In the United States, for example, the AICPA Code of Professional Conduct has a Confidential Client Information Rule, which in general terms asks members in public practice to protect confidential client information and to get the client’s specific consent before disclosing it, with limited exceptions.
• The law where you work. US tax return preparers are subject to Internal Revenue Code section 7216, which the IRS explains in general terms as limiting the use and disclosure of tax return information unless an exception applies or the taxpayer gives a written consent that meets specific requirements. Other countries have their own privacy and data protection rules.

These details are for your advisor. A good habit is to write down what you were told, and by whom.

What we do and do not do

At Thomas & Finch, we do not receive your files, and we do not ask for them. The fit check runs in your browser, and the details you type into our request form reach us by email, as described in our privacy policy. If you apply, you do so through the buyer’s own process, the buyer reviews your firm, and acceptance is never guaranteed.

We are paid by the buyer, and only if an accepted firm came through our referral. We would rather tell you honestly that your firm is not a fit, or that the privacy questions are not settled, than push anyone forward.

Where to start

If you would like to see whether your firm looks like it meets the usual criteria, take the free fit check. To understand the bigger picture first, read how AI training data collection works.