The AI tools nobody in finance checked.

Whether AI trains on your company data can depend on the plan the tool sits on. In our reading, the answer sits in that plan’s terms. An opinion piece in Accounting Today, published on 9 November 2026, lists a question accountants should put to their clients: “Are any accounts on consumer or individual plans that allow training on user inputs?” We think many finance teams could not answer it today.

The piece is written for accounting firms looking at their clients. In our view, the same questions apply to any finance lead looking at their own company.

An AI tool can arrive through an expense claim

The article starts with a familiar path. “A staff member signs up for an AI tool, pays with a company card and gets reimbursed.” Each charge “sat below any approval threshold, so no one read the terms or assessed the vendor.” At that moment, as the author puts it, “nobody in finance sees it as a matter for the auditors.”

Then comes what went in. Transcription services, document analysis tools and research assistants have spread across businesses of every size, and “some of these tools now hold contracts, payroll files and draft financial statements.” The article notes that these tools will not appear in an IT asset register or a single sign-on log, because staff usually signed up with personal logins.

The risk is the data, not the tool

We do not think a transcription service is dangerous in itself. The exposure is a payroll file sitting in an account finance does not know exists, under terms nobody read. The tool holds the data. The data is where the risk lies.

The article gives plain examples. A client employee might upload the accountant’s review notes to a free assistant for a quick summary. Many AI notetakers join calls automatically once someone links them to a calendar, and many keep the full recording as well as the written summary. We think the problem is rarely a bad tool. It is a reasonable tool, on the wrong plan, given the wrong document.

Three questions before anything goes in

Before a contract, a payroll export or a draft set of accounts is uploaded anywhere, a finance lead can ask three questions. Ask them of every vendor, including the ones selling to you directly.

Does it train on our data? The answer should be in writing, in the terms of the plan you actually pay for. The article includes this among its questions, and lists moving a tool to a business plan with proper terms as one of the choices for each tool found.

How long is it kept, and where? Ask how long uploads and recordings are retained, in which country they are stored, and who can have them deleted. The article puts the last part directly: “Who can cancel each account and ask the vendor to delete stored data?” We would treat an account nobody can answer for as data nobody manages.

Can each answer be traced to its source document? This question is ours, not the article’s. A summary of a contract that cannot point back to the clause it came from has to be checked from scratch, so it saves little. A system worth trusting with draft accounts should be able to show, for every figure or statement, the document it came from.

Two rules answer two of the three questions

There is a second route. Instead of sending documents out to a subscription, the AI can be built to work inside the company’s own systems, where the data already lives. That is how we build at Syntexia, under two rules: client data never trains a model, and every result points back to its source document. Those rules answer the first and third questions before the build begins.

The second question does not answer itself. How long data is kept, where it sits and who can delete it still need to be agreed and written down, whoever builds the system. Ask your vendor to show you where that is set out.

Finding the tools takes the expenses and a conversation

For tools already in the business, the article starts with the money: “Pull 12 months of card and reimbursement line items rather than summaries.” It also warns where that search stops. “Payments routed through Apple or Google Play hide the vendor’s name completely,” and for those, “You will need to ask the employee who submitted the claim.”

The larger gap is the one with no charge at all. In the author’s words, “Free accounts leave no payment trail, and they can carry the highest risk of all.” An expense review is a start, not a complete list.

For each tool found, the article offers three outcomes: “The client can cancel it, move it to a business plan with proper terms, or keep it with a documented reason.” When raising the subject with clients, its advice is to “Begin with notetakers”. It also asks: “Does the business have a written AI use policy, and when was it last reviewed?”

Where to start

Start with the expense line items, then ask staff about what those cannot show, free accounts included. Read the terms of each plan, and ask the three questions before the next contract goes in. If you are weighing whether to build AI inside your own systems instead, and want to understand how that is done so the data stays where it is, we would be glad to talk it through.