Trust in the Numbers: What a Firm Should Agree With Each Client Before AI Touches Their Ledger

Technology | November 18, 2026

Trust in the Numbers: What a Firm Should Agree With Each Client Before AI Touches Their Ledger

A firm that can explain plainly what its AI does, what it does not do, and who stands behind the numbers is offering something clients want: confidence that faster does not mean less careful.

Adam Riccoboni

When a firm brings artificial intelligence into bookkeeping, write-up, or the monthly close, most of the preparation happens inside the firm. Partners pick a tool, staff are trained, and someone writes an internal policy. That work matters. But in my experience, the conversation that most often gets skipped is the one with the client.

The client owns the ledger. Their bank feeds, their vendor list, their payroll data, and their customer records are what the AI will read, and in some cases change. If the firm and the client have not agreed in advance what the AI may touch, who checks its work, and what happens when it gets something wrong, the firm is carrying a risk that nobody has priced and nobody has signed for.

I have spent more than a decade building custom AI systems, much of it for the finance function. The technical problems are usually solvable. The problems that cause real trouble tend to be about expectations that were never written down. Here is what I would want agreed with each client before AI goes anywhere near their books.

1. What the AI can see

Start with access, because it is the easiest to get wrong quietly. Many AI tools connect through the same integrations the firm already uses, which means they can inherit far broader access than the task needs. A tool that categorizes bank transactions does not need to read payroll. A tool that drafts the monthly commentary does not need the ability to post journals.

Agree with the client, in writing, which data sources the AI will connect to, and keep that list as short as the work allows. Ask the client whether any data is off limits: some will be relaxed about vendor invoices but uneasy about employee pay or customer details. You will learn something about the client, and they will see that you take their data seriously.

This also sits on top of obligations firms already have. Tax professionals are expected to keep a written information security plan, and the IRS guidance in Publication 4557, Safeguarding Taxpayer Data, is a sensible baseline. If a new AI connection is not reflected in that plan, the plan is out of date.

2. What the AI can change

Reading data and changing data are different levels of trust, and the agreement should say which one applies. I find it helpful to describe three modes to clients:

  • Suggest: The AI proposes a categorization, a match or a draft, and a person accepts or rejects it.
  • Act with review: The AI posts the entry, and a person reviews a sample or an exception report before the period is closed.
  • Act alone: The AI posts and nobody looks unless something breaks.

For most firms and most clients, the first two are where AI belongs today. The third may be reasonable for narrow, repetitive, low-value tasks with clear rules, but it should be a decision the client has agreed to, not one they discover later. Write down which tasks sit in which mode, and agree that nothing moves up a level without a conversation.

3. Who signs off the output

The point I would press hardest is that AI does not sign anything. A person does. Before any AI-assisted work goes to a client, a lender, or a tax authority, a named person at the firm should have reviewed it, and the client should know who that person is.

This sounds obvious, but it changes in practice when the work gets faster. When a close used to take a week, review was built into the rhythm. When a tool produces a draft set of numbers overnight, there is a quiet temptation to treat the draft as finished. The engagement should make clear that the reviewer’s sign-off, not the tool’s output, is what the client is relying on.

It is also worth agreeing what the client signs off. If the AI has categorized transactions the client knows more about than you do, such as a related-party payment or an unusual receipt, the client’s confirmation should be part of the process, not an afterthought.

4. What gets recorded

If a number is questioned six months from now, can the firm show where it came from? With AI in the workflow, that means more than a final ledger. It means being able to see what the tool proposed, what a person changed, who approved it, and when.

Agree with the client what record will be kept and for how long, and make sure your tool can actually produce it before you promise it. Some tools keep a clear trail of suggestions and approvals; others keep only the end result. The NIST AI Risk Management Framework is a useful reference here, not because a small firm needs to adopt it wholesale, but because its emphasis on documentation and accountability gives you a vocabulary for the conversation.

5. What happens when it goes wrong

Every AI system will eventually make a confident mistake. A recurring payment gets miscategorized for three months. A supplier is matched to the wrong entity. The question is not whether this happens but whether the firm and the client have agreed what happens next.

Cover three things. How errors will be spotted, whether through review, reconciliation, or exception reports. How quickly the client will be told. And how corrections will be made and recorded, so that a fixed error does not leave a confusing trail behind it. A client who has been told in advance how mistakes are handled is far more forgiving than one who finds out from their lender.

6. Where the data goes

Clients increasingly ask whether their data is used to train someone else’s model, where it is stored, and who at the vendor can see it. The firm should be able to answer those questions for every tool it uses, and the answers should be in writing. If the vendor’s terms are unclear, that is worth knowing before the client asks, not after.

Putting it in the engagement letter

None of this needs a separate contract. Most of it fits in a short AI section of the engagement letter, or a one-page schedule attached to it, covering access, permitted actions, review and sign-off, records, error handling, and data use. Review it at renewal, and whenever you add a new tool or change how an existing one is used.

The conversation itself has value beyond the paperwork. Clients are hearing a great deal about AI, much of it either breathless or alarming. A firm that can explain plainly what its AI does, what it does not do, and who stands behind the numbers is offering something clients want: confidence that faster does not mean less careful.

In my work, the AI projects that last are rarely the ones with the cleverest technology. They are the ones where everyone agreed, at the start, who was responsible for what. For an accounting firm, that agreement starts with the client.

ABOUT THE AUTHOR:

Adam Riccoboni is founder and CEO of Critical Future, an AI strategy and engineering consultancy with over a decade of custom AI development and a specialism in AI for the finance function. He is the author of The AI Age (2019) and co-author and editor of Engineering Mathematics and Artificial Intelligence (Routledge, 2023). He lectures in artificial intelligence at ESCP Business School and convenes the CFO Council on AI.

Photo credit: Radission US/Unsplash

Sign in to get access to this free resource, and all of our whitepapers and reports.

Download this content today!

Register to get free access to this content, as well as newsletters, continuing education, podcasts, and more…

Leave a Reply