Where AI shows up in client work, and how to keep it from going wrong

Most of the conversation about AI in accounting is about back-office automation: data entry, bank feeds, categorizing transactions at scale. This guide is about something else. It is about AI doing the client work itself: the forecasting, the analysis, the commentary, the answers to client questions.

In the firms we talk to, that kind of AI shows up in two places. Most partners are already in the first, often without calling it a strategy. The second is one most firms want and are wary of, and the wariness is justified.

Both places need two things. The AI needs context about the client, because without it the answers are confidently wrong. And every answer needs its lineage: where it came from, which numbers it used, what it assumed. Context makes a good answer possible. Lineage is how you find out whether you got one. The short version of the fixes:

• When the firm uses AI on its own work, build a shared skill library so the firm learns once instead of every person learning alone, and check the lineage of every answer.
• When clients ask AI directly, test it hard before launch, then give it clear rules for handing a question to a person instead of answering.

The two places AI shows up

One: the firm uses AI on its own client work. A CPA or fractional CFO pastes a client’s numbers into a general tool like Claude or ChatGPT and asks it to build a forecast, explain a variance, or draft the monthly commentary. In our conversations, this is most firms. The large majority use general-purpose tools directly; a smaller group has built a custom assistant or agent. Either way, the person reading the output is the professional.

Two: the client asks AI directly, through the firm. The firm opens a chat channel under its own brand, where a client can ask questions between check-ins instead of waiting for the next call. Here the person reading the output is the client, and the firm’s name is on the answer.

The dividing line is one question: who reads the answer, the professional or the client? That decides what a mistake costs.

Context is essential, and the answer still has to be checked

General AI is good at arithmetic and fluent at writing. What it lacks is context about this particular client. It does not know that two labels mean the same account, which customer has started paying late, or that a large invoice is due next week. It produces a confident, well-formed answer anyway.

Jack Holloway, commercial director at the audit software company Caseware, described the result in a sponsored AccountingWEB article: when accountants paste accounts into a general model, they get noise, differences flagged that are not real issues, and omissions, real issues missed. His diagnosis was missing context. He also made the point that reviewers need to be able to show what was checked and why it was cleared, which a one-prompt-at-a-time chatbot does not give you.

Context is not the whole story, though. A tool can have every relevant number and still draw a conclusion the numbers do not support. So the second requirement is lineage. For any answer you rely on, you should be able to see which data it used, which lines produced each figure, and what it assumed. An answer that cannot show how it got there has not earned your trust, however well it reads.

Place one: the firm using AI on its own work

Paste three statements into a general tool and ask for analysis, and you get something that reads well. The totals look plausible. The commentary sounds sure. Some of it is wrong in ways that are hard to spot.

Three failures come up again and again:

• Miscategorization hidden by a reasonable total. The tool sorts inflows and outflows and gets most of it right. A few items land in the wrong bucket, and because the total still looks sensible, nothing jumps out. This is the dangerous one: the number at the top feels right while five lines underneath are wrong.
• An invented cause. Asked to write commentary, the tool explains why the numbers moved: “margins fell on rising costs.” The fall in margin is in the data. The reason is not. The tool had the numbers and still wrote an explanation nothing supports.
• Noise and omission in variance. It flags differences that do not matter and stays quiet on the one that does, because knowing which variance is material is judgement about the business, not a calculation.

Forecasting is the hardest case. A variance or a categorization can be checked against something that already exists. A forecast cannot, because the outcome has not happened yet. That is why a clean-looking forecast is the easiest one to trust by mistake. We cover this in the forecast blind spot.

Tweaking is not checking

A fractional CFO we work with described her experience honestly: she had to keep tweaking the tool until it gave her the right answer. Iterating is fine. That is how these tools work.

The trap is that tweaking until an answer looks right is not the same as confirming it is right. And a one-off conversation with a general tool has no memory. You fix the same mistakes this month, and next month you start from zero and fix them again. Multiply that by ten people in a firm, each learning the same lessons separately, and the checking never gets cheaper.

The fix: a shared skill library, and lineage on every answer

Instead of ten people each working it out alone, write down what the firm learns once and reuse it. Most general tools now let you save reusable instructions (Claude calls them skills; ChatGPT has custom GPTs and projects), and a shared document works too. This is not a technical project. It is the firm’s know-how, written where the tool can use it.

Keep two kinds of knowledge apart. The method is shared across the firm. Each client’s facts stay in that client’s own restricted workspace. The method is how your firm does the work. The facts are how one client’s accounts map, what counts as their revenue versus an owner’s draw, which of their customers pays late. Those belong with that client, visible only to the people who serve them.

A useful library entry for a piece of client work holds three things:

1. What good and bad look like. One example of commentary you would send and one you would not. These tools follow examples far better than descriptions.
2. The lineage you require. Every figure has to trace to the lines that produced it, so you spot-check the lines behind the totals rather than trusting the total. Every explanation needs evidence for the why, not just the number that changed. If margins fell on rising costs, show which cost lines rose, the invoices or price changes behind them. A number proves that something changed; it does not prove why. If the evidence is not there, the explanation comes out.
3. The corrections that keep coming back. When someone fixes the same mistake twice, it goes into the library.

Then track those recurring corrections. They should decline over time. Saving an instruction does not guarantee the tool follows it, which is why the lineage checks stay in place even after the library grows. If the corrections are not falling, the library is not working, and you know it from evidence rather than a feeling. That falling cost of checking is the measure we argue for in what AI can actually do in client work.

Place two: the client asking AI directly

Now move the same tool to the other side of the relationship. The firm opens a chat channel under its brand so clients can ask questions between calls. It is a good idea, and few firms have done it, because the fear is obvious: the AI gives a wrong answer, under the firm’s name, and the client acts on it.

The failures are the same ones, with nobody expert reading first.

Take a question every founder asks: “Can I afford to hire two contractors?”

• Thin context. The agent sees only today’s bank balance. It says no. The owner passes on the hire and loses the project. It should never have answered: it could not see receivables, which is exactly the information the question turns on.
• Rich context. The agent can see the forecast, the client’s payment patterns, the seasonality. A good answer is conditional and shows its lineage: “Yes, if the two invoices due next week arrive as expected. They come from customers who have paid on time for the past year. If either slips past the end of the month, the answer changes. I’ve flagged this for your CPA to confirm.”

Same question, very different answers. The rich-context one is not better because it says yes. It is better because it says what the yes rests on and what would change it.

The fix: test first, then clear rules for handing off

Two steps, in this order.

Before launch, test it against the questions clients actually ask. Write down every question you can think of, with the correct answer for a sample client, and run the agent against all of them. Include the hard ones: the hire question, “should I take this loan”, questions the data cannot settle. Check that it gives the analysis where the data supports it, states its conditions, and hands off the decision. Run the same set again after every change. People who build these systems call this an eval. For a firm, it is a test you insist on before anything reaches a client.

After launch, give it an escape hatch with clear rules. AI models are built to give an answer, and they will answer even when they should not. Telling the agent to hand off “when unsure” leaves the safeguard in the hands of the same tool that is confidently wrong. Give it concrete reasons instead. The agent routes a question to a person at the firm, with the client’s context attached, whenever:

1. Information is missing. The answer depends on data it cannot see, like receivables it has no access to.
2. The evidence conflicts. The books say one thing and the bank says another, or two sources disagree.
3. The question is outside the channel’s agreed scope. The firm decides in advance what the channel answers. The line is between supporting a decision and making it. Within scope, the channel can explain whether a hire is affordable, model what a loan would do to cash, and state the conditions an answer depends on. Anything that needs the firm’s approval or judgement, like whether to go ahead with the hire, take the loan or make a tax election, goes to a person.

Those two examples — the hire and the loan — are worked through from the owner’s side in what an AI CFO can’t sign off on.

The rule holds even when the agent could produce a confident answer. “Which of my invoices are still unpaid” gets answered from the data, with the invoices listed. “What would this loan do to my cash over the next three months” gets a scenario with its assumptions shown. “Should I take this loan” goes to a person, every time.

Testing catches the problems you can foresee. The handoff rules cover the ones you cannot. Every question that gets handed off becomes a new test, so the channel gets more reliable the longer it runs. As a practising CPA put it to us, it is better to fail than to pretend to work.

The questions to ask in both places

Whether the AI is on your desk or in front of your client, the test is the same. Not “does this answer sound right,” because it always will. Instead: what did it know about this client, can it show how it got the answer, and what does it do when it shouldn’t answer?

On your own work, the shared library carries the method, each client’s workspace carries the facts, and lineage lets you check the rest. In front of a client, nobody is there to catch a mistake, so testing and clear handoff rules have to do that job.

We built AgentLink to do this automatically: it keeps each client’s context, shows where every answer came from, and sends what the data cannot settle to the firm. Everything above works without it.

Frequently asked questions

Where is AI actually used in client-facing accounting work?

In two places. First, the firm using AI on its own work: forecasting, variance analysis and monthly commentary, usually with a general tool like Claude or ChatGPT. Second, clients asking AI directly through a chat channel the firm provides. The first is common today; the second is something most firms want but are wary of building.

Why does general-purpose AI give wrong financial answers that look right?

Mostly because it lacks context about the specific client: which customer pays late, that two account labels mean the same thing, that a large invoice is due next week. But even with the right numbers it can draw conclusions the data does not support, such as explaining why margins fell without any evidence for the cause. That is why every answer needs to show its lineage.

What does it mean for an AI answer to have lineage?

It means you can see how the answer was reached: which data it used, which lines produced each figure, what it assumed, and what evidence supports any explanation. A number shows that something changed; lineage shows whether the stated reason holds up. An answer without lineage has not earned your trust.

How can a firm make AI more reliable on its own client work?

Build a shared skill library holding the firm’s method: examples of good and bad output, the lineage every answer must show, and the corrections that keep recurring. Keep each client’s specific facts in that client’s restricted workspace. Then track recurring corrections; they should decline over time.

Is it safe to give clients a chat channel that answers with AI?

Only with two safeguards. Test it before launch against the questions clients really ask, including ones it should hand off. Then give it concrete handoff rules: route the question to a person when information is missing, when evidence conflicts, or when it needs the firm’s approval or judgement, even if it could produce a confident answer. The channel can still support a decision with analysis and scenarios; the firm makes the call.

How is this different from AI for bookkeeping automation?

Bookkeeping automation handles back-office tasks like data entry and bank-feed categorization. This guide is about AI doing the client work itself: analysis, forecasting, commentary and answering client questions. The stakes are higher because the output is something a firm or client will act on.

About the author

Nilanjan Raychaudhuri is the founder of AgentLink (Tublian LLC, Columbus, Ohio), which builds controller-layer software sold through CPA and fractional CFO firms. He has spent the past year interviewing practitioners at fractional CPA and CFO practices about how advisory work is actually delivered. Team page

The practitioner observations in this guide come from design-partner conversations. Corrections to nilanjan@agentlink.finance.