What AI can actually do in client work

AI can already do a lot of real client work. It can categorize transactions, support reconciliation by proposing matches between payments and invoices, explain why this month’s numbers moved against last month’s, forecast a client’s cash over the next thirteen weeks, keep track of what was agreed on client calls, and draft plain-English answers to the questions clients ask. None of that is speculative. Firms are using all of it today, and the tools keep getting better. For a broader view of the products themselves, see our map of AI tools for CPA firms.

So “what can AI do?” is mostly settled. It can do almost anything you would put in front of it. That is also why the question has stopped being useful. Every vendor is selling the same list.

The question that decides whether AI earns its place in your firm is a different one: for each thing it does, can you check the work, and when? Your name goes on what reaches the client. If you cannot check the output, it does not matter how capable the system is.

This guide sorts AI client work into three groups by that question. By the end you will have a way to look at any AI task, yours or a vendor’s, and know how much to hand over.

“Verify” is hiding three different situations

Everyone says “trust but verify.” It usually arrives as a warning at the end of an article: the AI can make mistakes, so check its work.

That advice pictures one action. The AI gives you an answer, you look it over, you catch the mistake. It works when the right answer already exists somewhere for you to compare against. A lot of client work is not like that.

There are really three situations hiding inside the word “verify.”

• Checkable now. The answer is already in the ledger or the bank. You can compare the AI’s output against it today.
• Checkable later. The answer does not exist yet. A forecast is the obvious case. You find out whether it was right when the future arrives, and only if you wrote down what you expected beforehand.
• Needs judgement. The numbers can be right and the call can still belong to a person, because it rests on things about this client that the data cannot see.

One task can touch all three. A single recommendation might rest on inputs you can check today, a prediction you can check next month, and a final call that still needs the firm. The groups sort the parts of the work, not whole tools.

Which group a piece of work falls into tells you how much rope to give the AI.

Group one: checkable now, and the cost nobody counts

This is the group everyone means when they talk about AI in accounting. Categorizing transactions. Matching payments to invoices. Answering “what did we spend on software last quarter?” The right answer exists, so you can check it.

That sounds like the easy case. It hides the cost most AI adoption never counts.

A bookkeeper we know uses the automatic categorization built into a popular accounting platform. His experience is that it is wrong often enough that he checks every line. He did not get his time back. He now does his old job plus supervising the AI doing it.

Checking is work. Most firms measure what the AI produces: how many transactions it categorized, how many matches it proposed. Almost none measure the time spent checking what it produced. When checking every output takes about as long as doing the work, the AI has moved the job, not removed it.

You also cannot rely on how it feels. In a 2025 study by the research group METR, experienced software developers expected AI tools to speed them up by 24%. After working with them, they believed the tools had sped them up by 20%. Measured on the clock, they had taken 19% longer. METR has since said those results are out of date, and its follow-up suggests newer tools may be faster, though it calls its own new estimate unreliable. The study was about developers, not accountants. The lasting lesson is the gap between what people felt and what was measured: the original study found people got even the direction wrong.

The property that separates good systems from bad ones

The useful question is not whether an AI system is accurate on day one. It is whether the amount of checking it costs you falls over time.

In a bad system, you check every output forever. The same kind of mistake comes back every month, and you correct it every month. In a good system, a correction made once is remembered, so the same check is not needed twice. Your checking load shrinks.

Measure it

Most firms running AI can tell you whether it is switched on and how much it produced. Very few can tell you whether it is getting better. You can, with one number: how often the firm corrects the same kind of thing, tracked month over month. If that count is falling, the system is learning. If it is flat, you are paying the checking cost indefinitely, however impressive the output looks.

Why failures have to be visible

One of our design partners, a practising CPA, has a rule we built our system around: better to fail than to pretend to work.

It matters here because the checking load can only fall if failures are visible. A system that is unsure but buries its guess inside a confident answer cannot be corrected, because nobody can see what to correct. So it never learns, and your checking never shrinks. A system that says “I could not tell where this belongs” gives you something to fix once.

Hidden failures cannot be measured. What cannot be measured cannot improve. We cover how to judge this inside a system in how to tell whether an AI system can be trusted with client numbers.

Group two: checkable later, and the pattern that changes under you

The second group is anything where the answer does not exist yet when the AI produces it. Cash forecasts. When a client’s customers will pay. Whether a commitment made on a call, like holding off on a hire until collections recover, actually held.

You cannot check any of these today, however carefully you read them. You can only check them when the future arrives.

A real example

One client’s largest customer is invoiced on 30-day terms, and for years paid around day 30. When the client connected, its payment history showed that clearly, so the forecast started in the right place.

Then the customer started to slip. Payments landed around day 40, then around day 50. A forecast still leaning on the old pattern expected cash that was not there yet, and missed a dip it should have shown.

Each miss was recorded against what had been expected. After a few of them, the system stopped trusting the old pattern for that customer and forecast the new one.

Learning a pattern from history is easy. What’s harder is noticing the pattern has changed and learning from the misses instead of holding on to the past.

The same invoice, two ways of checking

You might ask whether the aging report already tells you this. It tells you part of it.

The aging report shows that an invoice is overdue. That is a fact about the past. You can check it against the ledger today, so it belongs in group one.

What the aging report does not tell you is when the money will arrive. The forecast does. It commits to a date, and it gets scored when that date comes. Same invoice, two different ways of checking: one against what has happened, one against what has not happened yet.

What your best bookkeeper already does

Good bookkeepers do this in their heads. When a reliable customer starts paying late, nobody tells them. They notice it month after month, and they quietly adjust what they expect.

That knowledge is valuable and fragile. It lives in one person’s head. It is not written down. It leaves when they leave, and it cannot be applied to forty clients at once. The aging report never captures it, because the aging report resets every month and learns nothing.

A forecasting system that learns from its misses does what that bookkeeper does, by watching outcomes, except it writes the pattern down. Then it does not forget, it does not leave, and it applies to every client.

This is also why a smarter AI model does not solve forecasting on its own. “This customer has started paying three weeks late” is not in any model’s training data. It is specific to this one business. A better model helps with what generalizes across businesses. Only checking forecasts against what actually happened, over and over, learns what is particular to this one.

Where this has limits

A late payment is not automatically a new pattern. The system has to distinguish a one-off delay from a changing payment pattern, the same as a person would, before it changes the forecast.

Where history is thin, such as a new customer or a short trading record, forecasts should carry less confidence and say so.

Some expected money has no invoice behind it at all: “I’m expecting a payment from this customer.” When money arrives, the bank shows it. What can be missing is enough evidence to tie that receipt to the expectation. Where there is not, the honest move is to ask on the next call rather than guess.

For more on this group, see the forecast blind spot, learning from forecasts and tracking client commitments.

Group three: right numbers, still a person’s call

The third group is where AI can do the analysis well and still should not make the decision. That split is the whole subject of what an AI CFO can’t sign off on, if you are weighing one up.

Here is a case a fractional CFO we work with raised while we were designing recommendations for working capital. A client’s days sales outstanding is climbing, so money from customers is arriving more slowly. Its days payable outstanding is short and stable, so it pays its own vendors quickly. The textbook recommendation writes itself: stretch vendor payments from 8 days to 30 or 60, and hold onto cash longer.

The arithmetic is right. An AI system can find that recommendation, size it and show what it does to cash.

And the client may say no: “That vendor is critical to us. I won’t risk the relationship.” Nothing in the ledger knows that. The numbers said stretch, the relationship said don’t, and the person who knows the client is right.

Outcomes can inform that kind of judgement later. They will not conclusively prove whether risking the relationship would have been right. If the client had stretched payments and the relationship soured, was it the advice, or something else? That is why the call belongs with the firm, even when the analysis is sound. It does not mean advice cannot be evaluated. We cover how to do that in measuring advisory impact.

The rejection is the lesson

What the system does with that “no” matters as much as the recommendation.

The rejection should be recorded, and the hard question is its scope. Some rejections are permanent: “This vendor is critical. Never suggest delaying them.” That belongs to the vendor for good. Others are temporary: “We held payments to them during the cash crunch.” That was only true while cash was tight, and should stop applying when it isn’t.

Get the scope wrong one way and the system keeps suggesting something the client already refused, which is exactly how AI loses a firm’s trust. Get it wrong the other way and a one-off situation quietly suppresses a good recommendation forever.

So the system should propose the scope, and the firm should confirm it in one step. In this group, the AI’s job is to lay out the facts, make the recommendation, and carry the firm’s correction forward at the right scope. It should never pretend it could have made the call itself.

Sort before you buy

When you look at an AI tool for client work, the list of what it can do will be long and mostly true. Use the three groups instead. Four questions do most of the work.

1. Which of these tasks can I check today, which only later, and which still require professional judgement? A vendor who has thought about this will answer quickly. One who hasn’t will tell you it is all accurate.
2. For the work I can only check later, what does it write down now so I can score it then? If nothing is recorded before the outcome, nobody can say afterwards whether it was right.
3. Is the amount of checking falling month over month, and can you show me? This is the difference between a system that learns and one that taxes you forever.
4. When it does not know, does it say so? A system that always produces a confident answer is guessing some of the time, and you will not know which time.

The examples in this guide come from building AgentLink, an AI controller layer that forecasts and tracks client commitments, checks them against what actually happens, and brings the firm only what the data cannot settle. It is sold through CPA and fractional CFO firms. The firm keeps the client, the pricing and the judgement.

Frequently asked questions

What can AI actually do for an accounting firm?

AI can categorize transactions, propose matches for reconciliation, explain variances between periods, forecast a client’s cash, track commitments made on client calls and draft answers to client questions. The more useful question is which of those outputs the firm can check, and when: some against the ledger today, some only when the future arrives, and some that still need professional judgement.

Does AI save accountants time, or create more checking?

It can do either. If the firm has to check every output, and the same mistakes come back every month, AI moves the work rather than removing it. A good system remembers corrections, so the amount of checking falls over time. The way to know is to track how often the firm corrects the same kind of thing, month over month.

How do you check an AI cash forecast?

Not on the day it is made, because the outcome has not happened yet. The forecast has to record what it expected, such as when a customer would pay, so that it can be compared with what actually happened. A forecasting system should learn from those misses, especially when a customer’s payment behavior changes.

Can AI make recommendations to clients?

AI can find and size recommendations, such as extending vendor payment terms to hold cash longer, and the arithmetic can be right. Whether the client should act on them often depends on things the data cannot see, such as how important a vendor relationship is. That call belongs with the firm, and the system should remember the firm’s decision at the right scope.

How do I know if our AI system is getting better?

Measure how often the firm corrects the same kind of thing, month over month. If that number is falling, the system is learning. It can only fall if the system shows when it is unsure, because hidden mistakes cannot be corrected.

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 and are used with permission where attributed. Corrections to nilanjan@agentlink.finance.