AI controller for CPA firms: how a small firm runs one across its client book

If you run a small CPA, accounting or fractional CFO firm, you already know the bind. You are still in the client work yourself. You carry fifteen, twenty, maybe thirty clients. Those clients want more from you between calls: a quick answer when cash looks tight, a heads-up before a problem lands, someone who remembers what was agreed last month. And the only way most firms have ever added that kind of attention is to add a person.

An AI controller is a different answer to that bind.

An AI controller does a controller’s daily work across a firm’s client book. It is software that watches each client’s bank and ledger every day, keeps the forecast current, checks the plan against cash, and carries what was decided into the next call. The firm keeps the judgement.

There is already a page on what an AI controller is. This guide is about something more practical: how a small firm actually runs one. Which clients go on, how it connects, what you tell clients, what changes in your week, and what stays with you. It describes AgentLink’s firm-led model, so some specifics below, like read-only connections and the firm’s name on everything clients see, are how AgentLink works rather than properties of every AI controller.

Why would a small accounting firm run an AI controller?

Because attention is the thing you sell, and today attention costs hours.

Think about what it takes to walk into a client call properly prepared. Pull the books. Check the bank. Remember what you agreed last time and whether it happened. Look at whether the forecast still holds. Across most firms we talk to, that preparation runs two hours to half a day per client per cycle. Multiply that by your client list and you can see why the work between calls gets squeezed. It isn’t a lack of care. It’s arithmetic.

The usual fix is to hire a controller. That helps, but the ceiling just moves. A controller is one person with a finite week, and a part-time controller can only carry so many clients before the same squeeze returns. You add cost, and you add a hire to manage, and the gap between calls is still there for whoever doesn’t fit in that person’s week.

An AI controller takes on the daily watching itself. It doesn’t get tired at client twenty-two, and it doesn’t need a week off. That frees your hours, and your staff’s hours, for what clients actually pay you for: judgement, conversations, and advice they act on. For a small firm, that is the difference between offering more and offering more only to a few.

Which clients should go on first?

This is the question every firm asks, and it is the wrong question.

It sounds sensible. Start with the clients who value advisory work, the ones already paying for more, the ones who would notice. Leave the bookkeeping-only clients alone, because they never asked for anything beyond clean books.

But look at why that split exists in the first place. Firms didn’t sort clients into advisory and bookkeeping because some clients deserve attention and others don’t. They sorted them because attention costs a person’s time, and there is only so much of it. The split is rationing. It was never a judgement about who would benefit.

Every business you serve has cash that moves, plans that drift, and decisions that need following through. A client who only buys bookkeeping today isn’t a client who doesn’t need more. More often, it’s a client nobody has offered more to yet. As a seasoned CFO put it to us recently, advisory isn’t optional for any business. Some owners just don’t know they need it until someone shows them what they were missing.

So the answer is: all your clients. Once the daily watching no longer costs a person’s hours, there’s no reason to ration it. Every client gets the same standing attention, and you decide where to spend your conversations based on what the watching turns up, not on who happens to be paying for the top tier.

That doesn’t mean everything happens on day one. In practice, start with the client who asks you the most questions between calls. That’s where the time goes today, so it’s where you’ll feel the relief first. Then bring the rest on in the weeks that follow. The principle is every client. The rollout starts where it hurts.

Who connects it, and how long does it take?

The firm connects it, and adding a client to AgentLink takes under five minutes. Setup follows the way a firm already works, so the client barely has to do anything.

The books and the bank. The firm connects both, using the access it already has to the client’s accounting system and bank. That’s how firms already operate, so there’s no new permission to chase. In AgentLink, both connections are read-only. Nothing changes in the client’s books, and it doesn’t post entries or move money. If the firm wants the client to have their own access later, it can add them. How we handle your clients’ data covers the detail.

What was agreed on each call. The firm uploads its call transcripts on its normal cadence. AgentLink then pulls out the decisions worth tracking: the targets, promises and commitments that can later be checked against the bank and the books. That extraction is AgentLink’s own technology, built to separate the decisions that matter from the rest of the conversation, so the firm never has to hand-build a list. It’s what lets the AI controller check, weeks later, whether what was decided actually happened.

The brief. Once a client is connected, the prepared brief arrives the day before each scheduled call, so the firm walks in with the picture already built.

Questions that need a person. By default, they go to the person at the firm who set up the client, usually the CPA or fractional CFO. The firm can also add the business owner to that client, and they’ll receive the same notifications if the firm gives them access.

One thing does take longer than setup. The AI controller reads each client’s history to learn how that client really runs: which customers pay late, which bills land when, what a normal month looks like. Its forecasts get sharper as it checks its own predictions against what the bank later shows. Connecting a client takes minutes. Getting to know the client takes what it takes a new staff member: time with the numbers.

What does the firm tell clients, and whose name is on it?

Your name is on it. Clients hear from the firm, not from a software vendor.

This matters more for a small firm than a large one. Your client relationships are personal, often built over years, and they are the business. An AI controller that sent clients its own branded messages would put a stranger between you and the people you serve.

So in AgentLink, everything a client sees arrives under the firm’s name. When a client asks a question between calls, like whether they can afford to pay a large bill this week, the answer comes from the firm, answered under the firm’s name and drawn from the client’s actual numbers. Questions that call for judgement go to a person at the firm instead of being answered.

What do you tell clients? Something simple and true: the firm now watches their numbers every day, not just before each call. That’s a service upgrade they can feel, and it’s yours to offer.

What changes in the firm’s week?

The biggest change is where the hours go.

Before each client call, the preparation arrives already assembled: what changed since you last spoke, what was expected and whether it happened, what is coming up, and what is worth raising. You spend the hour on the conversation, not on rebuilding the picture.

Across the whole client book, the firm gets a standing view of which clients need attention this week and why, so problems surface early instead of at the next scheduled call. And what was decided on each call stays connected to what happens afterwards, so the next conversation starts where the last one ended instead of starting over. Our guide on why client calls keep starting over explains why that matters so much.

None of this asks the firm to change how it works. Calls happen on the same cadence. The tools the firm uses for the books stay the same. What changes is how much of each week goes to assembling and how much goes to advising.

What does the firm still own?

Everything that requires judgement.

An AI controller supplies attention: what changed, what was expected, and whether it happened. It can model a hire, show what a loan would do to cash, and lay out the options with its assumptions visible. The firm owns the analysis, the recommendation, and the line for what gets reviewed by a person. The client approves the decision itself. Whether to take the loan, make the hire or delay the purchase is the owner’s call, unless they have explicitly handed that authority to the firm. What the firm puts its name and its accountability to is the advice.

The firm also owns the boundary: which questions the AI controller answers on its own, and which go to a named person before anyone acts. Decisions that are costly or hard to reverse should always go to a person. We cover where that line sits, and why accountability can’t be handed to software, in What an AI CFO can’t sign off on.

And the firm owns the relationship. The AI controller makes the firm more present between calls. It doesn’t stand in for the people the client trusts.

What won’t an AI controller do?

It’s worth being clear about the limits, because a tool that oversells itself is worse than one that doesn’t.

• It won’t keep the books. It works from the records bookkeeping produces. If the books are behind, it says so rather than pretending its answers are current. What a controller does that a bookkeeper does not explains why the two jobs are different.
• It won’t make the decision. It prepares the ground and shows its working. The firm recommends, and the client approves.
• It won’t know what was never shared. If something was agreed in a hallway conversation and never recorded, the AI controller can’t follow it through. It’s only as complete as the firm’s record of what was decided.
• It won’t go around the firm. AgentLink isn’t sold directly to business owners. Clients get it through their CPA or fractional CFO firm, under that firm’s name.

Frequently asked questions

What is an AI controller for a CPA firm?

An AI controller does a controller’s daily work across a firm’s client book. It is software that watches each client’s bank and ledger every day, keeps the forecast current, checks the plan against cash, and carries what was decided into the next call. The firm keeps the judgement and the client relationship.

Which clients should a CPA firm put on an AI controller?

All of them. Firms have traditionally limited advisory attention to a few clients because attention costs staff hours. When the daily watching no longer costs a person’s time, there’s no reason to ration it. In practice, start with the client who asks the most questions between calls, then bring the rest on over the following weeks.

Does AgentLink change the client’s books?

No. AgentLink connects read-only to the firm’s accounting system and the client’s bank feed. It doesn’t post entries or move money. It works from the records the firm already keeps.

How long does it take to add a client to AgentLink?

Under five minutes. The firm connects the client’s accounting system and bank using the access it already has, both read-only. The first prepared brief arrives the day before the client’s next scheduled call.

Do clients see AgentLink’s brand?

No. In AgentLink, everything a client sees arrives under the firm’s name. Questions clients ask between calls are answered under the firm’s name, and questions that need judgement go to a person at the firm.

Do we still need a controller or bookkeeper on staff?

The AI controller doesn’t keep the books, so bookkeeping still has to happen. It takes on the daily watching that otherwise fills a controller’s week, so the firm’s people can spend their time on judgement and client conversations instead of assembling the picture before every call.

How is an AI controller different from hiring a fractional controller?

A fractional controller is a person with a finite number of hours, which caps how many clients they can carry. An AI controller does the daily watching across every client at once. They work well together: the AI controller keeps the picture current, and the person brings the experience and judgement.

About the author

Nilanjan Raychaudhuri is the founder of AgentLink (Tublian LLC, Columbus, Ohio), which builds an AI controller sold through CPA and fractional CFO firms. He has spent the past year working with practitioners at small CPA and fractional CFO practices on how client work is actually delivered between calls. Team page

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