Most conversations about artificial intelligence in accounting firms start with the tools. Which product reconciles the bank feed, which one codes the transactions, which one drafts the month-end commentary. It is an understandable place to start, because that is what the demos show. In my experience it is the wrong place to look for risk.
A recent piece on CPA Practice Advisor, “The 2026 AI Finance Stack,” made the point that the lasting value sits in the workflow layer between systems: the part that moves data reliably from one place to another and keeps a full record of what happened along the way. I agree, and I would go a step further. The workflow layer is also where AI quietly changes who is responsible for what. If a firm does not redesign its review points and its audit trail around that change, it ends up with faster closes that nobody can fully explain.
Recommended Articles
Accounting October 6, 2026
The 2026 AI Finance Stack
I have spent more than a decade building custom AI systems, with a particular focus on the finance function. What follows is the practical view I share with firm leaders who want the speed without losing the ability to sign their name to the result.
The close was always a chain of handoffs
Before any AI arrived, a month-end close was a series of handoffs. A bookkeeper codes transactions. A senior reviews the reconciliations. A manager looks at the variances. A partner signs off the pack. Each handoff is also a checkpoint, and each checkpoint has a person attached to it who can answer the question “why does this number look like that?”
AI does not remove the chain. It moves the checkpoints. When a tool auto-matches bank lines, suggests accruals, or reclassifies a batch of expenses, it is doing work that used to pass through a human pair of eyes on the way to the next stage. The output still lands in the next person’s queue, but the reasoning that produced it is no longer sitting in a colleague’s head. Unless someone deliberately decides where the new checkpoints are, they simply disappear.
Three places the review point quietly moves
- Matching and reconciliation: Automated matching is one of the most useful things AI does in the close, and one of the easiest to over-trust. A tool that matches the vast majority of lines leaves a small set of exceptions, and the team naturally focuses on those. The risk is in the lines it matched with confidence and got wrong: a payment applied to the wrong invoice because the amounts agreed, or two similar supplier names merged into one. The review point needs to move from “clear the exceptions” to “sample the confident matches as well”.
- Coding and classification: When a model learns a client’s coding from history, it learns the history’s mistakes too. If a previous bookkeeper consistently put software subscriptions in the wrong expense line, the AI will do it faster and more consistently. The review point here is not each transaction. It is the rules the system has inferred, checked periodically by someone who knows how the client’s chart of accounts is meant to work.
- Commentary and explanations: Language models can now write variance commentary that reads very well. That is exactly why it needs a reviewer who checks the claim, not the prose. A fluent sentence explaining that costs rose because of a supplier price increase is only useful if the supplier price actually increased. The review point is the link between each explanation and the evidence behind it.
What a usable audit trail looks like
Firms already understand audit trails for human work. The question is what the equivalent looks like when part of the work is done by a model. In practice, I suggest firms make sure they can answer four questions for any AI-assisted step in the close.
First, what did the system see? The data it used, from which source, as at which date. If the bank feed was refreshed after the model ran, the record should show it.
Second, what did it do? Not a vague note that “AI reconciled the account,” but the actual matches, codings, or adjustments it proposed.
Third, what did a person change? If a reviewer overrode a match or rewrote a line of commentary, keep both the original and the change, with a short reason. This is the cheapest control a firm can put in place, and over a few months it becomes the best evidence available of where the tool is reliable and where it is not.
Fourth, who accepted it? Every AI-assisted output that ends up in a client deliverable should have a named person who accepted it. Not because anyone wants a culture of blame, but because a client, a regulator, or a future partner will one day ask, and “the system did it” is not an answer a firm can give.
Design the workflow before choosing the tool
The firms that get this right tend to draw the workflow first and pick tools second. They map the close as it really runs today, mark each handoff, and decide in advance which checkpoints stay human, which can be sampled and which can be automated outright. Only then do they ask which tools fit that design.
Firms that work the other way round usually discover the gaps at the worst moment: during a busy season, on a client file where something has gone wrong, with no clear record of which step introduced the error. At Critical Future we spend a surprising share of our finance work on exactly this mapping exercise, because the model is rarely the weak point. The handoffs around it usually are.
It also helps to start small. Pick one recurring process for a handful of clients with clean data, run the AI alongside the existing method for a few cycles, and compare. Shadow running costs a little extra time up front and tells you far more than any demo about where the review points need to sit.
People and machines, each doing their part
Back in 2018 I wrote about what I called Collective AI: the idea that the best results come from people and machines working together, each doing what they do best, rather than from machines replacing people. The close is a good example. Let the model do the high-volume matching, coding, and first drafts. Keep people responsible for judgment, for the client relationship, and for the moment someone signs.
The firms that will benefit most from AI are not necessarily the ones with the most tools. They are the ones that can still explain, line by line, how a number got into the pack, even when a machine did most of the work to get it there.
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: vectorjuice/Freepik
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In
