AI agents tend to show up in the numbers before they show up on the organisation chart. Someone tries a tool on a reconciliation; it saves an afternoon and within a quarter it drafts half the variance commentary. Nobody has decided who owns it, where its work is recorded or what it costs to check. The hours have moved but the management information has not.
In my software business we now use AI agents for nearly everything in our day-to-day delivery work. We got there gradually: code assistance first, then our own internal work, then smaller client tasks. We still keep agents away from security-heavy client work.
When developers check work they did not write, in a style they do not recognise, review time goes up
Software is not accounting, but the pattern maps closely on to a finance team, and the lessons are about record keeping more than technology.
Here are four that I think apply to anyone who signs off numbers.
1. Review time belongs in management information
Our surprise was not what the agents could produce; it was how long review took. When developers check work they did not write, in a style they do not recognise, review time goes up a lot. Our plans had not allowed for that. The drafting got faster, the checking quietly got slower and for a while only the first half showed up anywhere.
Finance teams will meet the same thing. Picture a month-end close where an agent drafts the bank and intercompany reconciliations, and a first cut of the board pack commentary. The draft arrives in minutes. Then a qualified accountant spends two hours working out why one plausible explanation of a margin movement cites the wrong cost line.
An agent with no owner is a process with no owner, and auditors tend to have views about those
If the management information shows only that the close finished a day earlier, it is telling half the story. Review hours are a real cost; they sit with some of your most expensive people and they are the control that makes the output usable. Measure them like any other cost.
2. Record agent work and human review separately
The next step is to record review time where it can be seen. We track the share of our code that is AI-assisted as a percentage and discuss it in our lead developer meetings. The figure itself matters less than the conversation it forces: where is it high, where is review heavy and where are we not getting the value we expected?
In a finance team, the equivalent is simple coding discipline. Use the cost centre or client matter codes you already have, and split each one into two lines: agent-run and human review. A practice preparing year-end accounts might record the agent’s drafting run and then three hours of manager review against the same client code. A shared service centre might do the same by cost centre for purchase invoice matching.
After a few months, you can see which processes genuinely benefit and which simply move work from preparers to reviewers. That helps pricing, resourcing and, frankly, honesty with clients about what they pay for.
3. Every agent needs a named owner and a clear scope
An agent with no owner is a process with no owner, and auditors tend to have views about those. At WeCode, agents are used on both our side and the client’s side, for mock-ups and requirements, so it matters a great deal who is accountable for what an agent produced and who agreed to rely on it.
In finance, I would write it down plainly. For example: the financial controller owns the reconciliation agent. Its scope is to draft reconciliations and flag unmatched items. It never posts journals, never approves payments and never changes master data. A named person reviews its output before anything reaches the ledger. The FP&A lead might own a separate agent that drafts variance commentary, with its scope ending at ‘draft for review’ and sign-off staying with a human.
The point of separating agent hours from review hours is to see cost and control, not to rank
If you cannot name the owner and describe the scope in two sentences, the agent probably should not be running yet.
4. The record is about work, not watching staff
This is the part that worries people, rightly. Recording agent and review time can sound like the first step towards monitoring every click. It does not need to be. At WeCode, people record their own time. We do not take screenshots, log keystrokes or record screens. The record exists to understand the work: what was done, for whom, by whom or by what, and what it took to check.
The same principle should hold in finance. The point of separating agent hours from review hours is to see cost and control, not to rank accountants by speed. If a team feels watched, people stop recording honestly the moment review overruns, which is precisely when you most need to know. Trust is a data quality control, too.
Questions for boards
Boards and audit committees do not need a technical briefing on AI. They need one answer: for each agent in use, who owns it, what is it allowed to do, and how much qualified review time does its output consume? If finance can answer that from its own records, the agents are part of the team. If it cannot, they are working off the books, which is not where finance usually likes things to be.
These are practical observations from running a software agency, not accounting advice or ACCA guidance. But the habit of recording who did the work, and who checked it, is one finance professionals already understand better than most.