
What Changes in Your Controls When AI Enters a Process
Who reviews the output, what evidence exists, and what the control actually relies on now.
A finance team starts using an AI tool to match invoices to purchase orders. Accuracy goes up, the backlog disappears, and everyone is pleased. Then someone asks who is reviewing the matches now.
That question tends to come up late, often during an audit, and often after the tool has been running for months. The process changed. Nobody went back to check whether the controls around it changed with it.
- When AI takes over a step, the control that sat on that step does not automatically move with it.
- Review controls change most. Checking AI output is a different job from checking a colleague's work.
- Policies are common. Monitoring and evidence are where most organizations are still behind.
The Step Changed. Did the Control?
Most controls were designed around a person doing a task. A clerk matched the invoice, a supervisor reviewed a sample, and the risks were familiar ones: fatigue, shortcuts, the odd keying error.
When AI takes over that step, the risk changes shape. Errors tend to become less frequent and more consistent. A person makes a different mistake on a Tuesday than on a Friday. A model that has learned the wrong rule makes the same mistake every time, at volume, and quietly.
That changes what a good control needs to look for. The old review was built to catch random slips. The new one has to catch patterns.

Review Controls Need the Most Rethinking
A reviewer checking a colleague's work has a feel for what a mistake looks like. AI output rarely gives that signal. It is formatted cleanly, it is complete, and it looks equally confident whether it is right or wrong.
So "reviewed and approved" can mean very little. If the reviewer is scrolling through four thousand matches a day, the review is a signature, not a control. The useful questions are more specific. What gets a human look and what does not? What threshold sends an item for review? What does the reviewer compare the output against?
A shared services team used an AI model to code supplier invoices to general ledger accounts. Spot checks looked clean for months. Then a quarter-end analytic showed one supplier's invoices had been coded to the wrong expense account ever since the supplier changed its trading name. Every invoice, consistently. The reviewer had been checking that the codes looked reasonable, not that they matched the contract.
Evidence and Change Management
Two practical questions come up quickly. The first is evidence. When a control relies on AI, what shows that it operated? The output alone usually does not. Logs of what was processed, which items were routed to people, which version of the model or rules was running, and what the reviewer actually did are what an auditor will ask for.
The second is change. When a vendor updates its model or a team adjusts a prompt or threshold, that is a change to how the control works. It may never go near the IT change management process, because nobody thinks of it as a system change.
Proving the control works is not.Monitoring trails policy by 27 points.
Policy Is the Easy Part
Recent survey data shows the gap clearly. In a 2025 Gradient Flow survey of organizations working with generative AI, three in four had an AI usage policy. Fewer than half monitored their AI systems in production for accuracy, drift or misuse. The policy exists. The control often does not.
Internal audit is in the same position. The IIA's 2025 Pulse survey found that 41% of internal audit teams use generative AI, mostly occasionally. The function being asked to assess AI controls is still learning to use the tools itself, which is all the more reason to build the right questions into the audit plan now.
The Short Version
AI changes the step. The control around it has to be looked at again on purpose, because it will not update itself. The organizations that get this right are not the ones with the longest AI policy. They are the ones that can show, for a specific process, who checked the output and what they found.
For each AI step, ask
A short inventory goes a long way. For each process that now includes an AI step, including ones a vendor added to software you already use, ask:
- What decision does the AI make, or influence?
- Which control used to cover that step, and does it still apply?
- Who reviews the output, how much of it, and against what?
- What evidence would show the review happened?
- What happens when the model, the prompt or the settings change?
None of these require deep technical knowledge. They are the same questions a good auditor asks of any process. They just need asking again.
The IIA, 2025 North American Pulse of Internal Audit.
Start a Conversation
Have a Specific Problem to Work Through?
An aging finding, control issue, documentation gap, or workstream short on capacity is enough of a starting point. Tell us what is happening and what you need to be true at the end.
Tell us what is stuck, what changed, or what deadline is driving the need. We can determine whether BPS Advisory Group is the right fit and what a sensible scope would look like.
Or write to info@bpsadvisoryus.com