Take an interim audit request list that arrives in early September with one new line. Please identify any processes relevant to financial reporting in which AI tools are used, and describe the related controls. The controller can name the bank reconciliation tool that suggests matches. She is less sure whether the FP&A team's commentary drafts count, or what changed when the AP coding assistant went live in May.
That line is where AI internal controls start for most finance teams. The underlying principles have not changed: a control has to be designed to catch a material error, operate consistently, and leave evidence. What changes is that part of the work inside the control is now done by software whose behavior can shift without anyone touching the ERP.
This guide is practical rather than legal. Requirements differ by jurisdiction, by listing status and by audit firm, so treat it as a list of topics to discuss with your auditors, not as a statement of what they will require.
Where AI touches financial reporting
Start by sorting each AI use into one of three roles. The control implications differ.
| AI's role | Example | What the control must cover |
|---|---|---|
| Performs a step | Auto-matching bank lines to ledger entries above a confidence threshold | The rules and thresholds, testing of the matches, and review of what was matched automatically |
| Assists a person who performs the control | Suggesting invoice coding or drafting flux commentary for a reviewer to accept or change | What the reviewer checks and against what source, so the review is real and not a click-through |
| Produces information a control relies on | An AI-generated exceptions list that a reviewer works through | Completeness and accuracy of the list, because items it misses are never reviewed |
The third role is the one most often missed. Auditors already test whether the reports used in a control are complete and accurate, and an AI-produced list is a report like any other. If the tool quietly drops a category of exception, the review that follows will look clean and miss it.
Build a simple register: process, step, tool, role, control owner, related control. For most mid-sized companies it fits on one page, and it answers the auditor's request directly.
Designing AI internal controls for assisted steps
Decide what the control actually is. In most AI-assisted steps the control is still the human review. The AI output is an input to that review. Write the control description that way, so it is clear that the reviewer is accountable and the tool is not.
Make the review precise. "Review AI-suggested journal entries" will not satisfy anyone. Say what the reviewer checks: supporting documents, account, amount, period, and whether the entry is unusual for that account. Set thresholds, such as every entry above a set amount checked to source, with a sample below it.
Test what the tool leaves out. For any AI-produced list or match, check periodically for completeness. Take a sample of items the tool did not flag and confirm it was right not to. This is the step reviewers skip most often, because nothing on the screen prompts it.
Keep the reviewer able to disagree. If the AI suggestion is pre-filled and accepting it takes one click, acceptance rates will drift towards 100%. Track how often reviewers change or reject suggestions. A rate near zero may mean the tool is excellent, or that nobody is looking. Find out which.
Do not stack AI on AI without a person. One tool checking another's output can catch some consistency errors. It does not replace a reviewer with the domain knowledge to see what both tools missed, such as a contract term that never reached the ledger.
Evidence an auditor can re-perform
Evidence for an AI-assisted step has to let someone else see what the tool was given, what it produced and what the person did with it.
| Evidence item | Why it is needed |
|---|---|
| Input, or a reference to the data used | Shows what the tool was working from, and allows the output to be re-created or checked |
| Tool, prompt or configuration version, and model version where known | Shows which version of the logic produced the output |
| Output as produced | Separates what the tool suggested from what the reviewer changed |
| Reviewer, changes made and decision | Shows the review happened and what judgment was applied |
| Approval timestamp and approver | Ties the step to the control and the period |
| Exceptions and overrides | Shows how cases outside the rules were handled |
Keep it for as long as the financial records it supports are kept. Where a vendor runs the tool, confirm that you can get this evidence out of it, and in what form, before you rely on it at year-end.
Change management for prompts and models
A traditional system changes when someone changes it. An AI-assisted step can change in four ways, only one of which touches your own configuration.
The prompt or instructions change. Someone improves the wording to fix a problem. The output shifts for every case, not only the one that prompted the change.
The source library changes. A new policy is added or an old one removed. Answers that relied on it change.
The model changes. The vendor updates the underlying model. Your configuration is untouched and the behavior still moves.
Access or data changes. The tool gains or loses access to a data source, and its outputs change accordingly.
Treat prompts, instructions, source libraries and access settings as configuration under change control: requested, tested, approved by someone other than the person who made the change, and recorded. For the model itself, ask the vendor how updates are announced, whether you can pin a version, and whether their SOC 1 report, if they provide one, covers the AI service.
The practical tool is a regression test set: a fixed set of cases with known correct outcomes, run after any change of any kind. Keep the results with the change record. If the outputs move, someone decides whether that is acceptable before the tool is used in the next close.
Segregation of duties when software does the work
Give each AI agent its own identity. Automated tools that act on financial data should run under their own accounts, not a shared one and not a person's. That way their rights can be reviewed and their actions traced.
Put agents in the SoD matrix. An agent that can both create and approve a transaction is a segregation of duties conflict, exactly as it would be for a person. So is an agent that can change vendor bank details and propose payments to that vendor.
Separate the builder from the reviewer. The person who maintains a tool's prompts and configuration should not be the only person reviewing its output in the same process. Otherwise one person can change the logic and approve its results.
Include AI access in periodic reviews. Access reviews should list every AI tool and agent account with access to financial systems, what each can read, and what each can write. Read-only by default makes this review short.
What auditors are likely to ask
There is no single rulebook yet. The PCAOB's staff spotlight of 22 July 2024, based on outreach to audit firms and preparers, observed that GenAI use in audits "appears to be focused primarily on administrative and research activities". Among preparers, GenAI in accounting and financial reporting "appears to be secondary to integrating GenAI in operational and customer-facing areas". That was a snapshot from two years ago, and use inside finance has grown since.
KPMG's Global AI in Finance 2026 report, based on responses from 1,013 senior finance leaders across 20 countries and published on 11 May 2026, found that "Active use of AI in finance has more than doubled since 2024, rising from 30 percent to 75 percent". It put the share of organizations "fully assurance-ready for AI-enabled finance processes" at 42 percent, fewer than half. Use has moved faster than the controls around it.
Expect questions along these lines:
For US-listed companies these questions sit inside management's assessment of internal control over financial reporting. For private companies with an external audit, the same topics come up in the auditor's understanding of controls. Either way, discuss them with your auditors before year-end, while there is still time to agree the approach and the testing without the closing timetable pressing on both sides.
What AI internal controls will not do
Controls do not make an AI tool more accurate. They detect and correct errors and leave evidence that someone looked. The quality of the output still depends on the source data, the design of the step and the experience of the reviewer.
They will not settle questions of interpretation for you. Whether a particular AI-assisted step is a key control, how much testing it needs and how a deficiency would be classified are matters for management and the auditors to work through together, with the facts of your company in front of them.
Frequently asked questions
What changes in internal controls when AI is used in financial reporting?
The control around an AI-assisted step has to cover the AI's output: what the reviewer checks and against what source, whether AI-produced information is complete and accurate, how changes to prompts and models are controlled, and what evidence is kept. The basic principles of control design stay the same.
Is the AI or the reviewer the control?
In most AI-assisted steps the human review is the control, and the AI output is an input to it. Write the control description so the reviewer is clearly accountable and the review is specific about what is checked.
How should prompts and AI models be handled under change management?
Treat prompts, instructions, source libraries and access settings as configuration: changes are requested, tested against a fixed set of cases, approved by someone other than the person who made them, and recorded. Ask vendors how model updates are announced and whether a version can be pinned.
Can an AI agent create a segregation of duties conflict?
Yes, an AI agent with rights to both create and approve a transaction, or to change vendor bank details and propose payments, is a conflict just as it would be for a person. Give each agent its own account and include it in the SoD matrix and access reviews.
What will auditors ask about AI in financial reporting?
Expect questions about where AI is used and in what role, the related controls, the completeness and accuracy of AI-produced information, change management for prompts and models, system access, and retained evidence. Discuss these with your auditors before year-end.
Are most finance teams ready for audit of AI-enabled processes?
Fewer than half say they are: in KPMG's May 2026 survey, 42% of organizations were fully assurance-ready for AI-enabled finance processes. The rest have work to do on controls, evidence or both before relying on AI in reporting.
Where CFOLogic fits
CFOLogic is not a CPA firm, and our work is not assurance. We design and build AI-assisted tools for finance teams with the review step and the record built in: read-only on the systems already in place, with a person approving what leaves the team. Command Center is what we call a management layer that sits over one finance process. It surfaces the items that need a person, with the owner and the waiting time, and records what was done. The P2P Command Center demo shows it on a fictional company's purchase-to-pay process.