Keeping a Human in the Loop: AI Controls for Finance

Of all the ways a finance team can use AI, one principle does more than any other to keep that use safe: keeping a human in the loop. A finance professional keeps the judgement, the verification and the accountability, and AI assists rather than replaces them. It is the central control because it guards against AI’s fundamental risk, that it can produce output which is fluent, confident and wrong, by making sure a person checks and owns the work instead of trusting it blindly.

This guide is for finance teams and the people who lead them. It covers what keeping a human in the loop means, why it matters in finance, how to build it as a real control rather than a slogan, where it matters most, and how it underpins every other safe use of AI. It then goes further into the practical detail that controllers and auditors ask about: review tiers by risk, segregation between preparer and reviewer when AI produces the first draft, automation bias, approval gates for agentic AI, and the logging and evidence that show a review actually happened. The aim is a control that captures AI’s value without letting its errors flow into the numbers.

What Keeping a Human in the Loop Means

Keeping a human in the loop means a finance professional stays actively involved in, and responsible for, the work AI assists with. AI produces drafts, analysis, suggested matches or proposed journals; the finance professional reviews, verifies and decides. The human is the one who ultimately owns the output, with AI acting as an assistant rather than an autonomous operator.

The opposite is taking the human out of the loop: using AI output without verification, or letting AI act on systems without oversight. That removes the one check that catches AI’s errors. Because AI can be confidently wrong, keeping a human in the loop is what stops those errors reaching the ledger, the management pack or the statutory accounts.

It helps to separate three levels of involvement, because teams often blur them:

• Human in the loop — a person approves each output or action before it takes effect. AI suggests a bank match; a person accepts or rejects it.
• Human on the loop — AI acts within set limits and a person monitors results and exceptions after the event, with power to intervene. Suitable only for low-value, reversible, well-tested tasks.
• Human out of the loop — AI acts with no review. For anything that affects the books, payments or external reporting, this is not acceptable in a finance function.

Most of this guide is about choosing the right level for each use and proving that the chosen level operates in practice.

Why It Is Essential in Finance

Finance is a domain where AI’s fundamental risk is unusually costly. Output that is plausible but wrong can misstate a balance, mislead a board, trigger a payment or breach a covenant. Accuracy is not a nice-to-have, and errors have legal, regulatory and commercial consequences. Human verification is what catches those errors before anyone relies on the work.

Finance also rests on judgement and accountability that cannot be delegated to software. Directors remain responsible for the accounts; a Financial Controller signs off reconciliations; a reviewer puts their name to a journal. AI cannot carry that accountability or exercise professional judgement about estimates, provisions or disclosures. The profession’s own guidance points the same way: the ICAEW’s generative AI guide on ethics expects accountants to understand how a tool works, including its risks and limitations, before relying on it, and flags self-review as an ethical threat, while the Financial Reporting Council’s guidance on AI in audit, published in June 2025, sets expectations for proportionate documentation of AI tools. Without a human in the loop, finance work would rest on unverified output with nobody accountable for it, which is untenable.

How to Implement It as a Control

A principle only becomes a control when it is built into how work is done, so that it happens reliably rather than when someone remembers. The core requirement is simple: AI output is verified by a competent finance professional before it is relied upon, and anything factual or consequential is checked against source data, not just read for sense. That requirement belongs in the team’s AI usage policy and in the procedures for each process that uses AI.

Implementation also means the finance professional keeps the decisions. AI informs judgements about estimates, provisions or write-offs; it does not make them. Where AI can act on systems, consequential actions need human authorisation, as covered in our guide to connecting AI to finance data safely. Policy, training and a shared understanding of why the checks matter turn this from a principle into something the team actually does.

Design it like any other control

The easiest way to make human review stick is to document it in the same way as any other key control in the financial control framework. For each AI-assisted process, record:

• The risk — what could go wrong if the AI output were wrong (misstatement, wrong payment, misleading commentary, data leak).
• The control owner — a named role, not “the team”.
• The review step — what the reviewer checks, against what source, and to what threshold.
• Frequency — every item, every batch, or a sample.
• Evidence — what record shows the review happened (sign-off, system approval, annotated working).

Growing businesses that have not yet formalised their controls can start with the approach in our guide to internal controls for growing businesses and add AI steps to it, rather than building a separate AI framework that nobody maintains.

Where It Matters Most

Human review matters across all AI use, but it matters most where the stakes are highest, and knowing where helps a team focus its effort. It matters most where output is consequential, feeding the financial statements, board decisions or external reporting, because an uncaught error costs most there. It matters most where AI can act rather than advise, because an unchecked action causes direct harm. It matters most for accuracy-critical work where the numbers must be right, and wherever sensitive data or serious consequences are involved.

Review tiers by risk

A tiered approach makes “proportionate” concrete. The table below is a starting point; each team should map its own AI uses to tiers and record the result in its policy.

The tiers apply across processes covered in our guides to AI for reconciliations, AI in the month-end close and AI for AP and AR. A use can move up a tier if the output goes external, or down a tier once it has a long clean track record and robust automated checks, but that decision should be made deliberately and recorded.

How It Underlies the Safe Use of AI

Keeping a human in the loop is the common thread through every safe use of AI in finance. Whether the use is the close, reconciliations, reporting, analysis or forecasting, safe use means AI assists while the finance professional verifies, judges and takes ownership. Every other safeguard — data security, tool selection, prompt discipline — sits around this one.

A team that adopts human review as its core AI control has a foundation for every application, because whatever the use, the control guards against AI’s risk and supplies the judgement and accountability. A team that abandons it anywhere loses that safety for that use, however good its other safeguards are. This connects to the broader view in where AI helps and where it is dangerous and the governance in AI usage policies for finance teams; data handling is covered separately in AI data security in finance.

Human in the Loop Without Negating the Value

Human involvement has to be heavy enough to provide real safety and light enough to leave AI’s time saving intact. If every AI-drafted email gets a two-person review, the benefit disappears; if AI-proposed journals get a glance, the safety disappears. The goal is meaningful verification where it matters and lighter, still genuine, oversight where it does not.

That is what the tiers above achieve: rigorous review for consequential, high-stakes or action-enabled uses, and proportionate checks for low-stakes ones. Two further techniques help. First, design AI outputs to be easy to check — ask for figures with their source references, or for a list of the assumptions made — so verification takes minutes, not hours; our finance prompt library and guide to prompt engineering for finance include prompts written this way. Second, combine human review with automated checks (control totals, sum-checks, tie-outs to the trial balance) so the reviewer’s attention goes to judgement, not arithmetic. A team that calibrates like this gets both the safety and the value.

Segregation Between Preparer and Reviewer When AI Drafts

Segregation of duties assumes two people: one prepares, another reviews. AI quietly changes that. When a Management Accountant uses AI to draft a reconciliation or a journal and then “reviews” the AI’s work, the same person has effectively prepared and checked it. The AI is not a preparer in any accountable sense, so the self-review is not an independent check.

Practical rules that keep segregation meaningful:

• The person who ran the AI is the preparer. They are accountable for the output as if they had written it by hand, and they sign as preparer.
• Tier 3 and 4 work needs a second person. The reviewer should not be the person who wrote the prompt or accepted the AI’s suggestions.
• Reviewers should know AI was used. Flag AI-assisted workings so the reviewer looks for AI-typical errors: invented figures, misread columns, plausible but wrong accounting treatments.
• Small teams need compensating controls. Where there is no second accountant, use a Financial Controller or Finance Director review, an outsourced reviewer, or a deeper monthly analytical review of AI-touched balances.
• System permissions must reflect this. An AI service account should never hold approval rights, and a user should not be able to approve a journal their own AI session created.

This is where the role of a Financial Controller using AI differs from that of a preparer: the controller’s value increasingly lies in independent review of AI-assisted work, not in producing it.

Automation Bias: Why Reviews Fail

The biggest threat to a human-in-the-loop control is not missing policy but a reviewer who stops genuinely reviewing. Automation bias is the well-documented tendency to over-trust automated output — accepting a suggestion because the system made it, and failing to notice errors the system did not flag. AI makes it worse because its output is fluent and well formatted, so wrong answers look as polished as right ones. After weeks of correct suggestions, a reviewer can slide into approving on sight.

Ways to counter it:

• Check to source, not to sense. The review step should require tracing figures to the ledger, bank statement or contract, not reading the narrative and agreeing with it.
• Make the reviewer state something. A short note on what was checked is better than a tick, and it creates evidence.
• Watch acceptance rates. If a reviewer accepts every AI-suggested match or journal over many months, that is a signal to test the review, not proof the AI is perfect.
• Re-perform a sample. Periodically have someone redo AI-assisted work from scratch and compare.
• Rotate and refresh. Rotate reviewers on high-tier processes, and share examples of AI errors the team has caught so people know what to look for.

Agentic AI: Approval Gates Before AI Acts

Most finance AI use today is advisory: a person asks, AI answers, and the person decides what to do. Agentic AI is different. These tools can take steps on their own — opening files, querying the ledger, drafting and submitting journals, updating spreadsheets or sending emails. That moves the risk from bad advice to bad actions, and the control needs to move with it.

The control is an approval gate: a point where the agent stops and a named person must approve before anything consequential happens. Good gate design includes:

• Read-only by default. Agents get read access first; write access is granted per process, per system, after testing.
• Gates on every write that matters. Posting journals, releasing payments, changing supplier bank details, editing master data or sending anything external always needs human approval.
• Approval within delegated authority. The person approving an agent’s action must hold the same authority they would need to do it themselves; the agent does not create new authority.
• Show the approver what will happen. The gate should display the exact entry, amount and accounts, plus the agent’s reasoning and sources, not a summary.
• Limits and kill switch. Cap values and volumes the agent can propose, and make it simple to suspend an agent immediately.
• No self-approval, no chaining around gates. An agent must not be able to approve its own action or break a large action into small ones below a threshold.

The UK’s NCSC guidelines for secure AI system development are a useful reference for the IT side of this, and the integration points are covered in our guide to connecting AI to finance data.

Logging and Evidence of Review for Auditors

A review that left no trace did not happen, as far as an auditor is concerned. External auditors increasingly ask which processes use AI, how output is checked and what evidence exists; our guide to what auditors ask about AI covers the questions in detail. This section covers the records that answer them.

What to log

• An AI use register — each process using AI, the tool, the risk tier, the control owner and the review step.
• Inputs and outputs for tier 3 and 4 work — the prompt or instruction, the data supplied and the output, retained with the working papers for the period.
• Reviewer evidence — who reviewed, when, what was checked to which source, and what was changed.
• System audit trails — for agentic tools, the requested action, the approver, the timestamp and the outcome, ideally in the ledger’s own approval workflow.
• Errors caught and incidents — a short log of material AI errors found in review and any that got through, with the fix. This is powerful evidence that the control works.
• Tool and version changes — when a model or tool changes, note it, because behaviour can change too.

Keep it proportionate. Tier 1 use needs no special logging. The FRC’s AI in audit guidance stresses proportionate documentation for audit firms’ own tools, and the same thinking works for a finance team: enough to show the control operated, not a transcript of every chat. Our guide to AI for audit preparation covers using AI to get ready for the audit itself.

Legal and Regulatory Context

Human oversight is also, in places, a legal requirement. Under UK data protection law, significant decisions about individuals made solely by automated means must come with safeguards, including the ability to obtain meaningful human intervention and to contest the decision; the rules were revised by the Data (Use and Access) Act 2025, with the changes taking effect in February 2026. The ICO’s guidance on automated decision-making is the place to check. That can apply to finance processes such as automated credit decisions on customers or supplier vetting.

Groups operating in the EU should also be aware that the EU AI Act requires human oversight for systems it classes as high-risk (Article 14), a category that includes some credit-scoring uses. Most everyday finance-team AI use will not fall into that category, but legal advice is worth taking where a process makes decisions about people. FCA-regulated firms have the added overlay of senior manager accountability, covered in our guide to SM&CR for finance teams.

What This Means When You Hire

The control is only as good as the people operating it. When we assess candidates for Financial Controller, Finance Manager and accountant roles, the useful question is not whether someone uses AI but how they check it. Strong candidates can describe a specific AI error they caught, explain how they evidence their reviews and say where they would not use AI at all. That shift is covered in our article on how AI is changing what employers look for in an FC. For senior regulated roles where personal accountability is formal, sister firm SMF Capital handles SMF appointments, and FD Capital handles FD and CFO searches.

Frequently Asked Questions

Does every AI output need a second reviewer?

No. Low-risk uses such as drafting internal emails or explaining a formula only need the user to read and correct the output. A second, independent reviewer is needed for high-risk work: journals, provisions, statutory or regulatory reporting, tax computations and any AI action on live systems.

If I check the AI’s work myself, is that a review?

It is a preparer’s check, not an independent review. The person who ran the AI is the preparer and is accountable for the output. For higher-risk work, someone who did not run the AI step should review it.

What evidence will our auditors want?

Typically a list of processes using AI, the review step for each, and evidence that the review happened: sign-offs, system approvals and, for significant areas, the retained input and output. See what auditors ask about AI for the full set of questions.

Can an AI agent post journals if a person approves them?

It can, provided the approval gate is real: the approver sees the exact entry, holds the authority to post it themselves, and the agent cannot self-approve or split entries to avoid limits. Start read-only and grant write access process by process.

How do we stop reviewers rubber-stamping AI output?

Require checks to source rather than to sense, ask reviewers to note what they checked, monitor acceptance rates, re-perform a sample periodically and share examples of AI errors the team has caught.

Where should human-in-the-loop rules be written down?

In the team’s AI usage policy, which sets the principle and the tiers, and in each process’s procedures and control documentation, which set the specific review step and evidence.