Skip to content
Articles

Human Review of AI Compliance Checks: What Reviewers Decide

Human review of AI compliance checks works when reviewers see the evidence, can change the outcome and record why. What to route, show and log.

By ApioskAgent workflows7 min read

Human review of AI compliance checks is meaningful when three things are true: a named reviewer sees the evidence behind the result, that reviewer can change the outcome before it takes effect, and the decision is recorded with its reason. Route results to a person on defined triggers rather than by habit, and keep the source facts, the machine's matches and the AI's interpretation separate from the human decision.

This guide is for compliance, procurement and onboarding teams that let software or AI agents run company verification, sanctions screening and ownership checks. It explains which results should reach a reviewer, what the reviewer needs in front of them and how to keep the review from becoming a formality.

Human review of AI compliance checks, step by step

A review step answers one question: is this outcome right for this counterparty, given the evidence? The machine gathers and compares. The reviewer decides whether the comparison supports the outcome.

In practice the process has four parts:

  • A trigger decides which results go to a person.
  • A review packet shows the evidence, not just a verdict.
  • The reviewer accepts, changes or escalates the outcome.
  • A decision record captures who decided, on what evidence and why.

What you can expect is a smaller, better-informed queue, not the removal of judgement. Automation handles the clean, fully evidenced cases your policy allows it to close. People handle the cases where the evidence is incomplete, ambiguous or consequential.

Fast answers. Thin sign-offs.

Most failed reviews do not look like failures. They look like a reviewer clicking approve on a screen that said everything was fine.

The common patterns are familiar:

  • The reviewer sees a summary such as "no issues found" and never sees that one source timed out.
  • A possible sanctions match is closed because the name looked different, without comparing dates of birth, addresses or registration numbers.
  • Every result goes to review, the queue grows, and reviewers start approving in bulk.
  • The decision is recorded as "approved" with no reason, so nobody can reconstruct it a year later.

The underlying problem has a name. Article 14 of the EU AI Act asks that people overseeing high-risk AI systems stay aware of automation bias, the tendency to rely or over-rely on a system's output. A company-verification check is not automatically a high-risk system under that Act, but the warning applies to any review step where a confident machine answer sits in front of a busy person.

Decide who can decide

Before routing anything, write down who has authority over which outcomes. A reviewer who cannot reject a supplier, or who needs three approvals to do so, is under pressure to accept.

The mandatory inputs are short:

  • The decisions in scope, such as onboarding, refusing, continuing or ending a relationship, or changing the depth of checks.
  • The named role that may make each decision, and the role it escalates to.
  • The evidence your policy requires for each decision, marked mandatory or optional.
  • The triggers that send a result to review.

Optional but useful: written guidance on how to resolve a possible match, examples of past decisions and access to the original sources. Our guide to reviewing a sanctions screening name match covers that comparison in detail.

Regulation sets a floor in some jurisdictions. In the EU, Article 76(5) of the AML Regulation (EU) 2024/1624 allows obliged entities to base decisions on automated processes, including AI systems, but requires meaningful human intervention for decisions to enter into, refuse or maintain a business relationship, or to increase or decrease customer due diligence. The customer can obtain an explanation and challenge the decision. The regulation applies from 10 July 2027, and until then national rules based on the earlier AML directives continue to apply.

Machines flag. People decide.

The technical goal is a review packet where every layer is labelled for what it is. Keep these four apart:

  • Source facts: what a register or list returned, unchanged, with source name and retrieval time.
  • Algorithmic results: match scores, name similarity and rule outcomes, with the rule that produced them.
  • AI interpretation: summaries or explanations, labelled as interpretation and linked to the records they rely on.
  • Human decision: the outcome, the reviewer, the time and the reason.

Then define triggers. Good candidates are a possible match on a sanctions or PEP list, a mandatory item that was not returned, a source failure, a conflict between two sources, a change since the last check and any decision to refuse or end a relationship. Missing information is a trigger, never a pass. Our guide to missing company registry data explains how to label each gap.

The packet should let the reviewer check each claim against a record without rerunning the work. If the agent says the company is active, the packet links the register entry. If the agent says a match is unlikely, it shows the details that differ and the ones that were unavailable. Our article on source-backed company checks for AI agents lists fields worth keeping in that handoff.

Finally, record the decision as a new entry. Do not overwrite the machine's output with the reviewer's conclusion. The UK ICO's guidance on automated decision-making says human involvement must be active, not a token gesture, that the reviewer needs the authority to alter the decision, and that you should keep a record of how the human was involved. That guidance concerns decisions about individuals, which in company checks includes directors, sole traders and beneficial owners.

One queue. Then calibrated rules.

Implementation is mostly testing. Before relying on the review step, run it against cases where you already know the right answer.

  • Seed the queue with a true match, a false positive, a timed-out source, a missing mandatory field and a clean result.
  • Check that each one reaches the right reviewer with the evidence needed to resolve it.
  • Track how often reviewers change the machine's outcome. A rate near zero can mean the triggers are too broad or the reviewers are not looking.
  • Sample results that were closed without review, because errors there never reach the queue.
  • Set a time limit for open items, and escalate rather than auto-approve when it passes.
  • Re-review when something changes, such as a new list entry or a change in ownership.

A fictional example shows the flow. An agent checks Kestrel Brook Packaging GmbH before a first payment. The register confirms an active company. The sanctions screen returns a possible match on a similarly named entity in another country, and the ownership source fails to answer. Two triggers fire. The reviewer opens the list entry, sees a different country and a different registration number, and records "match discounted: identifiers differ" with the entry reference. The ownership item stays open, the payment stays on hold, and the check is retried the next morning. Nothing is approved until the mandatory item is resolved.

There are trade-offs. Narrow triggers keep the queue small but let more errors pass unseen. Broad triggers catch more but tire reviewers, which brings back the bulk approvals you were trying to avoid. Sampling and tracking changed outcomes is how you find the balance for your own volume and risk.

How we approach it

We built Apiosk to supply the evidence a reviewer needs, not to replace the reviewer. As our agents page puts it, Apiosk retrieves and combines the evidence; a person decides, or your own rules do.

Each result names the sources it used, and when a source has no record, the answer says the record is missing rather than guessing. Sanctions and PEP screening returns match or no match with the score, the list any match came from and a timestamped record of the check. That is the input to a review, not its conclusion. You can see which registers and lists are connected on our sources page.

Connected agents work within your team's spending limits and approval rules, and the cost of a check is shown and approved before it runs. For items that change after onboarding, change alerts deliver only what changed, which gives a reviewer a reason to look again. Your routing rules, reviewer roles and decision records remain your own process, in your own systems.

Reviews turning into rubber stamps? Show the evidence.

If your reviewers approve results they cannot trace, see the checks Apiosk runs and give every decision a source a person can open.

This article was prepared with AI assistance using the linked EUR-Lex, ICO and EU AI Act texts and the current Apiosk product pages. It is not legal advice, and Kestrel Brook Packaging GmbH is fictional.

Frequently asked questions

Does EU law require a person to review every AI compliance check?

Not every check. Article 76(5) of the EU AML Regulation, which applies from 10 July 2027, lets obliged entities use automated processes, including AI systems, but requires meaningful human intervention for decisions to enter into, refuse or maintain a business relationship, and for decisions to increase or reduce customer due diligence. Retrieving a record or computing a match score is not such a decision. Check which rules apply to your entity and sector, and take legal advice where it matters.

What makes a human review meaningful rather than a formality?

The reviewer sees the evidence the result rests on, acts before the outcome takes effect, has the authority to change it and records why. The UK ICO's guidance on automated decision-making says involvement must be active and not a token gesture, and that the reviewer needs the discretion and authority to alter the decision. A person who approves a summary without opening the source is closer to a signature than a review.

Can an AI agent clear a no-match sanctions result on its own?

That depends on your own policy, not on the agent. Some teams let rules close a clean result automatically and review only possible matches. Before allowing that, confirm the check actually completed, that the right entity and lists were screened, and that you sample cleared results. A no-match result means the screened lists returned nothing for those inputs. It does not show that the counterparty carries no risk.

Should the reviewer be able to see the AI's reasoning?

Yes, labelled as interpretation. Show the AI summary next to the source records it claims to summarise, never in place of them. The reviewer should be able to check each statement against a record, and to disagree with the summary without having to rerun the whole check.