Why Compliance AI Can't Fully Automate (Human in the Loop)

AI already screens customers, scores transactions and closes alerts faster than any compliance team could. So why not let it run on autopilot? Because in compliance the hard part is making a decision you can defend to a regulator, a customer and, sometimes, a court.

In some cases human oversight is a legal requirement. From July 2027, the EU's new Anti-Money Laundering Regulation will require meaningful human intervention in automated decisions about customers. In others, it is a design choice. And a 2024 meta-analysis of 106 experiments shows that simply putting a person in front of an AI recommendation does not reliably improve results.

In June 2026, Bank of England Deputy Governor Sarah Breeden said that relying on a human in the loop for all agent actions is "unlikely to be realistic" (Reuters). The real question is where humans should stay involved, how, and with what evidence.

Key takeaways

• Some compliance decisions legally require a human. From 10 July 2027, the EU's AMLR requires meaningful human intervention in automated decisions to accept, refuse or end a customer relationship.
• The AI Act covers only part of compliance AI. Its human oversight rules apply to high-risk systems such as credit scoring from December 2027, and many anti-money laundering and fraud tools fall outside that category.
• AI pays off on volume. Google Cloud reported that HSBC found two to four times more suspicious activity with over 60% fewer alerts.
• A human in the loop is not enough on its own. A meta-analysis of 106 experiments found human–AI teams did worse, on average, than the better of the two working alone.
• Oversight has to be designed. Route decisions by risk, show reviewers the model's confidence and measure how often they override it.
• Accountability stays with the firm. A firm remains fully liable for outsourced tasks, and the audit trail is how it proves oversight happened.

What is human in the loop (HITL)?

Human in the loop (HITL) is a model of AI oversight in which a person must review, approve, correct or reject an AI output before it takes effect. The AI system proposes; a human decides.

This differs from "there's a human somewhere in the process". If an analyst only reads a monthly report about decisions the model has already executed, the human is not in the loop.

The European Commission's High-Level Expert Group on AI described this spectrum back in 2019 in its Ethics Guidelines for Trustworthy AI. It also set a useful rule: the less oversight a human can exercise over an AI system, the more extensive testing and stricter governance it needs. Every step you automate fully needs evidence that it works without a human.

Human-in-the-loop vs human-on-the-loop vs human-over-the-loop

These terms are often used as synonyms, but they mean different degrees of human involvement. Frameworks also don't fully agree on them. The EU guidelines use "in the loop", "on the loop" and "in command", while Singapore's Model AI Governance Framework (2020) uses "in", "over" and "out of the loop".

In practice, a mature compliance program uses the first three models at once. Human-in-the-loop for decisions with legal effect on a customer, human-on-the-loop for high-volume, low-risk work, and human-over-the-loop for the system as a whole. The real design work is deciding which decisions belong to which model.

Where AI genuinely helps compliance

Arguing for human oversight doesn't mean arguing against AI. In its 2021 report on new technologies, FATF noted that technology makes AML/CFT faster and better, but manual review is still needed for residual risk. AI earns its place where volume overwhelms people.

AML and transaction monitoring

Rule-based transaction monitoring generates large volumes of alerts that analysts must clear by hand. Machine learning models can rank alerts by risk, spot patterns across accounts and surface links that static rules miss. Google Cloud reported that HSBC, using its anti-money laundering AI, identified two to four times more suspicious activity while cutting alert volumes by more than 60%.

KYC/KYB, sanctions and PEP screening

Screening customers and businesses against sanctions lists and PEP databases involves fuzzy name matching and incomplete data. AI resolves entities across spellings and pre-populates customer files for an analyst.

Alert triage and fewer false positives

AI can enrich each alert with context, score its risk and auto-close only the clearly benign ones. The analyst's queue gets shorter and the cases in it are more likely to matter.

Document review

Generative AI can extract data from registration documents, contracts and proof-of-address files and flag inconsistencies, so a reviewer starts from a structured draft instead of a pile of PDFs. In each of these cases, AI does the preparatory work. The decision that carries legal or reputational weight still needs human judgment.

Why compliance AI can't be fully automated

The limits of compliance AI fall into six groups. Some come from the law, some from the technology and some from the people who use it.

Grey areas and context

Compliance decisions often hinge on things the data does not show: intent, the plausibility of a business story, the reputational risk of a client who is legal but controversial. A model can tell you that a transaction is unusual. It cannot tell you whether the customer's explanation makes sense. That is why the EU's Anti-Money Laundering Regulation (AMLR), applicable from 10 July 2027, allows automated processes and AI. It requires meaningful human intervention in decisions to enter into, refuse or end a business relationship, carry out a transaction, or change the extent of due diligence.

Explainability: regulators and customers need a reason

A risk score is not a reason. The AMLR gives customers the right to obtain an explanation of a decision and to challenge it, with the exception of suspicious activity reports. Under GDPR Article 22, the EU Court of Justice has gone further. In the SCHUFA case (December 2023), it held that a credit score can itself count as an automated decision when a lender relies heavily on it. In Dun & Bradstreet Austria (February 2025), it ruled that the data controller must explain the procedure and principles actually applied. Handing over a complex formula is not enough. A black box model that nobody in the firm can explain becomes a legal problem.

Bias in data and automation bias in people

Models learn from historical decisions, so training data can carry yesterday's blind spots into tomorrow's alerts. Humans add a second problem on top: automation bias, the tendency to trust a machine's output even when it is wrong. The EU AI Act names this risk explicitly in Article 14, which requires people overseeing high-risk AI systems to stay aware of it.

The effect is real even among experts. In a 2023 study in Radiology, experienced radiologists' accuracy fell from 82% to 45.5% when a supposed AI suggested the wrong category. The study concerned mammography, not compliance, but the mechanism is the same: a confident AI suggestion pulls the human reviewer with it.

"A human in the loop tends to over-rely on AI, most often because the process isn't defined: nobody has decided what can be delegated to the model or how to check that it works. When a task feels unstructured and ad hoc, people skip planning the AI's work and setting acceptance criteria, so they can't tell whether the output is correct. You can see it right away with developers: some ask critical questions about LLM-generated code and review it, others accept the changes without thinking about edge cases or consequences."

Dawid Dziedzic Python Developer at Miquido

You can't hand accountability to an algorithm

Under the AMLR, a firm remains fully liable for any task it outsources, including to a technology provider. It must also show its supervisor that it understands how that provider's tools work. A regulator will not accept "the model decided" as an answer.

The UK's FCA has shown what that means in practice, even for traditional rule-based systems. In 2024 it fined Metro Bank £16.7 million because its automated system failed to monitor over 60 million transactions worth more than £51 billion. Staff had raised the problem in 2017 and 2018, but it took years to fix. What failed were the controls and escalation around the system.

New patterns, adversarial attacks and zero-days

Models detect what resembles the past. Criminals adapt to the future. A new typology, a sanctions-evasion scheme or a zero-day exploit may look normal to a model trained on yesterday's data. AI systems can also be attacked directly. NIST's taxonomy of adversarial machine learning (2025) lists attacks such as evasion, data poisoning and prompt injection against AI agents. Attackers already use AI too: according to IBM's 2026 Cost of a Data Breach report, one in four malicious breaches was AI-enabled, 56% more than a year earlier.

Shadow AI

The oversight problem doesn't stop at official systems. Employees paste customer data into public chatbots or connect unapproved AI tools to internal workflows. IBM’s 2025 study looked at 600 organisations that had suffered a data breach. One in five of them said the breach was due to shadow AI. Those with high levels of shadow AI saw breach costs $670,000 higher on average than those with low levels or none. You cannot audit a decision made in a tool you don't know exists.

How AI SOC teams split work between agents and analysts

Security operations teams face the same dilemma as compliance teams, at machine speed. Alert volume keeps growing, and agentic AI promises to triage, investigate and even contain threats on its own. The emerging AI SOC model splits the work roughly like this:

• AI agents enrich and correlate alerts, close obvious false positives and draft incident summaries with suggested response actions.
• Analysts decide on actions with real consequences: isolating a production server, disabling an executive's account, notifying a regulator. They also handle anything the agent flags as low confidence or unfamiliar.
• Engineers and SOC leads stay over the loop: they tune detection rules, set the agent's permissions and review its mistakes.

Gartner's 2026 Hype Cycle for Security Operations, as reported by Cybersecurity Insiders, places AI SOC agents at the peak of inflated expectations, with 1–5% market penetration and five to ten years to mainstream adoption. For the foreseeable future, AI SOC human oversight remains the operating model.

How to design human oversight that actually works

Adding a reviewer is easy; making the review meaningful takes design. A 2024 meta-analysis in Nature Human Behaviour covering 106 experiments found that human–AI combinations performed, on average, worse than the better of the two working alone. Gains appeared mainly in tasks where humans had an edge over the AI. The lesson for compliance is that a human in the loop needs to be placed where they add value.

Route by risk, not by volume

Define in advance which decisions always go to a compliance officer or analyst: those with legal effect on a customer, those above a value or risk threshold, and those where the model's confidence is low. Everything else can run human-on-the-loop, with sampled quality checks.

"Human oversight is needed wherever you need a flesh-and-blood person to take responsibility for what the AI does, and that’s what compliance comes down to: naming the person accountable for a given area. AI can make that person’s work faster and let them cover more ground at once. In evaluations, for example, a model checks hundreds of qualitative datasets against agreed criteria. A human writes the instructions and spot-checks the model’s verdicts, selectively or statistically rather than case by case. It’s far less work than manual evaluation, but it isn’t blind trust that the model will catch every case on the first prompt."

Dawid Dziedzic Python Developer at Miquido

Show confidence, not persuasion

How an AI recommendation is presented changes how people judge it. A study by Microsoft Research and the University of Washington found that AI explanations made people more likely to accept recommendations whether or not they were correct, and did not outperform simply showing the model's confidence. Reviewers need the evidence and the uncertainty, not a convincing story.

Measure overrides and agreement

If reviewers agree with the AI 99.9% of the time, it can mean the model is excellent, or that nobody is really checking. Track override rates, review times and agreement rates by reviewer. Too few overrides may signal blind trust in the model. Too many may signal that the model or the thresholds need tuning.

Test the oversight before you rely on it

Legal scholar Ben Green, after analysing 41 policies that mandate human oversight of algorithms, concluded that people often cannot perform the oversight these policies assume. He recommends checking whether they can before a system goes live. In practice: run a pilot, plant known errors in the queue and check whether reviewers catch them.

Oversight rules only work if the system enforces them: escalation logic, confidence thresholds, permissions for AI agents and the interface the reviewer uses. At Miquido, we design these layers together with the model in our AI development services.

Audit trail and accountability

An audit trail is the proof that oversight happened. Without it, human review is just a claim. Regulators will ask not only what decision was made, but who made it, on what basis and whether the AI output was challenged.

A useful audit trail for an AI-assisted compliance decision records:

• The input data and its source
• The model and its version
• The AI recommendation and its confidence score
• The escalation rule that sent the case to a human, or let it close automatically
• Who reviewed it and when
• The final decision and its rationale
• Whether the reviewer overrode the AI

This record serves three audiences. Supervisors can verify that the firm understands and controls its tools, as the AMLR requires. Customers can receive an explanation they are entitled to. And the firm can learn from its own overrides and errors.

If your team is planning AI for screening, monitoring or alert triage, start with the oversight map: which decisions stay human, what triggers escalation and what gets logged. We help design that map into the system from day one. Talk to our AI team.

Top AI innovations delivered monthly!

The administrator of your personal data is Miquido sp. z o.o. sp.k., with its ... registered office in Kraków at Zabłocie 43A, 30 - 701. We process the provided information in order to send you a newsletter. The basis for processing of your data is your consent and Miquido’s legitimate interest.You may withdraw your consent at any time by contacting us at marketing@miquido.com. You have the right to object, the right to access your data, the right to request rectification, deletion or restriction of data processing. For detailed information on the processing of your personal data, please see Privacy Policy.

Show more