Which AI decisions need human approval in financial services, what US regulators expect from human-in-the-loop review, and how to avoid rubber-stamp approvals.
Human-in-the-loop AI helps financial services firms maintain meaningful human oversight of AI-driven decisions. Lenders, insurers, and advisers must determine which decisions require human approval, establish effective review processes, and ensure reviewers can challenge or change AI outcomes.
Many financial firms already keep a person close to their AI decisions. In the Bank of England and FCA's 2026 survey, 80% of UK firms answering questions on AI agents had agents in production that recommend actions while "humans decide and execute." Only 2% let agents act with broad autonomy. In the US, 48% of adviser firms have a formal policy, plan, or procedure for human-in-the-loop oversight of AI outputs, according to the IAA, ACA Group, and YCC survey.
New state rules set a higher bar for that person. California's rules, for example, ask whether the reviewer understands the AI's output, considers other relevant information, and has the authority to change the decision.
This guide to human-in-the-loop AI sets out which AI decisions at lenders, insurers, and advisers need human approval, what US rules expect from a review, and how to make that review work at scale. For the wider picture, see our guide to AI governance in financial services.
These three terms describe how much a person controls an AI decision. A firm can use all three for different decisions.
The right model depends on the decision, which is what the rest of this guide covers.

No US rule lists which AI decisions need a person. Several now describe what meaningful human oversight of AI looks like.
Outside the US, the FCA has warned against "treating human review as a substitute for addressing weaknesses in the underlying system" (Clifford Chance). For firms with EU exposure, the EU AI Act requires high-risk systems such as credit scoring to be designed for effective human oversight, with those rules now applying from December 2, 2027 (Mondaq).
Two questions decide where a person belongs in an AI decision.
What does a wrong decision cost? Consider the customer, the firm, and the regulator. A wrongly declined loan application costs the applicant, and it can raise fair lending questions. A clumsy internal summary costs very little.
How easily can it be reversed? A card blocked by mistake can be unblocked in minutes. A loan offer sent to the wrong customer, or an investment trade, is much harder to take back.
High cost and hard to reverse means a person approves before the decision takes effect. High cost but easy to reverse means the AI can act while a person monitors and corrects quickly. Low-cost decisions can be sampled, before release if they're hard to reverse, or afterward if they're easy to fix.
Here's how that plays out for common decisions in financial services, from AI credit decisioning to customer messages. Your own answers will depend on your products, customers, and risk appetite.
Decision | Cost of a wrong decision | Easy to reverse? | Human role |
|---|---|---|---|
Loan approval or decline | High for the applicant, with fair lending exposure | Hard once communicated | A person approves declines and edge cases. Approvals inside policy can run automatically with regular sampling |
Insurance claim denial | High for the policyholder | Hard once communicated | A person approves every denial |
Fraud alert that blocks a card or payment | Medium, mainly customer friction | Easy | The AI acts, and a person monitors and handles appeals quickly |
AML transaction alert | High, mainly regulatory | Easy at the alert stage | The AI flags and prioritizes, and a human investigator decides what to escalate |
Investment recommendation to a client | High, with suitability risk | Hard once the client acts | An adviser approves it before it reaches the client |
Customer message drafted by generative AI | Medium to high if it includes a commitment, price, or disclosure | Hard once sent | A person reviews regulated content before sending, and routine replies are sampled |
Internal summary or research note | Low to medium | Easy | Sample after the fact |

A reviewer who approves almost everything the AI suggests adds cost and very little control. The FCA has described a case where human review "became a box-ticking exercise" as reviewers accepted AI recommendations without enough scrutiny. Gradient Labs describes the same pattern from its founders' time at Monzo. A reply-suggestion tool was right about 90% of the time, and reviewers "learned to click accept, quality drifted, and efficiency barely moved" (Gradient Labs).
Review works when four conditions are met.
"Most review screens show the AI's answer and an approve button. A reviewer also needs the reasons and the evidence behind them. Without those, approving becomes the only realistic option."
Mukul Juneja, Founding Partner & CTO, Muoro.
Track a few numbers for each type of decision, and look at them monthly.
Measure | What it tells you |
|---|---|
Override rate | How often reviewers change the AI's decision. A rate near zero needs investigating |
Second-review disagreement | How often a second reviewer disagrees with the first, on a sample of cases |
Time per review | Whether reviewers have time to judge each case properly |
Decisions overturned on appeal | Whether reviewed decisions hold up when customers challenge them |
Errors found in later sampling | What review is missing |
Log each human review alongside the AI decision. Record who reviewed it, when, what the AI recommended and why, what the reviewer saw, their decision, and their reason for any change. That record supports Colorado's adverse outcome explanations, California's access requests, and FINRA's effective practice of storing prompt and output logs. It also shows examiners the review actually happened.
At high volumes, review capacity becomes the practical constraint. At Colorado's rulemaking hearing in October, Affirm said it makes 37.8 million credit decisions a year and declines about 15% of applicants. Affirm estimated that could mean 11 human reviews a minute under Colorado's law, according to The Sum and Substance.
To plan for it, estimate review volumes from your decline or denial rates, set service levels for each decision type, and sample lower-risk decisions instead of reviewing every one. Feed reviewer overrides back to the model team as labeled examples. That kind of human feedback in AI governance can improve the model and, over time, reduce how many cases need a person.
Some agents act on their own instead of recommending, which changes where a person belongs. FINRA lists agents acting beyond their "intended scope and authority" among the risks firms should consider. Put an approval gate in front of any action that's costly or hard to reverse, such as moving money, changing a customer's account, or sending a commitment. Then name a person who can pause the agent, and test that they can. For who should own each stage of approval, see our AI governance checklist.
Not sure which of your AI decisions need a person, or whether your reviewers can change the outcome? Muoro's AI governance assessment maps your AI-influenced decisions, checks how human review works in practice, and gives you a prioritized plan to close the gaps. Book your AI governance assessment.
A person reviews and approves each AI decision or output before it takes effect. In financial services, that's usually a credit officer, claims handler, or adviser who can accept, change, or reject what the AI recommends.
With human-in-the-loop, a person approves each decision. With human-on-the-loop, the AI acts while a person monitors and can step in. Many firms use both, depending on how costly and how reversible each decision is.
Yes, for many lower-risk decisions, such as routing a query or flagging a transaction. For consequential decisions like credit or claims denials, Colorado and California give consumers a route to human review, and many firms keep a person involved in the final decision.
Give reviewers the authority to change the outcome, the information to judge it, and the time to do it. Then track how often they override the AI, and investigate if the rate falls close to zero.
No. Base it on what a wrong decision would cost and how easily it can be reversed. Have a person approve high-impact decisions, monitor medium-risk ones, and sample the rest.
Someone with the expertise to understand the output and the authority to change the decision, as California's rules require. In most firms, that person sits in the business team that owns the decision.
Move beyond pilots. Build AI systems that drive real decisions and measurable outcomes.
TALK TO US