# What is human-in-the-loop approval?

Human-in-the-loop approval is a control in which an AI system may propose an action but a named person must review and approve it before it takes effect. It is effective when the approval is enforced by the system, the approver is independent, and the decision is recorded.

- Source: https://finray.tech/intelligence/glossary/human-in-the-loop-approval/
- Published: 2026-11-21
- Updated: 2026-11-21
- Publisher: Finray Technologies Ltd, Cyprus Companies Registry HE 445903

---

## What is human-in-the-loop approval?

Human-in-the-loop approval is a control in which an AI system may propose an action but a named person must review and approve it before it takes effect. In financial software the action is typically a payment, a journal entry or a change to a limit. The control is only meaningful if the system enforces the approval, the approver is independent of the proposer, and the decision is recorded.

The term is often used loosely. A dashboard that shows what an agent did after the fact is monitoring, not approval. A button the agent can press for itself is not an independent check. The test is simple: can the action take effect without a person deciding?

## How does human-in-the-loop approval work?

A well-formed approval flow has these elements.

1. **Proposal.** The AI produces a proposal in a distinct state, separate from the executed action.
2. **Routing.** Rules decide whether the proposal needs approval and by whom: by amount, counterparty, account type or risk signal.
3. **Review surface.** The approver sees the proposal, the source material, and the results of automated checks.
4. **Decision.** The approver approves, edits or rejects. Edits are recorded as edits, not hidden.
5. **Execution.** Only an approved proposal proceeds. The execution is linked to the approval.
6. **Record.** Proposer, approver, time, rule version and outcome are stored in a form that is hard to alter and easy to query.

## Why does it matter for AI agents?

AI agents can produce plausible but wrong output, and can be influenced by text they read. The Model Context Protocol specification states that tools represent arbitrary code execution, that hosts must obtain explicit user consent before invoking a tool, and that descriptions of tool behaviour should be considered untrusted unless obtained from a trusted server. The same specification notes that MCP itself cannot enforce these principles at the protocol level, which means the consent and authorisation flows have to be built into the applications around it.

For payments, the AP2 announcement distinguishes human-present purchases, where the user approves the final cart, from delegated tasks where the user signs rules in advance. Both are forms of human authority; the second moves the human decision from the moment of purchase to the moment of delegation, which makes the limits in the delegation the control that matters.

## What do regulators expect?

The EU AI Act, for AI systems classified as high-risk, requires in Article 14 that such systems be designed so they can be effectively overseen by natural persons while in use. The article lists capabilities the overseer should have: to understand the system's capacities and limitations, to remain aware of the tendency to rely automatically on the output, to interpret the output correctly, to decide not to use the system or to override its output, and to intervene or stop the system. Whether a particular system is high-risk depends on its use and classification, and application dates are staged. Article 14 requires that oversight be possible; it does not require a person to approve each output.

FINMA's Guidance 08/2024 does not prescribe a specific approval mechanism. It expects supervised institutions to identify, assess, manage and monitor AI risks, and names operational and model risks, data risks, IT and cyber risks, third-party dependency and legal and reputational risks among its focus areas. An approval gate is one common way to manage the model-risk component for consequential actions.

## What can go wrong?

| Failure | Why it happens | Mitigation |
|---|---|---|
| Rubber-stamping | Volume is high and the review surface hides the decisive facts | Show source and key fields; sample approvals; track edit and rejection rates |
| Self-approval | The proposer and approver share an identity or role | Enforce different identities in the system |
| Bypass | An integration path skips the approval state | Make approval a property of the action in the data layer, not of one screen |
| Quiet policy change | Thresholds are changed without review | Treat changes to the approval policy as approved changes |
| Approval without authority | The approver cannot judge the matter | Route by competence; require escalation for out-of-range items |
| Lost context | The record shows approval but not what was approved | Store the proposal as approved, including edits |

## What is a worked example?

The numbers are synthetic. The example illustrates the control pattern; it is not a description of any one product's features. A finance team lets an assistant prepare supplier payments. The rules route any payment above CHF 5,000, any new payee, and any payment that differs from the invoice amount by more than CHF 50 to a named approver.

The assistant proposes three payments: CHF 2,300 to a known supplier, CHF 8,750 to a known supplier, and EUR 1,200 to a payee added yesterday. The first is below every threshold, so under these rules it proceeds without a person. The second is routed to an approver because of its amount. The third is routed because the payee is new. The approver for the third sees that the bank details differ from those on the supplier's earlier invoices, rejects it, and records the reason. The approver of the second sees the invoice, the proposal and the screening result, and approves; the record stores the proposal as approved.

The record can later show which payments needed a person, which did not, why, and what each person decided.

## What belongs outside the model?

- **The routing rules.** Evaluated by code, versioned, and changed under change control.
- **The state machine.** A proposal cannot execute until it is approved; the system enforces the transition.
- **Separation of duties.** The approver differs from the proposer and from the person who operates the AI.
- **The record.** Append-only, linked to the executed action and, where it exists, the ledger posting.
- **Escalation and timeouts.** An unapproved proposal expires or escalates; it does not default to executing.

## How does Finray relate to human-in-the-loop approval?

**Today.** Several Finray products are built around approval before effect. In [Corebanq Accounting](/platforms/corebanq-accounting/), postings and bills post only after a second person approves by default, as a company policy whose change is itself an approved change. A maker cannot approve their own change, and the database enforces this. Payroll runs need second-person approval, and VAT returns are prepared, independently reviewed and exported for the user to upload in the ESTV portal. The assistant is off by default, runs on a model the customer hosts locally, has no posting authority, and its proposals need a person's confirmation. Corebanq Accounting offers chat access from Claude or ChatGPT. Questions are read-only. Documents arrive as drafts; a person approves before anything posts. Chat access is off by default and activated per company. [Ordinis](/platforms/ordinis/) models multi-party approvals as state machines. [Corebanq](/platforms/corebanq/) role-based access control records the actor and the policy version with every action.

**Not yet.** Finray does not describe autonomous agents acting in these systems and does not claim agent integrations through the Model Context Protocol. The default in Corebanq Accounting is a company policy; the software does not claim that every change always needs four eyes.

Related reading: [agent mandate](/intelligence/glossary/agent-mandate/), [agentic payments](/intelligence/glossary/agentic-payments/), [agentic commerce](/intelligence/glossary/agentic-commerce/), [Model Context Protocol](/intelligence/glossary/model-context-protocol/) and the note [Agentic accounting under four eyes](/notes/agentic-accounting-four-eyes/).

## Frequently asked questions

### What is human-in-the-loop approval?

Human-in-the-loop approval is a control where an AI system can propose an action but a named person has to review and approve it before it takes effect. The approval is part of the workflow, enforced by the system, and recorded.

### Is human-in-the-loop the same as four eyes?

They overlap. Four eyes means a second person approves a change made by a first. Human-in-the-loop means a person approves an action proposed by an AI. When both apply, the approver should be independent of whoever operates or configured the AI.

### Does human review guarantee safety?

No. Reviewers can over-rely on a system's output, approve without reading, or lack the information needed to judge. The EU AI Act names this over-reliance risk in its human-oversight article. Review design and sampling matter as much as the existence of an approval step.

### When is approval required for an AI action in finance?

That is a risk decision for the institution. Common triggers are amount thresholds, new counterparties, unusual patterns, and any action that changes accounting records or moves money. The rules should be written down, enforced by the system, and changed only under change control.

### Can an AI agent approve its own work?

An AI agent should not approve its own work. Approval means an accountable person takes responsibility for the outcome. An approval given by the same system that proposed the action adds no independence.

## Sources

- [Regulation (EU) 2024/1689 (AI Act), Article 14, human oversight (EUR-Lex)](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
- [Model Context Protocol specification, security and trust principles](https://modelcontextprotocol.io/specification/latest)
- [FINMA Guidance 08/2024, Governance and risk management when using artificial intelligence](https://www.finma.ch/en/news/2024/12/20241218-mm-finma-am-08-24/)
- [Google Cloud blog, Announcing Agent Payments Protocol (AP2), 16 September 2025](https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol)
