AI agents for business finance: what they can and can't do
What AI agents can do in bookkeeping, close, collections and compliance today, where a person must stay in charge, and how to judge a product.
Part of the guideAI agents for business finance: what they can and can't do
The short answer
A person should make, or approve, any finance decision that moves money above a set limit, changes who gets paid, signs a legal filing, decides something significant about an individual, or is new to the business. Below those lines an AI agent can act and log. The review only counts if the approver sees the evidence, has time to check it and can say no without friction.
"Human in the loop" is easy to say and easy to get wrong. Put a person on every step and the agent saves no time. Put a person nowhere and one bad input can move real money. This article sets out where the line should sit for common finance decisions. It builds on the governance section of our guide to AI agents for business finance, which explains mandates and ceilings in general terms.
Four levels of human involvement
It helps to stop thinking of human review as yes or no. Most well-run finance teams use four levels.
- Agent acts and logs
- The agent completes the task and records what it did. A person samples the log later.
- Agent prepares, one person approves
- The agent does the work and stops. A named person reviews and approves before anything happens.
- Agent prepares, two people approve
- As above, but two named people must approve, and they must be different people from whoever set the agent up.
- Person only
- The agent may gather information, but the decision and action belong to a person.
A decision map for finance
The table below is a starting point for a small or mid-sized business. Adjust the amounts to your size and risk appetite, and write the result into each agent's mandate.
| Decision | Suggested level | Why |
|---|---|---|
| Match a payment to an invoice | Agent acts and logs | Reversible, evidence-based, high volume |
| Send a payment reminder | Agent acts and logs | Low risk if tone rules and dispute stops are set |
| Pay an existing supplier, below ceiling | Agent acts and logs | Known payee, limited exposure |
| Pay an existing supplier, above ceiling | One approver | Larger exposure |
| Large or unusual payment | Two approvers | Hard to reverse once sent |
| Add or change a payee's bank details | Person only, confirmed through a separate channel | The main route for invoice-redirection fraud |
| Agree a write-off, settlement or payment plan | One approver, two above a limit | Changes money owed and the customer relationship |
| Post routine journals | Agent acts and logs, below ceiling | Reversible; sampled at close |
| Judgement journals (provisions, write-downs) | Person only | Accounting judgement and audit exposure |
| Submit a tax or regulatory return | Person only | A legal declaration by the business |
| Decline credit or service to an individual | Person only, or person review on request | Legal duties on automated decisions about people |
| File a suspicious activity report or freeze an account | Person only (a named officer) | Legal duties and harm from errors |
Ceilings: per transaction and per day
A per-transaction ceiling on its own is easy to get around, by accident or by design. An agent following a fake instruction to "pay these four invoices" could stay under the limit on each one. That is why ceilings should also apply per day and per payee. The OWASP security project calls the underlying problem "excessive agency": an agent holding more permission than its job needs, so that a manipulated or mistaken instruction can do real damage.
Worked example
An agent's mandate allows payments to existing suppliers of up to £5,000 each and £15,000 in total per day. It is asked to pay four invoices of £4,900 from the same supplier.
- Payment 1: £4,900. Under both limits. Paid and logged. Running total £4,900.
- Payment 2: £4,900. Running total £9,800. Paid and logged.
- Payment 3: £4,900. Running total £14,700. Paid and logged.
- Payment 4: £4,900. Running total would be £19,600, above the £15,000 daily ceiling. The agent stops and sends it for approval, with a note that four same-value invoices from one supplier arrived together.
A tighter rule would also flag the pattern itself: several identical amounts to one payee on one day is unusual and worth a person's attention before the first payment, not the fourth.
What the law already expects
Several regimes already require meaningful human involvement in some decisions.
- United Kingdom. The Data (Use and Access) Act 2025 relaxed the rules on solely automated decisions with significant effects on individuals, but kept safeguards: people must be told about such decisions, be able to challenge them and be able to obtain human intervention.
- European Union. Under the AI Act, organisations deploying high-risk systems, such as credit scoring of individuals, must assign human oversight to people with the competence, training and authority to exercise it. See what the EU AI Act means for finance teams.
- Payments fraud. Since 7 October 2024, UK payment firms have had to reimburse eligible victims of authorised push payment scams up to £85,000 per claim. That shifts cost onto firms and makes checks before money leaves more valuable, not less.
- Regulators' direction. The FCA's Mills Review, published in July 2026, expects AI to initiate and carry out financial decisions "within agreed parameters". The parameters, and who sets them, are the human part.
Making the review real
A person who approves forty agent requests a day without reading them is not a control. Signs of rubber-stamping include approvals in a few seconds, no rejections for months and approvers who cannot explain why they approved. To keep review meaningful:
- Show the approver the evidence and the agent's reasoning on one screen, not a bare amount.
- Highlight what is new or unusual: a first payment, a changed amount, a new country.
- Keep approval volumes low by setting ceilings sensibly; if a person approves everything, the ceiling is too low.
- Make rejection as easy as approval, and record the reason either way.
- Review the approval log itself each quarter. Our article on audit trails for AI agents lists what it should hold.
The same principle runs through every article in this series, from collections to compliance: the agent prepares, and a named person owns the decisions that matter. For the full framework, return to the parent guide on AI agents in finance.
Where Pryme fits
On Pryme business accounts, every agent acts inside a mandate with a ceiling, and above it a person decides. You can require a second named approver above set thresholds, and a transaction PIN is required. Sign-in uses a password and an authenticator app, with no SMS codes. See security and the Chief of Staff agent.
Outside Nigeria, Pryme's activities are carried out through regulated partners.
Questions people ask
What does human in the loop mean in finance?
It means a person reviews or approves an AI system's output before it takes effect, at points you define. In finance those points are usually payments above a limit, payee changes, filings and decisions about individuals.
Which payments should need two approvers?
Large payments, unusual ones and first payments to a new payee are common choices. The exact amount depends on your size, but it should be set per transaction and per day, and approvers should be different from whoever configured the agent.
Can an AI agent approve its own payments?
It should not. Preparing and approving the same payment is a breach of segregation of duties, whether the actor is a person or software.
How do I stop approvers rubber-stamping AI decisions?
Show evidence and reasoning with each request, flag what is new, keep volumes manageable with sensible ceilings and review approval patterns regularly. Fast approvals with no rejections are a warning sign.
Is fully automated decision-making legal in the UK?
For many business tasks, yes. For significant decisions about individuals, the UK GDPR as amended by the Data (Use and Access) Act 2025 requires safeguards, including the right to human intervention and to challenge the decision.
Sources
- GOV.UK: Data (Use and Access) Act 2025 - data protection and privacy changes · accessed Oct 2026
- European Commission AI Act Service Desk: Article 26, obligations of deployers of high-risk AI systems · accessed Oct 2026
- Payment Systems Regulator: PS24/7 APP scams reimbursement, confirming the maximum level · accessed Oct 2026
- FCA: AI and the future of retail financial services (The Mills Review) · accessed Oct 2026
- OWASP: LLM06:2025 Excessive Agency · accessed Oct 2026
This article is general information, not financial, legal or tax advice. Products described are provided under the terms on this site and are subject to regulation and eligibility.