Human-in-the-loop approval for AI agent payments
Human-in-the-loop approval for AI agent payments: the agent prepares the batch, a person approves what's material and irreversible. Limits, tiers and why.
By Valeria VahorovskaUpdated 29 September 2026
Human-in-the-loop approval for AI agent payments means the agent does the reasoning and a named person approves the act that moves money. The agent reads invoices, matches them, and prices and routes the batch. The approval sits at settlement because agents are probabilistic and settlement is final. Spending limits and risk tiers decide which payments reach a person.
There’s a line in the IMF’s April 2026 note, How Agentic AI Will Reshape Payments, that should sit above every agentic finance roadmap: a tension between probabilistic AI behavior and the deterministic requirements of payment infrastructures. That sentence is the whole reason Fynex is built the way it is.
Why do AI agent payments need human approval?
Because the two systems have opposite tolerances for error. An agent is usually right; a payment is permanently right or wrong.
An AI agent is a probabilistic machine. Its outputs are likely correct, and getting likelier, but “likely” is the category it lives in. At some frequency it will produce an answer that is fluent, confident and wrong.
Settlement is the opposite kind of machine. There is no “probably paid”. Once money moves on an instant rail, instant means irrevocable. No amount of model progress closes that gap. A better agent is still a probabilistic system pointed at a deterministic one. The issue was never only accuracy. It is that the consequence structure is asymmetric: a bad draft costs a correction, a bad payment costs the money.
What the risks actually look like
The IMF note describes payments shifting from click-to-pay, where a person initiates each transaction, to decide-to-pay. There, an agent decides and execution runs at machine speed. Its risk list is worth reading in full.
- Herding. Agents running similar optimisation logic act at the same time. The IMF warns this can produce correlated payment flows that strain settlement capacity.
- An expanded attack surface. Agents touch many tools and APIs, which widens exposure to data exfiltration, tool misuse and privilege escalation. An agent that can move money is a far better target than one that can only read.
- Opacity. When a chain of agent decisions ends in a payment, can you reconstruct why?
The loss channel already exists without agents. The FBI’s 2025 Internet Crime Report puts business email compromise losses at just over $3 billion in 2025 ($3,046,598,558). The scam works by persuading businesses to pay the wrong account. Weak controls also kill projects. In June 2025, Gartner predicted over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls.
How to set AI agent spending limits and approval tiers
Set limits by consequence, not by the agent’s confidence score. The question for each payment is how bad it would be if this one were wrong, and whether you could get it back.
| Tier | Example | What the agent does | What the human does |
|---|---|---|---|
| Prepare | Match invoices, price and route the batch, forecast cash | Runs freely; everything is a draft | Nothing until release |
| Clear inside limits | Recurring payment to a known supplier, under the per-payment ceiling | Queues it inside the approved batch | Releases the batch in one action |
| Escalate | New payee, changed bank details, amount far above normal | Flags it with the reason and the evidence | Reviews that payment individually |
| Pause | Monthly cap hit, pattern looks like fraud | Stops and holds authority | Investigates before anything moves |
A worked example
Say you pay 60 supplier bills a month, $310,000 in total. You set a per-payment ceiling of $15,000 and flag any new or changed bank details.
The agent reconciles all 60 bills against invoices and POs. Fifty-six are recurring, correctly matched and under the ceiling; they go into one batch. Three are over $15,000 and one supplier has “updated” its bank account by email. Those four escalate with their evidence. You release the batch of 56 in seconds and spend your attention on the four that deserve it, including the one that looks like email compromise.
The approval line, defended
Fynex’s agents run the reasoning half of finance. They read invoices and match them, reconcile payments into Xero, forecast cash, catch the duplicate and build the payout batch priced and routed. All of that is reversible. The one act held behind approval is the one that crosses into settlement.
- Irreversibility. The approval sits at the only boundary where a mistake can’t be taken back.
- A circuit-breaker against herding. A person on settlement doesn’t fire in perfect sync with every other agent in the market.
- A smaller attack surface. Compromising an agent that can prepare but not release gets an attacker a proposal, not a transfer.
- Traceability by construction. Every payment reaches a person as a decision with its reasoning attached. The approval is the audit trail.
Human-in-the-loop is one control, not the whole answer
A person at the door is not a magic safeguard, and it isn’t free. The IMF says so too: human oversight alone falls short under time-compressed or clustered failures, and approval delays can add liquidity risk.
That is why approval is one layer in a stack. The IMF’s list pairs tiered human oversight with separation of decision-making from execution, deterministic policy controls, audit trails and graduated kill switches. Fynex is built on that stack: agents reason in a layer that cannot execute, deterministic rules gate every instruction, and approval is reserved for what’s material and irreversible. A human on everything would be theatre.
The industry already works this way. In their 2024 survey, the Bank of England and the FCA found 55% of AI use cases in UK financial services have some degree of automated decision-making. Of those, 24% are semi-autonomous, designed to involve human oversight for critical or ambiguous decisions. Only 2% of use cases are fully autonomous.
Doesn’t approval slow payments down?
It slows precisely one thing: the irreversible act. Everything upstream runs continuously, at machine speed.
What reaches you is a finished batch with the reasoning attached, not a blank payment form. On a prepared run that’s seconds, not hours. Compare full autonomy through settlement, where the first time anyone sees a bad payment is after it has gone. If you’re evaluating approval workflows in a spend tool, our Ramp procure-to-pay evaluation covers where its approvals sit, and what an AI reconciliation agent does covers the preparation half.
Related: intent, authorization, settlement is the three-layer model this argument sits inside; know your agent covers what to record when an agent acts; agent-to-agent payments is the version with no human anywhere.
The category will spend the next few years deciding where the human sits relative to the money. We placed the human on the settlement boundary before the IMF note came out, and the Fynex product is built around it. Agentic everywhere it helps; deterministic, approved and traceable at the one place money becomes final.