Set the spending limit at the payment layer, not in the prompt, and give each agent a narrow mandate: what it may buy, how much, and for how long. One review of regulator, bank and consultancy documents found they converge on that mandate, a check at the moment of payment, escalation to a human above a threshold, and clear accountability. Then decide what happens when nobody answers: the same review says escalation should default to no.
Quick version
- Cap spending where money moves, not in the instructions the model reads.
- Scope each credential to a merchant, an amount and a time window.
- Set approval thresholds so large or unusual actions go to a person.
- Name one human owner for each agent.
- Log every action so it can be searched later.
- Start with small limits and raise them as the agent earns trust.
Several sources below are vendors. Ramp, Tipalti, Eco, PinkWallet's author and Zapx Labs all sell or build products in this area, so their advice leans toward their own features. The overlap between them is the useful part.
Why a cap matters before the first mistake
An agent does not get tired and does not get suspicious. If it loops or misreads a task, it repeats the error at machine speed. Ramp says an agent stuck in a loop can burn through $50,000 before anyone notices. Ramp sells agent spending controls, so weigh the framing, but the mechanism is easy to see.
Incidents are not rare. A 2026 Cloud Security Alliance survey, as quoted by Ramp, found 65% of enterprises running AI agents had at least one agent-related incident in the past 12 months. Of those, 35% reported direct financial losses.
The regulators' concern is similar. A European policy group (CERRE), as summarized in the DEV Community review, lists exceeding a spending limit through malfunction, hallucination or prompt injection as a concrete consumer risk. The author of that review works on a spend-control product, so read it as a good map of the documents rather than a neutral verdict.
Two kinds of agent spend
Finance teams usually have to cap two different things.
- Transactional spend. The agent buys something: software, supplies, a vendor invoice.
- Operating cost. The agent itself costs money through model usage.
Ramp's guide covers both. Ramp sells AI agent spending controls and virtual cards for agents, so read its advice with that in mind. On the operating side, it says a general-purpose agent using your most capable models can run $100 to $300 per day in token usage alone, while a scoped agent on a smaller model may cost under $10 per day. It also reports that average monthly token spend by Ramp customers has grown 13x since January 2025.
One practical lever follows from that. Ramp notes that a customer FAQ agent can run on a smaller, cheaper model, while a complex reasoning task may justify paying more. Matching the model to the stakes is a cost control, not just a quality choice.
The controls finance teams actually use
No single control covers everything. The table shows what each one does and where it comes from.
| Control | What it limits | Source |
|---|---|---|
| Single-use virtual card | One merchant and one dollar amount, enforced at the point of transaction | Ramp |
| Per-authorization or per-period limit with real-time approval | Each charge or each period, with a webhook that can approve or deny | Zapx Labs summary of Stripe Issuing |
| Per-action and aggregate value thresholds | Below the threshold the agent acts. Above it a human reviews or the action is blocked | MAS guidance, as summarized in DEV Community |
| Prefunded wallet with a capped total budget | The most the agent can ever spend | Citi Institute GPS, as summarized in DEV Community |
| Searchable audit trail | Nothing directly, but every action can be checked afterward | Tipalti |
Scope the credential, not just the dollar amount
A dollar cap alone is a blunt tool. Eco's help center says a working stack has three properties. Specificity means a credential carries the exact merchant, amount and time it authorizes, not blanket account access. Verifiability means the issuer and the merchant can independently confirm authorization at transaction time. Revocability means a user, a treasury admin or an automated risk rule can pull the agent's authority back without breaking the rest of the payment surface.
That is a good test for any setup you are reviewing. Can you answer all three? If an agent's credential is a shared company card, the answer to specificity and revocability is probably no.
Here is why the destination matters. Zapx Labs describes its own tool, Veyra, as using a capped allowance to a relayer. The cap limits how much can move but not where it goes. That is an honest description of a common limit: a cap controls the size of a loss, not its direction. Pair amount caps with merchant or category limits.
Keep the limit outside the model
It is tempting to write the budget into the agent's instructions. That is the weakest place for it. The DEV Community review puts it plainly: a budget in the system prompt is text the model can misread or be talked out of, while a mandate checked by the payment layer is not.
Our view: treat anything the model can read as advice, and anything the payment system enforces as a rule. Only rules belong in a finance workflow. The same logic is behind our piece on AI agent spending limits.
Set approval thresholds, and expect them to be needed
Tipalti tells finance leaders to define thresholds and policy triggers that force mandatory human review. It also suggests asking any vendor whether users can set custom approval thresholds for AI-driven actions. Tipalti sells finance automation, so it has an interest in that feature, but the question is a good one for any tool.
Thresholds beat reviewing everything. The Bank of England, as summarized in DEV Community, warns that keeping a human in the loop for every agent action is unlikely to be realistic. If every action needs a click, people approve without reading. Zapx Labs reports that most teams combine a hard cap with a human approval step above a threshold.
Three outcomes cover it: allow, escalate to a person, or block. For treasury work, J.P. Morgan, as summarized in the same review, expects agents to propose actions while humans keep the final say.
A made-up example of the logic
Here is a made-up example, not real data. An agent that renews small software subscriptions is allowed to pay a known vendor up to a set amount. Anything above that goes to its owner. A new vendor goes to its owner no matter the amount. If the owner does not answer, the payment does not go out. The numbers are yours to choose. The structure is what matters.
Name an owner and keep a trail
Ramp says every agent should have a named human owner who is accountable for its spending and output. Without one, a limit breach becomes everyone's problem and therefore no one's.
Tipalti adds that every action an AI agent takes should generate a searchable audit trail. It also says governance should be built into each step of the financial lifecycle, not applied as a final check before releasing payments.
There is a legal wrinkle worth knowing about. The Consumer Bankers Association, as summarized in DEV Community, flags that existing protection for unauthorized transactions may not fit agent payments cleanly. Check with your card issuer, bank or a qualified professional before assuming a disputed agent charge will be covered.
Watch for limits that drift
A cap set on paper can differ from what actually settles. Zapx Labs notes that when an agent uses compute providers or data APIs, the initial authorization hold often differs from final settlement once dynamic usage overages or sales tax clear two days later. Leave headroom in your reconciliation, and review settled amounts, not just approvals.
Expand autonomy slowly
Tipalti recommends that finance leaders gradually expand AI autonomy as trust is proven, keeping human oversight over high-risk workflows. That matches common sense. A new agent has no record. Give it a small mandate and tight thresholds, then look at the logs before loosening anything.
Our view: raise one limit at a time, and only after reading the audit trail. If you cannot explain what the agent did last month, it has not earned a bigger budget. Some tasks should stay with a person no matter what, as covered in what to never automate with an AI agent.
A checklist to run before launch
- Does the agent have its own credential, not a shared card?
- Is the credential scoped to a merchant or category, an amount and a time window?
- Is the limit enforced by the payment system, not the prompt?
- Can someone revoke the credential quickly without breaking other payments?
- Is there an approval threshold, and what happens if nobody responds?
- Is there one named human owner?
- Does every action land in a searchable log?
- Is there a cap on model usage costs, not only purchases?
- Is someone reviewing settled amounts against approved ones?
Where to go from here
The controls above are not complicated. They only work if you decide them before the agent is live. If you want a short version to keep next to you, Get Sunday Night Truths. The free newsletter sends The Agent Leash, a one-page checklist of 15 checks before an AI agent touches your money, the moment you sign up.
This article is general information, not financial or legal advice. Talk to a qualified professional about your own setup.
