Guide // MandateGuard
How do you set spending limits for an AI agent?
By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14
▸ Run the free live demo See pricing
Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.
Short answer
You set spending limits for an AI agent by writing them into a policy the agent cannot edit: a per-transaction cap sized to the real price of the job, a daily cap that assumes retries and double-fires, a vendor allowlist limited to sellers the job needs, and an approval threshold for amounts a human should see. In MandateGuard those four settings form the agent policy, and every spend request is checked against them before money moves. Start tight, read the denial receipts for a week, then loosen only the bound that is blocking real work — the receipts tell you which one that is.
What is an agent spending limit?
An agent spending limit is a rule enforced by software, not a hope expressed in a prompt. In MandateGuard a limit set has four parts: per-task cap, per-day cap, vendor allowlist and approve-over threshold. A request that fits all four is allowed; one that breaks a cap or the allowlist is denied with a reason; one that fits the caps but passes the threshold waits for a human. Each outcome is stamped into a hash-chained receipt.
Start from the price of the job, not a round number
The most common limit-setting mistake is picking a comfortable round number. Fifty dollars feels safe until the data provider the agent needs charges 62 dollars for the standard batch, and then every legitimate request needs an exception. Price the actual job first: list the purchases the agent will make in a normal week, with real prices from the vendors it will use. Set the per-task cap a little above the largest legitimate single purchase — enough headroom for taxes or a quantity step, not enough to buy the premium tier by accident. If two very different jobs share one agent, that is a signal to split the agent or use a task-specific spend envelope instead of inflating the cap.
Size the daily cap for the agent's worst honest day
Daily caps fail in the other direction: sized for an average day, they block the busy day; sized for the busiest day, they permit a slow leak. Work from the agent's failure modes. Agents retry. Agents loop. An agent that normally buys three 18-dollar batches might, on a confused day, attempt nine. A daily cap of roughly two to three times the normal daily spend catches runaway behaviour while leaving room for a genuinely heavy day. MandateGuard counts allowed spend against the day cap as decisions are made, so the cap is a live boundary, not an end-of-day surprise, and denied attempts still leave receipts that show you the pressure building.
Treat the allowlist as a limit, not paperwork
Caps limit how much; the allowlist limits who. An agent that may only pay the three vendors its job needs cannot be talked into buying from a lookalike domain, a sponsored result, or a marketplace seller with a similar name. Keep the list short and exact, and review it when the job changes rather than when an incident forces you to. When a request arrives for a vendor outside the list, MandateGuard denies it and records the attempt. A cluster of those denials is useful signal: either the job genuinely grew a new dependency, which deserves a deliberate allowlist change, or something is steering the agent, which deserves an investigation.
Place the approval threshold where judgement lives
Some spends are legal under the caps but still deserve a human: the first purchase from a newly allowlisted vendor, an amount near the top of the range, anything unusual for the task. The approve-over threshold catches those. Set it below the per-task cap — commonly at half to two-thirds of it — so routine small purchases flow and larger ones pause. Approvals land in the MandateGuard inbox and on your webhook if you configure one. Approving releases a one-time spend token bound to that vendor and amount, so approval cannot be reused for a different or larger purchase later.
Read the receipts and tune one bound at a time
Limits are a hypothesis; receipts are the experiment. After the first week, read the chain: which denials were the mandate saving you, and which were the mandate blocking real work? If the per-task cap blocked three legitimate purchases from an allowlisted vendor, raise that cap slightly or route that job through an envelope with its own maximum. If the daily cap never came close, consider lowering it — unused headroom is unpriced risk. Change one bound at a time and note the change, so next month's receipts can be read against a known policy. MandateGuard's verify-chain endpoint lets you prove the receipt history you tuned from is the history that actually happened.
Know when one policy is no longer enough
A single agent with a single policy is the free starting point and the right shape for one job. You have outgrown it when different jobs need different vendors, when one team's agent should not share another team's daily cap, or when you need duplicate-commitment detection across purchases. Pro and Team raise the agent and policy limits and add vendor trust from your own ledger, the kill-switch, evidence export and MCP tools. The limit-setting method does not change at that point — price the job, cap the day, allowlist the vendors, threshold the judgement calls — you simply get to apply it per agent instead of globally.
Five actions to take now
- List the agent's real weekly purchases with real vendor prices
- Set the per-task cap just above the largest legitimate purchase
- Set the daily cap at two to three times a normal day's spend
- Allowlist only the vendors this job needs, exactly named
- Set approve-over below the per-task cap and review denials weekly
Worked example: sizing limits for a cloud provisioning agent
An engineering team lets an agent spin up test infrastructure. Real prices: a standard test stack costs 14 dollars a run, the largest legitimate run is 38 dollars, and a normal day is four runs. They set a per-task cap of 45 dollars, a daily cap of 150 dollars, an allowlist containing only their cloud provider, and approve-over at 30 dollars. Most runs pass automatically; the 38-dollar runs pause for a human, which the team accepts because those are the runs worth a glance.
In week two the receipts show the agent attempted the same 14-dollar stack six times in eleven minutes after a tool error — the first four were allowed, the daily cap stopped the pattern at 150 dollars of attempted run value, and two denials were recorded with the cap reason. The team fixed the retry logic in the agent instead of raising the cap, which is exactly the decision the limit was designed to force.
Frequently asked questions
Should the daily cap be per agent or per team?
Per agent in MandateGuard, because each policy belongs to an agent. Teams that need a shared ceiling give sibling agents coordinated daily caps and review the combined receipts.
What if a legitimate purchase exceeds the per-task cap?
Split the purchase if the vendor allows it, route that job through a spend envelope with its own maximum, or raise the cap deliberately — the denial receipt tells you which case you are in.
How tight is too tight?
When denial receipts show legitimate, allowlisted work blocked several times a week, one bound is wrong. Loosen that bound only, by the smallest step that clears the real purchases.
Do limits apply to subscriptions the agent started earlier?
Every new charge the agent initiates is a new authorize call, so renewals the agent triggers are checked against the current mandate like any other spend.
Can I change limits without redeploying the agent?
Yes. The policy lives in MandateGuard, not in the agent. Editing caps, allowlists or thresholds takes effect on the next authorize call.
Keep reading
- What is an AI agent spending mandate?
- What is a per-transaction spending cap for an AI agent?
- How do daily spending caps stop AI agents leaking money?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides