Guide // MandateGuard
What is a per-transaction spending cap 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
A per-transaction cap — in MandateGuard, the per-task cap — is the largest single amount an agent may spend in one authorize call. Any request above it is denied outright, with the reason written to the receipt chain, no matter how much daily budget remains. It is the control that stops the single catastrophic purchase: the premium tier, the wrong quantity, the annual plan bought instead of the monthly one. Size it from the largest legitimate single purchase the job requires, keep it paired with a daily cap for slow leaks, and let the approval threshold — not a bigger cap — handle the legitimate exceptions.
What is a per-task cap?
MandateGuard's per-task cap is the maximum USD amount allowed in one authorization for one task reference. The engine compares the requested amount with the cap before anything else can spend. Above the cap: DENY, reason recorded. Below it, the request still faces the daily cap, the vendor allowlist and the approval threshold. The cap is a property of the policy, editable by the workspace owner at any time, effective on the next call.
The failure this cap exists to prevent
Most agent spending disasters are single events, not patterns. An agent asked to buy a dataset buys the enterprise tier. An agent told to book the cheapest suitable option books the refundable premium fare. An agent ordering ten units orders ten cases. In each case the daily budget might easily cover the amount, and the vendor might be perfectly legitimate — the error is the size of one decision. A per-transaction cap is aimed precisely at that shape of error. It converts the worst case from an unbounded charge into a denied request and a receipt, and it does so at the only moment that matters: before the money moves.
How the check runs inside an authorize call
When the agent calls the authorize endpoint, it presents its agent key, the vendor, the amount and a task reference. The engine loads the agent policy and evaluates the request against each bound. If the amount exceeds the per-task cap, the answer is DENY with the cap as the reason, and the receipt records the attempted amount — useful evidence when you later ask whether the cap is set correctly. If the amount fits, evaluation continues: is the vendor allowlisted, does the day still have room, does this amount need a human? The ordering matters because it keeps reasons honest: a denial names the first bound the request actually broke.
Sizing the cap from real prices
List the single purchases the job legitimately makes and take the largest. Add modest headroom for price steps, taxes or currency rounding — then stop. Every extra dollar of headroom is a dollar of single-decision risk you chose to keep. Where one job has a bimodal price shape — many 12-dollar purchases and one legitimate 300-dollar purchase each quarter — a high cap is the wrong tool. Keep the cap at the routine level and give the quarterly purchase its own route: a human approval if it fits under a higher threshold arrangement, or a task-specific spend envelope with its own maximum total. The cap should describe the normal shape of the job, not its rarest exception.
Per-task cap versus approval threshold
These two settings are often confused, and the difference is the difference between never and ask me. Above the per-task cap, the answer is always no; the agent should redesign the purchase, split it, or stop. Above the approval threshold but below the cap, the answer is ask a human: the request waits in the inbox, and approval releases a one-time token for exactly that vendor and amount. A healthy policy uses both: the threshold at perhaps half to two-thirds of the cap, so the top of the allowed range is supervised. If you find approvers rubber-stamping everything in that band, the threshold is theatre — either lower the cap or accept the band and raise the threshold to where judgement is real.
What the receipts teach you about your cap
Denied-over-cap receipts cluster in instructive ways. The same vendor and a similar amount repeatedly denied usually means the job outgrew the cap or a vendor raised prices — a deliberate cap review is due. Wildly varying amounts from an unlisted context usually mean the agent is improvising purchases outside its job description — the cap just saved you, and the agent prompt or toolset needs the fix. A single huge attempt after a vendor page interaction can indicate the page steered the agent toward a premium option. MandateGuard records attempted amounts on denials precisely so these patterns are visible instead of anecdotal.
Caps on the free and paid plans
The free plan includes one agent and one policy with the full authorize flow, so a per-task cap protecting a single agent costs nothing to run. Pro and Team let you hold multiple agents and policies, each with its own cap tuned to its own job — the purchasing agent's cap need not resemble the research agent's. Paid plans also add duplicate-commitment detection, which complements the cap: the cap stops one oversized purchase, while duplicate detection catches the same legitimate-looking purchase made twice inside your window.
Five actions to take now
- Find the largest legitimate single purchase and set the cap just above it
- Pair the cap with a daily cap — one without the other leaves a shape of loss open
- Set the approval threshold inside the cap so the top band is supervised
- Route rare oversized jobs through an envelope instead of raising the cap
- Review over-cap denial receipts monthly for price changes and steering
Worked example: one cap, three requests
A marketing agent has a per-task cap of 75 dollars, a daily cap of 200 dollars, an allowlist of two stock-media vendors, and approve-over at 40 dollars. Its first request, 22 dollars for a photo pack, is ALLOW. Its second, 58 dollars for a video bundle, is inside the cap but above the threshold, so it waits and is approved by the owner, who can see the vendor and amount on the approval card.
Its third request is 240 dollars for an annual creator-suite plan from an allowlisted vendor. The vendor is fine and the daily cap has room, but 240 dollars breaks the per-task cap, so the answer is DENY with the cap named as the reason. The agent falls back to the monthly plan at 24 dollars. Without the cap, every individual check would have passed and the annual commitment would have been bought by default.
Frequently asked questions
Is a per-task cap the same as a per-transaction cap?
Yes, in MandateGuard's terms. The cap is evaluated per authorize call, and the task reference ties the spend to the job it belongs to.
Can a vendor split a charge to slip under the cap?
Splitting produces several authorize calls, which then face the daily cap and, on paid plans, duplicate-commitment detection. The receipt chain makes the pattern plain.
What reason does the agent see when it hits the cap?
DENY with the cap identified as the reason, plus a signed receipt recording the attempted amount, vendor and task reference.
Should different agents share one cap value?
Each policy carries its own cap. Tune per job: a provisioning agent and a media-buying agent have different legitimate price shapes.
Does the cap apply inside spend envelopes and intents?
Envelopes carry their own maximum total and per-spend approval level; intents carry cumulative and per-period budgets. The policy cap still governs plain authorize calls.
Keep reading
- How do daily spending caps stop AI agents leaking money?
- How do you set spending limits for an AI agent?
- When should a human approve an AI agent's spending?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides