← Go back

Guide // MandateGuard

When should a human approve an AI agent's spending?

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 human should approve an agent's spending when the purchase is inside its mandate but large enough, new enough or unusual enough to deserve judgement. In MandateGuard that line is the approve-over threshold: requests above it return REQUIRE_APPROVAL instead of ALLOW, wait in the approval inbox, and — if approved — release a one-time spend token bound to that vendor and amount. Place the threshold below the per-task cap so the top of the allowed range is supervised, keep the queue small enough to review daily, and deny from the inbox when the answer is no: a decision either way is recorded in the receipt chain.

What is an approval threshold?

The approve-over threshold is a policy amount. Requests at or below it that pass every other check are allowed automatically. Requests above it pause for a human decision in the MandateGuard approval inbox (and on your webhook, if configured). Approval does not hand the agent general permission: it mints a single-use spend token tied to the approved vendor and amount, which the agent redeems exactly once.

The three answers and what each one protects

MandateGuard's engine returns ALLOW, DENY or REQUIRE_APPROVAL, and the middle answer is the one that makes the system usable. Pure automation with only allow and deny forces a bad choice: set caps high enough for real work and the agent can make large unsupervised purchases, or set them low and legitimate work stalls. The approval state resolves the tension by separating legality from judgement. Caps define what is legal. The threshold defines what is legal but worth a look. That separation is why thresholds should be tuned to human attention, not just to money: the right threshold is the one where an approver, seeing the vendor, amount and task reference on the card, can make a real decision in seconds.

Placing the threshold

A workable starting point is half to two-thirds of the per-task cap. Below the threshold, purchases are routine: known vendors, ordinary amounts, coherent tasks. Above it sit the purchases with consequence — the top of the price range, the quantity step, the first big order of the month. Some teams also route by novelty rather than size alone: the first purchase from a newly allowlisted vendor deserves review even if it is small. MandateGuard's threshold is amount-based, so achieve novelty review operationally — when you add a vendor, temporarily lower the threshold or watch its first receipts closely — and use spend envelopes when a specific task needs named approval conditions written onto the mandate itself.

What the approver actually sees

Judgement needs context, and the approval flow is built around the card: the vendor, the amount, and the task reference the agent supplied. That trio answers the approver's real questions — who is being paid, how much, and for what claimed purpose — and the receipt history behind the agent answers the fourth: what has this agent done so far today and this month. Approvals can be actioned from the web inbox, and a webhook can notify your own systems when an approval is pending or decided. What the approver never has to do is share credentials or raise the agent's limits: approving one purchase changes nothing about the policy.

The one-time token is the security detail that matters

A naive approval system flips a boolean and lets the agent spend. MandateGuard instead issues a spend token on approval: stored only as a hash, bound to the approved vendor and amount, redeemable exactly once via the redeem endpoint. Binding prevents the approved 40-dollar purchase from becoming a 400-dollar one. Single-use prevents the approval from becoming a standing line of credit the agent can draw on later for a different purpose. Hash storage means the token database is not itself a pile of live credentials. When the agent redeems, the spend completes the receipt that the REQUIRE_APPROVAL decision began, so the chain shows the full story: requested, approved, redeemed.

Running the queue like an operator

Approval systems fail socially before they fail technically: the queue grows, reviews get rushed, and approve becomes a reflex. Keep the queue honest by keeping it small — if more than a handful of items wait each day, either the threshold is too low for the job's real price shape or the agent is generating noise. Review denials too: a denied approval is also a decision, recorded, and it teaches the agent's operators where the mandate and the job disagree. Finally, treat approval latency as part of the design. Agents that must wait should be built to continue other work or stop gracefully; a mandate that pauses spending should never corrupt the underlying task.

Approvals inside envelopes and intents

The same idea appears in MandateGuard's other mandate shapes. A spend envelope can state that single spends over a named amount need approval, and its reviewer card shows the request against the envelope's vendor scope, period and conditions before deciding. Intents carry budgets and authority windows that issuer-side operators manage with suspend, resume and revoke. The inbox threshold is the standing-policy version: simple, amount-based and always on. Choosing between them is a question of scope — a standing agent gets a threshold; a defined project gets an envelope; a mandate shared with outside parties gets an intent.

Five actions to take now

  1. Set approve-over at half to two-thirds of the per-task cap
  2. Review the approval inbox daily; a stale queue teaches rubber-stamping
  3. Decide from the card: vendor, amount, task reference, recent receipts
  4. Deny in the inbox when the answer is no — never raise limits to clear a queue
  5. If approvals exceed a handful a day, retune the threshold to the job's real prices

Worked example: one morning in the approval inbox

A support-operations agent has a per-task cap of 100 dollars, a daily cap of 300 dollars and approve-over at 50 dollars. At 09:12 it requests 64 dollars from its allowlisted helpdesk vendor for an extra agent seat — inside the caps, above the threshold. The request waits as REQUIRE_APPROVAL and the owner gets a webhook notification. The card shows the vendor, the amount and the task reference for the support expansion, and the owner approves. The agent receives a one-time token for that vendor and 64 dollars, and redeems it once.

At 09:40 the same agent requests 88 dollars from the same vendor for an analytics add-on with a vague task reference. The owner denies it from the inbox. No token exists, the agent cannot proceed, and both decisions sit in the receipt chain with their outcomes — which is what lets the owner, at month end, show exactly which agent purchases a human personally authorised.

Frequently asked questions

How fast must approvals happen?

At human speed, by design. Build agents to park the purchase and continue other work. If a job truly cannot wait, its routine amounts belong below the threshold.

Can approval raise the amount?

No. The token is bound to the approved vendor and amount. A larger purchase is a new request and a new decision.

Who can approve?

Whoever holds the workspace credentials and works the inbox. MandateGuard does not delegate approval to the agent or the vendor.

What if nobody approves?

The spend does not happen and the request remains pending. Silence is a no, which is the safe default for money.

Are approval decisions auditable?

Yes. Requests, approvals, denials and redemptions are all part of the hash-chained receipt history, exportable as evidence on paid plans.

Keep reading

Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides