Guide // MandateGuard
How do you require human approval before an AI agent pays?
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 require human approval before an agent pays by setting the approve-over threshold in its MandateGuard policy: any request above that amount returns REQUIRE_APPROVAL, waits in the approval inbox, and can only proceed when a person approves it — which releases a one-time spend token bound to that vendor and amount. The requirement lives in the policy engine, not the agent's prompt, so the agent cannot argue, forget or route around it. Build the agent to wait gracefully on that state, keep the approval queue reviewed daily, and remember approval is per purchase: the next purchase asks again.
What does requiring approval actually mean technically?
Requiring approval means configuring the policy so that spends over a threshold cannot complete on the agent's authority alone. In MandateGuard the mechanism is the approve-over threshold plus the approval inbox: requests above the threshold return REQUIRE_APPROVAL and wait; an approver's decision either releases a one-time spend token bound to vendor and amount, or closes the request. The requirement is enforced by the engine at authorize time, not by an instruction in the agent's prompt.
Approval must be a mechanism, not an instruction
The weak version of approval lives in a prompt: ask me before spending over 50 dollars. It fails the ways prompts fail — context pressure, ambiguous amounts mid-task, a vendor page insisting the step is a free trial. The strong version lives in the engine: the agent literally cannot complete an over-threshold purchase because the authorize call returns REQUIRE_APPROVAL and no token exists until a human creates one. When you evaluate any agent platform's approval story, ask where the refusal lives. If it lives in the model's instructions, it is a preference. If it lives in a policy engine the agent cannot edit, it is a requirement. MandateGuard implements the second kind.
Configure the threshold as the requirement
Setting approve-over is the whole configuration, and its placement defines what the requirement means in practice. Set it at zero-ish values and you have chosen manual purchasing with agent assistance — fine for probation. Set it just under your per-task cap and you have chosen broad autonomy with a supervised top band. The productive middle for most teams is a threshold at half to two-thirds of the cap, tuned after a week of receipts. Remember the threshold interacts with the allowlist: approval should never be the control that admits a new vendor, because approvers deciding quickly may not notice the counterparty changed. New vendors enter through the allowlist deliberately; amounts enter through approval routinely.
Build the waiting behaviour into the agent
Requiring approval changes the agent's programming contract. On REQUIRE_APPROVAL the agent must stop that purchase line, record that it is waiting, and either continue independent work or end its turn gracefully with a summary of what is pending. It must not poll in a tight loop, invent a smaller substitute purchase to stay busy, or re-present the same request hoping for a different answer. When the human approves, the agent redeems the one-time token to complete that exact purchase; when denied, it reports the denial and moves on. These behaviours are simple, but they must be designed in — an agent built only for the success path will treat a pending approval as a malfunction.
Run approvals as a daily operational queue
The requirement is only as strong as the queue discipline behind it. Route the inbox into a real routine: a named approver checks it on a cadence matching the agent's work rhythm, with a named backup for absences. Webhook notifications can push pending items into the channels your team already watches. Decide from the card — vendor, amount, task reference — and use receipt history when the card looks unusual. Keep decisions timely: a queue that waits three days teaches the organisation to route around the agent, recreating uncontrolled purchasing in the name of speed. A small, current queue is both safer and faster than a bypassed one.
Why the token binding completes the requirement
Approval systems are often undermined at the last step: the human approves a description, and the agent executes something adjacent. MandateGuard binds the release to the facts approved. The spend token is minted on approval, stored as a hash, bound to the approved vendor and amount, and redeemable exactly once through the redeem endpoint. The agent cannot apply Tuesday's hotel approval to Wednesday's upgrade, cannot inflate the amount, and cannot save the token for a different vendor. Redemption closes the receipt that the original request opened, so the chain shows request, decision and execution as one story. That binding is what lets approvers approve quickly: the thing they approved is the thing that can happen.
Extending requirements to projects and partners
Standing thresholds cover standing agents. Two MandateGuard shapes extend the same idea. Spend envelopes let you write approval conditions onto a single task's mandate — single spends over a named amount need approval, with the reviewer card showing the request against the envelope's vendors, period and conditions. Intents carry the requirement across organisations: an issuer registers the mandate, merchants verify receipts against it, and suspend or revoke changes the authority state for every checker at once. The pattern is constant — a human decision, bound to facts, recorded in a chain — applied at the scope the work actually has.
Five actions to take now
- Set approve-over deliberately — the threshold is the requirement's strength
- Build the agent to wait on REQUIRE_APPROVAL without retrying
- Put the inbox in someone's daily routine and name a backup approver
- Use the one-time token as issued; never ask the agent to reuse approvals
- For defined projects, write approval conditions into a spend envelope instead
Worked example: approval is per purchase, not per vendor
A founder sets approve-over at 30 dollars for her booking agent, whose per-task cap is 120 dollars and daily cap 250 dollars. On Tuesday the agent assembles a trip: a 24-dollar train ticket is ALLOW and completes. A 96-dollar hotel night returns REQUIRE_APPROVAL. The founder sees the card — vendor, amount, task reference naming the trip — checks the dates match the conference, and approves. The token released is bound to that hotel vendor and 96 dollars.
On Wednesday the agent tries to add a 54-dollar room upgrade from the same vendor. It does not inherit Tuesday's approval: it is a new request, again above the threshold, and it waits. The founder denies it from the inbox. The trip costs what was approved, the upgrade costs nothing, and the receipts show both decisions with their reasons — the difference between an approval system and an approval memory.
Frequently asked questions
Can I require approval for every purchase?
Set approve-over very low and effectively everything pauses for a human. That is right for a new agent's probationary period; most teams relax it for routine small buys once receipts build trust.
What if the approver is unavailable?
The spend waits. Design agent jobs so waiting is safe; for genuinely urgent categories, pre-agree a second approver rather than weakening the threshold.
Does approval work through webhooks?
MandateGuard can notify your webhook when approvals are pending and decided, so the requirement can plug into your existing ops channels.
Can an agent split a purchase to dodge approval?
Splitting creates multiple authorize calls under the daily cap, and the pattern is obvious in receipts. It also tends to trip duplicate detection on paid plans. Treat a split as a policy conversation.
How is this different from an envelope's approval conditions?
The threshold is amount-based and standing. An envelope states conditions for one task — named amount levels plus written conditions on the mandate itself — and its reviewer card shows the request against them.
Keep reading
- When should a human approve an AI agent's spending?
- How does a one-time spend token work after approval?
- How do you write approval conditions on a spending mandate?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides