← Go back

Guide // MandateGuard

What is an AI agent spending mandate?

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 spending mandate is a signed, scoped permission you give an AI agent before it spends: a per-transaction cap, a daily cap, a list of vendors it may use, and a threshold above which a human must approve. Every request is checked against the mandate before money moves, and every answer — ALLOW, DENY or REQUIRE_APPROVAL — is written to a signed, hash-chained receipt ledger. In MandateGuard the mandate is the policy you attach to an agent; the agent holds a key, never your card, and it cannot spend outside the mandate because the check happens in the engine, not in the prompt.

What is a mandate, in plain terms?

A mandate is to an agent what a purchase order limit is to an employee: a written, bounded authority to spend. Payment protocols such as AP2, UCP and x402 prove that a payment was authorised. A mandate answers the earlier question — whether this agent, on this task, should spend this money with this vendor again. MandateGuard issues, enforces and audits mandates, and stamps each decision into a tamper-evident receipt chain.

Why agents need mandates, not just payment methods

An AI agent with a stored card or an API billing key has capability without boundaries. It can retry a failed purchase ten times, subscribe to the premium tier of a tool it only needed once, or buy from a lookalike vendor with a similar name. None of those failures is a payments failure; the payment itself was valid. They are authority failures: nobody ever wrote down what the agent was allowed to do. A mandate writes it down before the first spend. It names the vendors the agent may use, the most it may spend in one transaction, the most it may spend in a day, and the amount at which a person takes over. Once that document exists, every later argument about an agent purchase stops being a matter of opinion and becomes a lookup against the mandate.

The four parts of every MandateGuard mandate

A MandateGuard policy has four working parts. The per-task cap is the largest single amount the agent may spend on one task reference. The per-day cap is the largest total it may spend across a day, which is what stops slow leaks made of many small purchases. The vendor allowlist names the sellers the agent may pay; anything outside it is denied with the reason recorded. The approve-over threshold routes larger in-policy spends to a human inbox instead of letting them through automatically. Together the four parts cover the realistic failure modes: one huge mistake, many small mistakes, the wrong counterparty, and the judgement call that deserves a second pair of eyes.

How a mandate is enforced before money moves

Enforcement is a pre-spend check, not an after-the-fact report. Before paying, the agent calls the authorize endpoint with its agent key, the vendor, the amount and a task reference. The engine compares the request with the mandate and answers in milliseconds: ALLOW when the spend is inside every bound, DENY with a reason when it breaks one, or REQUIRE_APPROVAL when it is inside the caps but above the human threshold. Approving releases a one-time spend token bound to that vendor and amount, redeemable exactly once. Because the agent never holds payment credentials it can use freely, there is no path around the check — the mandate is enforced by the system, not requested by the prompt.

What the receipt ledger adds

Every decision, including denials, appends a receipt to a per-workspace hash chain: each receipt commits to the hash of the one before it, and the chain can be recomputed at any time to prove nothing was edited. That turns the mandate from a settings page into evidence. When finance asks why the agent spent 340 dollars in March, the answer is not a chat log; it is a sequence of signed decisions, each showing the vendor, the amount, the task reference and the rule that allowed or stopped it. Denied attempts matter as much as allowed ones, because they show the mandate working — and they show you which vendor or task keeps pushing against the boundary.

Mandates versus prompts and system instructions

Teams often try to control spending by writing instructions: never spend more than 50 dollars, only use these vendors. Instructions are advice the model weighs against everything else in its context, and they fail silently exactly when the context gets long, the tool output is confusing, or a vendor page talks the agent into an upgrade. A mandate is the same intent expressed as data the engine enforces. The prompt can still guide judgement; the mandate guarantees the boundary. If the agent misunderstands the task, the worst case is a denied request and a receipt explaining why, instead of a charge you discover on a statement weeks later.

Who should hold the mandate

The mandate belongs to the person accountable for the budget, not to the agent and not to the vendor. In MandateGuard that is the workspace owner: they create the agent, set the policy, and receive the approvals. On the free plan that covers one agent and one policy with the full authorize flow and receipts, which is enough to protect a single purchasing or research agent. Pro and Team add multiple agents and policies, vendor trust built from your own ledger, duplicate-commitment detection, a kill-switch, evidence export and the interoperable intent service for mandates shared with merchants and issuer-side checkers.

Five actions to take now

  1. Name the agent and the one job it is allowed to spend on
  2. Set a per-task cap from the real price of that job, not a round guess
  3. Set a daily cap that assumes the agent retries and double-fires
  4. Allowlist only the vendors the job genuinely needs
  5. Set an approve-over threshold and check the approval inbox daily

Worked example: a research agent with a 200-dollar mandate

A small consultancy gives its research agent a mandate: per-task cap 40 dollars, daily cap 100 dollars, allowlist of two data providers and one model API, approve over 25 dollars. On Monday the agent buys four enrichment batches at 18 dollars each — all ALLOW, four receipts. It then tries a fifth batch from a new broker at 60 dollars. That request is DENY twice over: the vendor is not allowlisted and the amount breaks the per-task cap, and the receipt records both facts.

Later the agent requests 32 dollars from an allowlisted provider for a rush job. The amount is inside both caps, but it is above the 25-dollar approval threshold, so the engine answers REQUIRE_APPROVAL. The owner approves from the inbox, the agent receives a one-time token bound to that vendor and 32 dollars, redeems it once, and the day closes at 104 dollars of attempted spend, 72 dollars allowed automatically, 32 dollars allowed by a human, and one denial that cost nothing.

Frequently asked questions

What is the difference between a mandate and a budget?

A budget is a number you watch. A mandate is a number — plus vendor scope and approval rules — that a system enforces before each spend, with a signed receipt for every decision.

Can an agent exceed its mandate if it asks nicely?

No. The check is in the policy engine, not in the conversation. A request outside the mandate is denied or routed to a human regardless of how the agent phrases it.

What happens when a mandate denies a spend?

The agent receives DENY with a reason such as over the per-task cap or vendor not allowlisted, and the attempt is written to the receipt chain so the pattern is visible later.

Do mandates work with AP2, UCP or x402?

Yes — those protocols prove a payment was authorised; the mandate decides whether it should be authorised in the first place. MandateGuard also offers an AP2-style intent service for shared mandates.

How much does a mandate cost to run?

Free covers one agent and one policy with the full flow. Pro is 29 dollars a month and Team is 79 dollars a month, adding multiple agents, vendor trust, duplicate detection, the kill-switch and evidence export.

Keep reading

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