← Go back

Guide // MandateGuard

How do you give an AI agent a budget safely?

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

Give an AI agent a budget by converting the amount into a mandate, not a balance the agent holds: a per-task cap from the job's largest real purchase, a daily cap that makes the monthly total structural, an allowlist of the vendors the budget is actually for, an approval threshold where you want to see purchases yourself, receipts for every decision and a freeze for the bad day. Close or gate every other path by which the agent could pay, brief the agent to branch on ALLOW, DENY and REQUIRE_APPROVAL, staff the inbox before go-live, and grow the budget only on receipt evidence, one bound at a time.

What is the safe way to hand an agent money?

Giving an agent a budget safely means converting a sum of money into a mandate the agent operates inside: caps for size and pace, an allowlist for counterparties, an approval threshold for judgement calls, receipts for memory, and a freeze for emergencies. In MandateGuard the budget lives as the agent's policy (plus envelopes for defined jobs and intents for shared arrangements), enforced at authorize time — the agent proposes each spend and the mandate, not the agent's optimism, decides.

A budget is a mandate, not a balance

The unsafe mental model treats a budget as a pile of money the agent holds: a card, a balance, a wallet. Piles get spent at the holder's discretion, and the owner's only instruments are the pile's size and the post-hoc statement. The safe model treats a budget as a set of decisions made in advance: how large any single spend may be, how much may accumulate per day, which counterparties may be paid, which sizes need a second mind, and what evidence every decision leaves. The agent never holds the budget; it proposes against it. That inversion — from holding to proposing — is the entire safety story, and everything else in this guide is the inversion's implementation detail.

Translate the amount into the four settings

Take your monthly figure and decompose it. Daily cap: the monthly budget divided by working days, adjusted to the busiest honest day — this is the setting that makes the monthly total structurally likely rather than aspirationally hoped. Per-task cap: the largest single purchase the job legitimately makes, from real prices. Allowlist: the vendors the budget is for, by charging identity — a budget for four tools is not a budget for the internet. Approval threshold: the size at which you want to see the purchase yourself, set below the cap. MandateGuard's free plan holds this whole translation for one agent. Write the four numbers and their reasons in your policy register; the reasons are what the first monthly review will tune from.

Close the other doors

A mandate governs the paths that pass through it. Inventory the agent's other routes to money — stored cards in browsers, marketplace accounts, cloud consoles, a second tool with its own balance — and either gate them behind the same authorize discipline or remove the payment capability. This step is unglamorous and decisive: most real-world agent overspend does not defeat the control, it walks around it, through a path nobody remembered granting. Where a path must remain (a rail the agent executes allowed spends on), the discipline is that execution follows an ALLOW or a redeemed token — never precedes it. Defence in depth is welcome (a low card limit as backstop), provided nobody mistakes the backstop for the governor.

Brief the agent, staff the human side

The agent should know its budget exists and how to behave inside it: propose through authorize, proceed on ALLOW, stop and report on DENY without retrying or substituting, wait gracefully on REQUIRE_APPROVAL, and always pass a real task reference. That briefing is prompt territory used correctly — guidance about a mechanism, not a substitute for one. On your side: put the inbox on your daily list before the agent's first purchase, decide who your backup approver is, and know where the freeze lives. The first week will produce a handful of approvals and probably a denial or two; treat both as the system introducing itself, and resist adjusting anything until a pattern, not an event, asks for it.

Grow the budget on evidence

Safe budgets grow the way credit should: on a record. A month of clean receipts, approvals granted thoughtfully, denials that read as the mandate working — that record justifies raising a bound that demonstrably binds real work, adding a vendor the job genuinely adopted, or graduating a recurring task into its own envelope with a larger total. Growth by evidence keeps every expansion defensible to future you, to finance and to auditors. The opposite pattern — budget raised in a hurry during a busy week, for reasons nobody wrote down — is how safe starts become the sprawling mandates that incident reviews later marvel at. One bound at a time, in the review, with the reason recorded: the whole method in one sentence.

The five ways budgets become unsafe

Name them so you can refuse them. One, the agent holds payment credentials it can use without a check. Two, the mandate exists but a second, ungated tool does the real buying. Three, caps copied from another job's prices, so denials train everyone to bypass. Four, approvals that became a reflex because the threshold was set where judgement is never real. Five, no review rhythm, so the budget's fit decays as prices and jobs move. Every item on that list is a process failure, not a technology failure — which is good news, because it means the safe version is available to any team willing to run the small disciplines: propose, bound, supervise, evidence, review. Give the agent its budget that way, and the budget stops being a risk you took and becomes a control you run.

Five actions to take now

  1. Start with the free plan and the agent whose spending worries you most
  2. Give the agent a mandate before its first purchase, not after its first incident
  3. Remove or gate every other path by which the agent could pay
  4. Brief the agent on its mandate: branch on decisions, never retry denials
  5. Put the approval inbox on your own daily list before go-live day

Worked example: 300 dollars, two ways

A founder wants her agent to handle a 300-dollar monthly tools budget. The unsafe version (which she almost ships): a virtual card with a 300-dollar limit stored in the agent's browser. The safe version takes twenty minutes longer: a MandateGuard policy — per-task cap 60 dollars (largest tool is 49), daily cap 90 dollars (no honest day needs more), allowlist of the four current tools' vendors, approve-over at 30 dollars (every new tool and tier change pauses), receipts flowing to her inbox routine.

In the first month the mandate earns its shape three times: a 79-dollar bundle is denied at the cap and bought monthly instead; a new tool at 25 dollars pauses for approval (granted — it is now tracked from birth); and a duplicate start on an existing tool is visible in receipts before its first renewal. The 300 dollars is spent to 212, fully explained. The card, meanwhile, still exists as a backstop limit — defence in depth, with the mandate doing the governing.

Frequently asked questions

Is a virtual card with a low limit not simpler?

Simpler to issue, weaker in every dimension that matters: no vendor logic, no per-task logic, no approvals, no decision evidence, and renewals charge it directly. As a backstop under a mandate, it is genuinely useful.

How big should the first budget be?

As small as the job's real prices allow for the first month. Authority grows on receipt evidence; starting tight is free information.

What if the agent needs to exceed the budget for a great opportunity?

Opportunities route through humans: an approval, an envelope for the specific job, or a deliberate policy change. A budget the agent can exceed on its own judgement is not a budget.

Can I give different budgets to different tasks?

Yes — that is the envelope pattern: standing policy for the agent's life, per-task envelopes with their own totals for defined work.

What is the first sign the budget is mis-set?

Denial patterns: legitimate work blocked repeatedly (too tight, in one named bound) or a month of zero denials and zero approvals (too loose to be a control).

Keep reading

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