← Go back

Guide // MandateGuard

What is a spend envelope 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 spend envelope is a mandate written for one task instead of for an agent's whole life: the intended task, a maximum total cost, approved vendors, a spending period and explicit approval conditions. Requests against it are evaluated by MandateGuard's real engine and shown to a reviewer on one card — within mandate, needs approval, or outside mandate — and every decision joins the same hash-chained receipt trail as policy spend. Use envelopes for sprints, projects and client-billed work whose budgets belong to the task; keep the standing policy for the agent's continuous operations.

What is a spend envelope?

A spend envelope is a task-specific mandate in MandateGuard: one intended task, one maximum total cost, an approved vendor set, a spending period and approval conditions, identified by an envelope ID. When an agent requests spend against it, MandateGuard evaluates the request against the envelope and presents the reviewer with a single card — within mandate, needs approval, or outside mandate — while every decision joins the same hash-chained receipt trail as policy decisions.

Why tasks need their own mandates

Standing policies describe an agent's normal life. Real work, though, arrives as tasks with shapes the normal life does not capture: a research sprint with a fixed budget, a migration weekend with unusual vendors, a client project whose costs must be provable to the client. Stretching the standing policy to fit each task — raising caps, widening allowlists, then remembering to tighten them again — is how mandates rot. The spend envelope solves the shape mismatch by scoping authority to the task itself: define the task once, fund it once, and let its authority expire with its period. The standing policy stays exactly as boring as it should be.

The five fields of an envelope

An envelope records the intended task in words, because the task is what a reviewer will judge requests against. It sets a maximum total cost — the envelope's budget, consumed as spends are approved against it. It names approved vendors (or deliberately leaves the list empty for no restriction). It may set a spending period with start and end dates, so authority cannot leak outside the task's actual window. And it records approval conditions — the written terms, such as single spends over a named amount needing approval — that the reviewer applies. Those five fields are the whole instrument, and their sufficiency is the point: everything a reviewer needs to judge a request is on the mandate, not in a chat thread.

The reviewer card: one screen, one decision

When a spend is evaluated against an envelope, the reviewer does not get a queue of fragments; they get one card showing the requested spend against the envelope: the vendor's status against the approved set, the period's validity, the conditions met or unmet, the envelope's spent-versus-maximum position, and the computed decision — within mandate, needs approval or outside mandate. That card design is a governance choice. Decisions made from complete context are faster and better than decisions assembled from three systems, and the card's completeness is what makes a human decision defensible later. The evaluation is the real policy engine's, not a mock-up: the decision it computes is the decision that is receipted.

Envelopes and the receipt trail

Envelope decisions are not a side ledger. They write into the same hash-chained receipt trail as authorize decisions, which means a task's spending story is verifiable with the rest of the workspace's history and exportable in the same evidence packs. For client-billed work this unity matters commercially: the envelope defines what was authorised, the receipts prove what happened, and the totals reconcile against invoices without a separate tracking spreadsheet that nobody quite trusts. When the task ends, its envelope remains as a closed, readable artefact — the mandate, the decisions and the total, in one place.

When envelopes beat policies, and vice versa

Reach for an envelope when the work has a nameable end, a budget that belongs to the task rather than the agent, vendors the standing policy should not permanently admit, or a third party (a client, a finance reviewer) who will later ask what this task was allowed to cost. Stay with the standing policy for the agent's continuous operational spending, where per-task and daily caps plus approvals already fit the shape. Many fleets use both constantly: the policy governs the agent's life, and each exceptional job gets an envelope laid over it, sized to the job. The anti-pattern to avoid is the permanent envelope — a task-scoped instrument quietly renewed forever, which is a policy that escaped review.

Defining envelopes well

The craft is in the writing. Name the task as a deliverable, not an activity (enrichment for the 400-lead list, not data work). Set the maximum from a quote or a bottom-up estimate, with the contingency visible rather than smuggled into a padded figure. Name vendors the task will genuinely pay; a vendor added in case is standing permission inside the task's period. Write conditions in decision language — single spends over 100 dollars need approval — not aspirations. And give the period real edges: the sprint's dates, not the quarter. An envelope written that carefully can be approved by a manager in one reading, which is the operational test of the whole instrument.

Five actions to take now

  1. Use an envelope when a task has a definable end and a definable total
  2. Name the maximum total from the task's quoted or estimated cost
  3. List the vendors the task will actually pay, not a category
  4. Write approval conditions a reviewer can apply in one reading
  5. Close the loop afterwards: compare envelope spend with the task's outcome

Worked example: one sprint, one envelope, 500 dollars

A consultancy defines an envelope for a client research sprint: task, data enrichment for a 400-lead list; maximum total 500 dollars; vendors, two data providers; period, the sprint's two weeks; condition, single spends over 100 dollars need approval. The agent's requests evaluate against it: a 60-dollar batch is within mandate; a 140-dollar rush batch needs approval and gets it from the card; a third vendor's 45-dollar offer is outside mandate, because the vendor is not in the envelope.

The sprint closes at 462 dollars spent of the 500-dollar envelope, every decision receipted. The client report cites the envelope totals directly — the mandate the consultancy set is the same artefact the client audits, which is why envelopes suit client-billed work so well.

Frequently asked questions

How is an envelope different from a policy?

A policy is standing law for an agent across all its work. An envelope is a mandate for one defined task, with its own total, vendors, period and conditions.

What does the reviewer see?

One card: the requested spend evaluated against the envelope — vendor in or out, period valid or not, conditions met or unmet — with the resulting decision.

Do envelope decisions create receipts?

Yes. Envelope evaluations write to the same hash-chained receipt trail as every other MandateGuard decision.

Can an envelope have no vendor restriction?

Yes — an empty approved-vendor list means no vendor restriction for that envelope; the agent's policy allowlist still governs ordinary authorize calls.

Which plans include envelopes?

Envelopes are a MandateGuard feature surface at /envelopes, part of the task-specific mandate capability described in the Tim Gabe Phase 8 work.

Keep reading

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