Guide // MandateGuard
How do you define a spend envelope, step by step?
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
Defining a spend envelope takes five decisions and one short form: the task written as a deliverable, the agent that will execute it, a maximum total derived from a quote or estimate with contingency visible, the charging vendors the task will pay (or a deliberately empty list), a real period, and an approval level at the task's judgement boundary, with approval conditions in decision language. Define it on the MandateGuard envelopes surface with your workspace key, note the envelope ID on the project record, and prove it with three test evaluations — within, needs approval, outside — before real spends depend on it.
What do I need to define an envelope?
Defining a spend envelope is the act of creating a task mandate on MandateGuard's envelopes surface: with a workspace key loaded, you state the intended task, the agent, the maximum total cost, approved vendors, an optional period, approval conditions and an optional single-spend approval level. The service returns an envelope ID (env_…) that spend requests then reference for evaluation on the reviewer card.
Before the form: the five decisions
The definition form is short; the thinking before it is the actual work. Decide the task sentence (what deliverable, for whom), the money (maximum total, derived from a quote or estimate rather than a mood), the counterparties (which charging vendors the task will pay), the time (period start and end that match the task's real schedule) and the supervision (which single spends deserve a human, expressed as an amount). Operators who make those five decisions on paper first complete the form in minutes and produce envelopes that judge cleanly. Operators who start from the form tend to produce envelopes that mirror their uncertainty — padded totals, vendor lists of everyone plausible, conditions written as hopes.
The task sentence and the agent
Write the intended task as a deliverable: enrich the 400-lead spring list, migrate the docs site, produce the launch video assets. The sentence is not decoration — reviewers judge ambiguous requests against it, and auditors read it to understand what the money was for. Bind the envelope to the agent that will do the work, using its agent ID from the workspace. Binding matters because evaluation and receipts then join that agent's history: the envelope's spends appear in the same chain as the agent's policy spends, attributable to one actor and one task. Loading your workspace key on the envelopes page lists your existing envelopes, which is worth a glance before defining a new one — duplicate envelopes for the same task are their own small chaos.
The maximum total: show your working
Set the maximum from the best cost evidence available: a vendor quote, the last similar task's receipts, a bottom-up estimate with quantities. Add contingency as a visible part of the figure (quote 280, contingency 70, maximum 350) rather than inflating line items — visible contingency can be released consciously; smuggled padding just becomes spend. Remember the maximum is a hard edge in evaluation: requests that would exceed the envelope's remaining total judge as outside the mandate, which is the behaviour you want at the task's end, and an annoyance mid-task if the figure was fantasy. If the task genuinely overruns, that is a redefinition conversation with the budget holder, held with receipts in hand.
Vendors and period: the quiet fields that do heavy work
The approved-vendor list decides which counterparties the task may pay at all. Name charging identities exactly as your agent will present them, confirmed from invoices or vendor dashboards where possible. Leaving the list empty is a legitimate choice — it means no vendor restriction for this envelope — but it should be a choice, made because the task's vendors are genuinely open-ended, and compensated by a lower approval level. The period fields deserve the same care: real start and end dates convert the envelope into a self-expiring grant, and stale-task spending (an agent charging a finished project's envelope) becomes an automatic outside-mandate verdict instead of a discovery at reconciliation.
Approval conditions and the approval level
The approval level is the envelope's numeric supervision: single spends over this amount need a human on the reviewer card. Set it at the task's judgement boundary — the size where a mistake would sting or a choice deserves a second mind — not at a round fraction of the total. The written approval conditions complement it with the qualitative rules: conditions the reviewer applies, stated in decision language. (In the current surface, conditions are recorded on the mandate as text the reviewer works from, alongside the computed checks.) After defining, run the test every new envelope deserves: evaluate one ordinary spend (expect within), one over the level (expect needs approval), one from an excluded vendor (expect outside). Three evaluations prove the mandate reads as you intended, before real money tests it for you.
Living with the envelope, then closing it
During the task, the envelope's own record is the status report: spent against maximum, decisions by state, time remaining in the period. Quote it in stand-ups instead of estimating. Resist mid-flight edits; if the task's facts change materially, define the change deliberately — a successor envelope for the new phase is cleaner than a silently stretched original — and note the reason where the team keeps its records. At close, do the five-minute retrospective: maximum versus actual, outside-mandate attempts and what they meant, approval volume versus the level's intent. File the envelope ID with the project's records. That small closing ritual is what turns envelopes from a control you use into a practice you improve.
Five actions to take now
- Load the workspace key, then write the task as a deliverable in one sentence
- Size the maximum from a quote or bottom-up estimate, contingency visible
- List the charging vendors the task will pay; leave empty only on purpose
- Set the period to the task's real start and end dates
- Set the approval level at the task's judgement boundary, then evaluate a test spend
Worked example: an envelope defined in fifteen minutes
A team defines an envelope for a website-migration task. Task sentence: migrate the docs site to the new host, including asset transfer and redirect checks. Maximum: 350 dollars, from the host's quote (280 dollars) plus a 70-dollar contingency for extra transfer volume. Vendors: the new host and the asset-delivery network — two entries, both charging identities confirmed from their invoices.
Period: the migration week's Monday to Sunday. Approval conditions: single spends over 100 dollars need approval; approval level set at 100 dollars. They evaluate a test spend of 40 dollars from the host: within mandate. A test spend from a domain registrar: outside mandate — correct, because domains were deliberately excluded. The envelope goes live with its ID recorded in the project ticket, and the migration's four real spends all judge cleanly against it.
Frequently asked questions
Do I need a paid plan to define envelopes?
Envelopes are part of MandateGuard's feature surface; check your workspace's plan for the task-specific mandate capability. The definition flow itself lives at /envelopes.
What if I do not know the exact vendors yet?
An empty vendor list means no vendor restriction for the envelope. Use it knowingly, and tighten the other fields (total, approval level) while the vendor set is open.
Can I edit an envelope after spends begin?
Treat mid-flight changes as redefinitions: deliberate, noted, and sparing. The receipts already written were judged against the terms as they stood.
What agent ID do I use?
The ID of the agent that will execute the task, from your workspace's agent list. Envelopes bind to an agent so evaluation joins the right history.
How do I close an envelope?
Let its period end, or stop referencing it and hold the closing review. The record remains readable for audit.
Keep reading
- What is a spend envelope for an AI agent?
- Why use a task-specific mandate? One task, one envelope
- How does the reviewer card decide within, needs approval, or outside mandate?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides