Guide // MandateGuard
What should an AI agent spending policy template include?
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
An agent spending policy template turns one good policy into a repeatable practice. It should capture: the agent's identity and the job it may spend on; the vendor allowlist with the need each vendor serves; per-task and daily caps, each with the real price observation that justifies it; the approval threshold, approver roster and written judgement conditions; the task-reference convention; incident contacts and freeze authority; and a review rhythm with a next date. MandateGuard enforces the four settings at its core; the template wraps them in the rationale that lets humans govern, audit and adjust them over time.
What belongs in a policy template?
An agent spending policy template is a reusable starting structure for a mandate: identity and purpose, vendor allowlist, per-task cap, daily cap, approval threshold and conditions, review rhythm and incident contacts. In MandateGuard the enforceable core is the four policy settings plus any envelope conditions for a specific task. The surrounding text — purpose, rationale, review dates — is your own governance record, and it is what makes the settings interpretable months later.
The template, section by section
A complete template has seven sections. Identity: the agent's name, its job in one sentence, and the accountable owner. Vendors: the allowlist with the charging identity of each seller and the job need it serves. Caps: per-task and daily figures, each followed by the real price observation that justifies it. Approvals: the threshold, the approver roster and backup, and any written conditions for judgement calls. Evidence: the task-reference convention and where receipts and exports are reviewed. Incidents: who may freeze, and the unfreeze rule. Review: the rhythm and the date of the next review. Seven short sections, each a paragraph at most — the template is a thinking tool, and its brevity is why teams actually complete it.
Numbers must carry their reasons
A cap without a reason is a future argument. Beside every figure, the template records the observation it came from: largest legitimate stack test, 52 dollars, observed October; busiest day, four runs. Those notes do three kinds of work. They let a reviewer approve the policy without re-deriving it. They let a future operator adjust intelligently when a vendor raises prices — the note says which observation to re-measure. And they expose incoherence at writing time: if you cannot name the observation behind a number, the number is a guess, and the template has just told you which homework to do before go-live.
Approval conditions in plain words
Thresholds route by amount; conditions route by judgement. The template's approvals section should state, in sentences an approver can apply quickly, what deserves denial even when legal: first purchase from a newly allowlisted vendor above a modest sum, quantities beyond the named project, purchases whose task reference does not match an open ticket. For standing policies these conditions live in your runbook and approver practice; for a defined task, MandateGuard spend envelopes let you write the conditions onto the mandate itself, where the reviewer card displays them beside the request. Either way, write them down: unwritten conditions are applied inconsistently and audited never.
Three filled shapes to start from
A research agent typically needs small caps, a short allowlist of data and model providers, and a low approval threshold, because its purchases are frequent and individually minor — aggregate control via the daily cap does the heavy lifting. A purchasing agent needs a higher per-task cap fitted to order values, supplier allowlists by charging identity, and thresholds at quantity-step boundaries. A provisioning agent needs caps fitted to infrastructure units, strict vendor singularity (one cloud), and approvals on full-environment runs. Same template, three different centres of gravity: frequency, order value and unit size respectively. Filling the template three times teaches a team more about its agents than any framework document.
Keeping templates and policies in sync
The standing risk is drift: policies edited in MandateGuard during incidents, templates unchanged in a wiki, until neither describes the other. Counter it with direction of truth: the live policy in MandateGuard is authoritative for behaviour, and the template document is updated at the monthly review to match it, with changes and reasons appended, never rewritten silently. When a new agent is created, the template is the checklist that prevents the classic omissions — no incident contact, no review date, a threshold above the cap. The review rhythm keeps both artefacts honest with little effort: fifteen minutes a month per policy, mostly spent reading receipts.
What the template cannot do
A template cannot price a job you have not studied, cannot staff your approval queue, and cannot make an agent supply good task references — those are operational commitments the document merely records. It also cannot substitute for the enforcement layer: a beautiful policy PDF with no engine behind it is a wish. The MandateGuard policy is the enforceable core; the template is the governance wrapper that makes the core explicable, reviewable and repeatable across a growing fleet. Treat the pair as one artefact and agent spending stops depending on any individual's memory — which is the entire point of writing policy down.
Five actions to take now
- Copy the template's structure, then replace every number with your job's real prices
- Write approval conditions in words a tired approver can apply at a glance
- Add a review date to the policy itself — policies expire by neglect, not by date
- Keep one policy per agent and job; fork the template for new jobs
- Record each change and its reason beside the receipts it will govern
Worked example: the template filled for a provisioning agent
Template applied to a provisioning agent: purpose, test infrastructure for the CI pipeline; allowlist, the team's cloud provider only; per-task cap 60 dollars (largest legitimate stack test, 52 dollars, plus headroom); daily cap 220 dollars (busiest honest day, four full runs); approve-over 40 dollars (full runs pause, component tests flow); incident contact, the platform lead; review, monthly.
Six weeks later the receipts show full-run approvals averaging five a week and zero denials except one frozen afternoon during a provider outage drill. The template's rationale notes let a new team member explain every number in one reading — and when the provider raises prices, the monthly review adjusts the cap to 70 dollars with the reason recorded, instead of the policy silently failing in production.
Frequently asked questions
Is the template itself enforced?
The four settings and envelope conditions are enforced by MandateGuard. The template's text sections are governance documentation — essential for humans, not parsed by the engine.
Can I reuse one template for all agents?
Reuse the structure, never the numbers. Each job's prices, vendors and risk shape differ; fork and fill per agent.
What numbers should a research agent start with?
From its real prices: typical data batches and API top-ups. Many small research agents run well with caps in the 40 to 60 dollar per-task range and approval near half the cap — validate against your vendors.
How often should the template be reviewed?
Monthly per policy, and the template itself quarterly: are we still capturing purpose, conditions and incident contacts for every new agent?
Where do envelope conditions fit?
As a per-task annex: a defined job gets an envelope with its own maximum, vendors, period and written approval conditions, referencing the standing policy.
Keep reading
- How do you set up your first AI agent spending policy?
- How do you write approval conditions on a spending mandate?
- What are the most common AI agent spending policy mistakes?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides