← Go back

Guide // MandateGuard

How do you control spending for AI agents that buy cloud services?

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

Control a cloud-buying agent by fitting its MandateGuard policy to your bill's anatomy: allowlist the provider's charging identity, set the per-task cap just above the largest unsupervised run, set the daily cap at the busiest honest day, and place the approval threshold at full-environment scale where configuration risk lives. Then reconcile allowed receipts against the monthly cloud bill — charges without receipts reveal spending paths outside the mandate, which is the finding the whole exercise exists to produce. Annual commits and reserved capacity belong in envelopes or human-led purchasing, never under an unsupervised agent's cap.

What do cloud agents actually spend on?

Spending controls for a cloud-buying agent are the MandateGuard policy applied to infrastructure purchases: the cloud provider on the allowlist, caps fitted to stack and run prices, approvals on environment-scale spends, and receipts reconciled against the cloud bill. The agent proposes each provisioning purchase through authorize; the mandate decides; the provider's own billing then corroborates the allowed set.

Why cloud spend is the sharp edge of agent autonomy

Cloud purchasing is where agent autonomy meets metered infinity. The catalogue is API-native, provisioning is instant, and the difference between a test and a commitment can be a single parameter — instance family, region, retention toggle. Agents are good at the mechanics and indifferent to the meter. Controls therefore matter here earlier than in most domains, and they must be fitted to the cloud's own price units: a cap means nothing until it is expressed against the real cost of a stack, a run, a day of the agent's actual workload. The MandateGuard policy gives you the four dials; this guide is about setting them from your bill's anatomy rather than from round numbers.

Fit caps to stacks, not to vibes

Inventory the agent's provisioning repertoire with prices: the component test, the integration run, the full environment, the data job. The per-task cap belongs just above the largest item the agent may start unsupervised — for many teams that is the integration run, with full environments deliberately above the approval threshold rather than above the cap, so they remain possible but supervised. The daily cap belongs at the busiest honest day: release days and incident reproductions included, retry storms excluded — those are what the cap is there to truncate. When a new workload shape appears (a GPU job, a load test), its first runs will hit the approval band; that friction is the system asking whether the shape belongs in the repertoire, and the answer should be a policy decision, not a cap shrug.

One provider, named correctly

Cloud allowlisting has a practical wrinkle: the charging identity. Providers bill through legal entities and marketplace constructs whose names differ from the product names the agent uses. Set the allowlist from the identity your invoices actually show — denial receipts will confirm the exact string the agent presents, and aligning the two is a one-time fix. Resist allowlisting a second provider just in case the agent wants to comparison-shop autonomously; multi-cloud purchasing by agent is a strategy decision with its own review, not a convenience entry. If a migration genuinely needs both, give the migration its own envelope with both vendors and a closing date.

Approvals where configuration risk lives

In cloud work, amount correlates imperfectly with risk: a 15-dollar misconfiguration that leaves resources running can out-cost a clean 200-dollar run. Approvals should therefore sit where environment-scale decisions happen — full stacks, new regions, persistent storage, anything with retention — even when their sticker price is modest. Express what you can in the mandate (approval threshold at the full-run boundary) and the rest in approver practice: the card's task reference should name the pipeline or change ticket, and approvers should deny references they cannot trace. Pair this with the cloud's own guardrails (budgets, teardown jobs) as defence in depth: MandateGuard governs the authority to start spend, cloud tooling governs the resources' lifetimes.

Reconcile receipts to the bill — the gap is the audit

The monthly cloud bill is a rare gift to governance: an independent record of what actually got charged. Reconcile it against allowed receipts. Matches build confidence. Receipts without charges suggest failed provisioning or credits (chase them; money leaks there too). Charges without receipts are the important direction: they prove a spending path exists outside the mandate — an engineer launching manually, a legacy automation, a forgotten service account. Each gap item deserves a decision: bring the path under the mandate, give it an envelope, or accept it explicitly in writing. An agent-spend programme that reconciles monthly will find its ungated paths in the first quarter; one that never reconciles will find them in an incident review.

Scale patterns for platform teams

As the fleet grows, keep the unit discipline: one agent per environment class (CI provisioning, data jobs, staging management), each with its own fitted policy, on Pro or Team. Duplicate detection starts to matter when several automations can request similar stacks; vendor trust, computed from your ledger, shows the provider relationship's real shape across all of them — useful at enterprise-agreement time, when your own receipts are the ground truth about consumption patterns. Evidence export per period gives FinOps a pack that joins mandate decisions to billing periods, which is the artefact cloud governance reviews actually ask for and almost never receive.

Five actions to take now

  1. Allowlist the cloud provider's charging identity, not its product names
  2. Cap per-task spend at the largest legitimate stack or run
  3. Set the daily cap for the busiest honest day, including retries
  4. Approve full-environment runs; let component tests flow
  5. Reconcile receipts against the cloud bill monthly — gaps mean ungated paths

Worked example: the reconciliation gap that changed a rule

A platform team's provisioning agent runs with a per-task cap of 75 dollars, a daily cap of 300 dollars, the team's single cloud provider allowlisted, and approve-over at 50 dollars. Component tests (8 to 22 dollars) flow automatically; full-stack runs (about 65 dollars) pause for approval, roughly three times a week — an accepted rhythm because full runs are where configuration mistakes get expensive.

The monthly reconciliation pays for the whole setup: receipts total 2,140 dollars of allowed provisioning, the cloud bill shows 2,310 dollars. The 170-dollar gap traces to a manually launched experiment by an engineer — legitimate, but now visible as the only spend outside the mandate. The team's rule changes accordingly: experiments go through the agent or get their own envelope, so next month's gap should be zero.

Frequently asked questions

Do caps work when cloud pricing is usage-based?

Authorise the provisioning actions and purchases the agent initiates with estimated amounts; reconcile against actual billing monthly. Estimates that drift show up as receipt-versus-bill variance worth investigating.

Should dev and prod agents share a policy?

No. Separate agents and policies keep blast radii and receipts separate; production deserves tighter approvals.

What about reserved instances and annual commits?

Treat them as envelope or human-led purchases: large, long commitments are exactly what per-task caps should refuse from an unsupervised agent.

Can the agent use two clouds?

It can, with both providers allowlisted — but ask whether one agent should hold both. Separate agents per cloud keep attribution clean.

How do I stop overnight runaway environments?

Daily caps bound purchase value; pair them with your cloud's own teardown automation. MandateGuard governs spend authority, not resource lifecycle.

Keep reading

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