Guide // MandateGuard
How do you choose AI agent spending control software?
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
Choose agent spending control software by drilling worst-day behaviour, not feature matrices. The criteria that decide outcomes: enforcement in an engine outside the model; three decisions (allow, deny, escalate) with approvals released as bound, single-use tokens; evidence as a hash-chained, self-verifiable, exportable ledger that includes refusals; one-action containment (freeze, suspend) with clean key and lifecycle hygiene; and interoperability — API, MCP, webhooks and honestly-described standards alignment. Pilot the winner on your riskiest agent for a month and let its receipts, not its slides, argue for the fleet rollout.
What should spending control software do?
Agent spending control software is the system that authorises agent purchases against mandates before execution, evidences every decision, and gives humans the approval and containment controls. Evaluation criteria: enforcement location (outside the model), decision completeness (allow, deny, escalate), evidence quality (chained, verifiable, exportable), human controls (bound approvals, kill-switch), interoperability (APIs, MCP, shared intents) and claim honesty (what the vendor's own site can prove).
Start with the failure you are buying against
Spending control is insurance-shaped software: its value concentrates on the worst day, not the average one. Choose against the incidents you actually fear — the premium-tier accident, the duplicate subscription year, the 3 a.m. retry storm, the audit you cannot evidence — and require every candidate to demonstrate its behaviour in those shapes. Feature matrices flatten the differences that matter (every vendor claims caps and approvals); behaviour under the drill does not. The criteria below are ordered by how much they determine worst-day outcomes, which is why enforcement location outranks dashboards and evidence outranks analytics.
Criterion one: enforcement outside the model
The control must execute in machinery the agent cannot influence: a policy engine evaluating concrete requests against stored mandates. Prompt-level rules, framework suggestions and post-hoc alerts all fail this test in the same way — they describe the boundary in the same medium the boundary must restrain. Ask the vendor to show you the refusal path: a request arrives, the engine evaluates, the purchase cannot proceed, and the refusal is recorded. Then ask the harder question: what could the agent do to spend anyway — another tool, a stored card, a different endpoint? The honest vendors will discuss path coverage with you; the others will return to the dashboard.
Criterion two: three decisions, not two
Binary allow/block systems force the bad choice between high caps (unsupervised risk) and low caps (blocked work). The third state — escalate to a human with the facts, resolving into a bound, single-use execution token — is what makes strict mandates livable. Evaluate its details: what does the approver see (vendor, amount, business reference, history)? Is the release bound to the approved vendor and amount and redeemable once, or is it a boolean the agent can lean on? Can approvals be notified by webhook into your ops flow? The quality of this one mechanism predicts daily usability more than any other feature, because it is where humans and the system meet every day.
Criterion three: evidence you can verify without the vendor
Ask what record a decision leaves and who can check it. The strong answer: every decision (including refusals) written contemporaneously to a tamper-evident, hash-chained ledger; verification runnable by you on demand; period exports for auditors; and, for shared mandates, receipt verification available to counterparties with read-only keys. The weak answers involve screenshots, proprietary viewers and trust-us assurances. Also check the data philosophy: does the vendor's trust or risk scoring use your data only, or pooled and third-party signals you cannot inspect? MandateGuard's vendor trust, computed solely from your own ledger with counts shown, is the pattern to demand — evidence whose provenance you can state in one sentence.
Criterion four: containment and lifecycle
Incidents test the controls nobody demoed: the kill-switch (one action, immediate, receipted refusals), freeze and suspension semantics for shared mandates, key management (show-once values, hashed storage, role-scoped counterparty keys, individual revocation), and offboarding (retiring an agent's authority cleanly). Drill the freeze live in evaluation; time it. Then ask about the quiet lifecycle: policy change history, key rotation, data export and deletion. A control plane is a long-lived trust position in your stack; its lifecycle hygiene is a security property, not an administrative detail.
Criterion five: interoperability and honest claims
The control must meet your agents where they are: a clean API for integrated stacks, MCP tools for tool-native agents, webhooks for human workflows, and — if you issue or accept mandates across organisations — a registered intent service with role-scoped verification, aligned honestly with the emerging standards (AP2-style shape, EMVCo draft) without certification theatre. Finally, audit the vendor's claims as a proxy for its engineering culture: are prices on the site the prices in the contract, are counts real, are draft alignments labelled drafts? A vendor careless with its own claims will be careless with your mandate's semantics. MandateGuard's materials, for instance, state plan limits, prices and the draft status of its standards alignment plainly enough to check — hold every candidate, including it, to that standard in the drill.
Five actions to take now
- Demo the refusal path, not just the dashboard: over-cap, unlisted vendor, frozen agent
- Confirm decisions are receipted in a tamper-evident chain you can verify yourself
- Check approval releases are bound and single-use, not a flag flip
- Verify the vendor's counts, prices and claims against the live site
- Pilot on your riskiest agent for a month before fleet rollout
Worked example: the three-drill evaluation
A platform team evaluates three tools with a scripted drill instead of slideware. Drill one: an over-cap purchase from an approved vendor. Tool A logs a warning and lets it through (prompt-level). Tool B blocks but keeps no record of the attempt. MandateGuard denies with a reason and a receipt. Drill two: an approved purchase re-presented at double the amount. Only the bound-token design refuses it. Drill three: produce last month's decisions for an auditor, verifiable without the vendor's help. Only the chained, exportable ledger answers in the meeting.
The team also checks claims the drills cannot: pricing against the live pricing page (Free, Pro 29, Team 79), the intent service's standards language (draft alignment, not certification), and whether vendor trust uses first-party data only. The selection memo writes itself from drill results — which is the point of drilling.
Frequently asked questions
What is the single most important criterion?
Enforcement location. If the control can be talked around by the agent or edited by the stack it governs, everything else is presentation.
Should we prefer an all-in-one agent platform?
Bundled controls tend to be shallow at the decision layer. A dedicated policy engine that integrates (API, MCP) usually evidences better. Judge by the drills, not the bundle.
How important is MCP support?
If your agents are MCP-native, the check must exist as a tool or it will be skipped. MandateGuard's MCP tools are part of its paid tiers.
What about open-source options?
Evaluate them by the same drills. The evidence and approval-binding criteria are where lightweight projects most often thin out.
Can we migrate later if we choose wrong?
Receipt history portability varies. Ask each vendor how decision history exports; a chain you cannot take with you is a lock-in with better marketing.
Keep reading
- AI agent spending policy vs prompt instructions: what actually enforces?
- What should an AI agent spending audit trail include?
- Free vs Pro vs Team: which MandateGuard plan fits your agents?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides