← Go back

Guide // MandateGuard

How do you audit AI agent spending?

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

You audit AI agent spending from the decision ledger, not from card statements: every authorize outcome — ALLOW, DENY, REQUIRE_APPROVAL — is a signed, hash-chained receipt naming the vendor, amount, task reference and reason. Verify the chain first to prove the record is intact, reconcile allowed totals per vendor against invoices to catch spending paths outside the mandate, then read approvals and denial clusters as the control narrative. On paid plans, evidence export packages receipts plus the chain verdict for external auditors. The discipline that makes it all work is set at integration time: meaningful task references on every call.

What makes agent spending auditable?

Auditing agent spending means reconstructing, from trustworthy records, what the agent proposed, what the policy decided and what was ultimately spent. In MandateGuard the audit substrate is the receipt chain: signed, hash-linked decision records with vendor, amount, task reference, decision and reason, verifiable on demand through verify-chain and packageable through evidence export on paid plans. The chain — not chat logs or card statements — is the primary source an audit should start from.

Start from decisions, not statements

Card statements and vendor invoices show what was charged; they cannot show what was refused, what a human approved, or which task a purchase belonged to. An audit of agent spending that starts from statements will always be partial, because the most important control events — denials and approvals — never touch a statement. Start from the decision ledger instead. MandateGuard receipts record every authorize outcome with the facts an auditor needs, and because each receipt is hash-chained and signed, the ledger carries its own integrity proof. Statements then play their proper role: corroborating that allowed decisions and actual charges agree, vendor by vendor.

Reconcile in three passes

A practical audit runs three passes. First, integrity: run verify-chain over the ledger and record the verdict and receipt count for the period. Second, completeness: total the allowed amounts per vendor from receipts and compare with invoices and statements; investigate gaps in both directions — charges without receipts mean a spending path outside the mandate, receipts without charges mean failed executions or undelivered purchases. Third, judgement: read the approvals and a sample of denials, checking that approved items show coherent task references and that denial patterns were actioned. Three passes, each with a defined question, beat the unfocused sampling that agent spending audits default to when the only records are chat transcripts.

Read denials as the control narrative

Auditors assessing a control want evidence it operated, not merely that it existed. Denied receipts are exactly that evidence: dated, reasoned refusals, with attempted amounts. A quarter with zero denials across a busy agent deserves a question — either the mandate is perfectly fitted or it is not being tested by reality. A quarter with denial clusters should show follow-up: a cap deliberately revised, a workflow fixed, an agent frozen during an incident. Trace two or three clusters to their resolutions during the audit. That trace — attempt, refusal, human response — is the story that converts a policy document into a demonstrated control, and it is only available because attempts are receipted.

Task references make or break the audit

The join between the ledger and the business is the task reference. With disciplined references — ticket numbers, order IDs, campaign codes — every receipt lands on a business object, and sampling becomes trivial: pick a ticket, read its decisions. Without them, auditors face amounts and vendors with no stated purpose, and the audit degrades into interviewing the agent's operators about what probably happened. If your references are weak, the audit's most valuable output is fixing them prospectively: enforce the convention in the calling code, refuse authorize calls without a reference, and let the next period audit itself. Retrospective reconstruction is the expensive habit this prevents.

Approvals and the human layer

Agent audits increasingly ask a governance question: which spends did a human personally authorise? The receipt chain answers it precisely, because REQUIRE_APPROVAL decisions and their outcomes are recorded with the facts the approver saw. Review approval latency and concentration as part of the audit: approvals granted within seconds in bulk suggest rubber-stamping; approvals concentrated in one person suggest a key-person risk worth a backup approver. Neither is a ledger failure — both are findings the ledger makes visible. The one-time token design helps here too: because tokens are bound and single-use, an approval in the chain corresponds to at most one executed purchase of exactly the approved shape.

Package the evidence before you are asked

Audits go better when the pack exists before the questions arrive. On paid plans, evidence export packages the receipts with the chain verification — the material an external auditor, insurer or enterprise customer's vendor review actually needs. Produce it at period end as a routine, verify the chain in the same sitting, and store it with the period's other financial evidence. The same pack serves incident response: after a bad day, the question is never what do we remember, but what the chain shows. Teams that export routinely discover that the audit stops being an event and becomes a by-product of running the mandate properly.

Five actions to take now

  1. Require task references that match your ticket or order system from day one
  2. Schedule a monthly receipt review with the chain verdict attached
  3. Sample-reconcile receipts against invoices every cycle
  4. Review denials and approvals as audit evidence, not noise
  5. Export the evidence pack at audit time and verify the chain in the same sitting

Worked example: a quarterly audit in one afternoon

A finance lead audits a quarter of agent spending in one afternoon. She starts with the chain verdict: verified, 486 receipts. She exports the pack (Pro plan) and reconciles allowed totals per vendor against invoices: the model provider matches to the cent; the data broker shows 3 allowed purchases with no matching delivery notes, worth 147 dollars, which she queries with the vendor using receipt numbers.

Denials tell the control story her auditor wants: 23 refused attempts including a frozen period during a tool incident, each stamped with its reason. Approvals show 31 human decisions with the approver-visible facts. The audit closes with two actions — a vendor credit for the undelivered purchases and a task-reference fix in one workflow — instead of the open-ended sampling exercise the previous, pre-mandate quarter required.

Frequently asked questions

What is the primary audit record?

The receipt chain: signed decisions with vendor, amount, task reference, decision and reason, hash-linked and verifiable. Statements and chat logs corroborate; the chain leads.

How do I prove records were not edited?

Run verify-chain: it recomputes hashes from stored rows. A verified verdict across the period is the integrity evidence; export it with the pack.

Do auditors care about denied attempts?

Yes — denials are control evidence. They show the mandate operating, and they quantify attempted overspend the business did not incur.

Can receipts reconcile to my accounting system?

Task references join receipts to your orders and tickets; allowed totals per vendor reconcile to invoices. The joining discipline must be set at integration time.

What does the free plan give an audit?

Full receipts and chain verification for one agent. Evidence export packaging is a paid-plan feature.

Keep reading

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