← Go back

Guide // MandateGuard

What are the most common AI agent spending policy mistakes?

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

The common agent spending policy mistakes are all visible in receipts before they become incidents: a per-task cap with no daily cap (or the reverse), numbers guessed instead of derived from real prices, an allowlist that has become a museum of dormant permissions, an approval threshold placed above the cap where it can never fire — or feeding an inbox nobody staffs — policy edited mid-incident instead of in the calm of a review, and a missing memory system (placeholder task references, no verification rhythm, no exports). Every one is cheap at setup: set both caps from prices, keep the list breathable, place and staff the threshold, and run the monthly review.

What are the most common policy mistakes?

The common policy mistakes are recurring configuration and governance errors in agent mandates: patterns visible across deployments — missing caps, guessed numbers, bloated allowlists, unstaffed approvals, thresholds placed where they can never fire, and policies edited reactively during incidents. Each maps to a MandateGuard setting or routine, and each is cheap to avoid at setup and expensive to discover in production.

Mistake one: only one cap

A per-task cap without a daily cap licenses a thousand legal purchases; a daily cap without a per-task cap licenses one purchase the size of the whole day. The two caps guard different geometries of loss — the single event and the aggregate pattern — and deployments that set only one usually set it from the failure they most recently imagined. The receipts make the gap obvious in hindsight: either a day full of small charges nobody bounded, or one large charge the day absorbed because nothing bounded the single decision. The fix costs one field. Whenever you review a policy — yours or a vendor's reference architecture — check for the pair before reading any of the numbers.

Mistake two: numbers from imagination

Guessed caps fail in both directions at once: too tight for the real job (manufacturing denial friction and exception habits) and too loose for the real risk (headroom nobody priced). The tell is round numbers with no provenance — nobody's largest legitimate purchase is exactly 500 dollars. The discipline that prevents it is the pricing step: list the job's purchases at real vendor prices, then set caps from the largest and the busiest. Policies written that way read differently (cap 120, because the full data pull is 112) and they survive reviews, because the number can be re-derived when challenged. A cap whose origin story is a shrug will be renegotiated by the first incident on terms the incident chooses.

Mistake three: the allowlist as a museum

Allowlists accrete: a vendor added for a one-off, a reseller added during an incident, a renamed company's old and new identities side by side. Each dormant entry is standing permission to pay a counterparty nobody currently needs — surface area with no upside. The museum's opposite failure also occurs: the list frozen so hard that legitimate new vendors generate weeks of denial friction until someone disables their attention to denials entirely. The maintenance rhythm prevents both: additions only with a named job need, removals at the monthly review for entries with no recent receipts, and denied-vendor attempts reviewed weekly so genuine needs surface as proposals rather than workarounds. An allowlist should always be describable in one breath: these vendors, for these jobs.

Mistake four: the dead threshold and the unstaffed inbox

A threshold at or above the per-task cap can never fire — REQUIRE_APPROVAL becomes a state the system documents but never enters. Nearly as bad is a live threshold feeding an inbox nobody works: requests age, operators learn the agent is where work goes to wait, and purchasing migrates back to uncontrolled channels with the mandate's blessing-by-neglect. Audit both mechanically: confirm approve-over is strictly below the cap and roughly at the judgement boundary; confirm a named approver and backup exist and that median approval latency is measured at reviews. The threshold is a promise that a human will look; staffing is keeping the promise. Between them they decide whether the human layer of the whole system is real.

Mistake five: editing policy mid-incident

Under incident pressure, teams reach for the policy: raise the cap to clear the backlog, widen the allowlist to let the agent finish, unfreeze to see if it happens again. Each edit destroys evidence (the mandate's state during the event becomes ambiguous) and usually rewards the anomaly with a wider mandate. The playbook's order exists precisely against this reflex: freeze first, read receipts, classify, fix at the causing layer — and let policy changes happen later, in the review, with the incident's receipts as the evidence and the change recorded with its reason. A policy that only ever changes in calm rooms, on evidence, stays interpretable for years; a policy edited in emergencies becomes a palimpsest nobody can audit.

Mistake six: no references, no reviews, no memory

The quietest mistakes are omissions of the memory system: agents calling authorize with placeholder task references, so receipts cannot join to business records; no chain verification rhythm, so integrity is asserted rather than demonstrated; no monthly review, so drift accumulates unobserved; no evidence export, so the first external audit starts from zero. None of these shows up as an incident. They show up as an organisation that cannot answer simple questions about its own agents — which is, over time, the most expensive state of all, because every other control's value depends on the record being readable. The fix is the routine itself: references enforced in code, verification and review on the calendar, exports filed per period.

Five actions to take now

  1. Set both caps — per-task without daily (or the reverse) is half a control
  2. Derive numbers from real prices; a guessed cap is a future incident
  3. Write the allowlist from the job, and never widen it mid-incident
  4. Place approve-over below the cap, and staff the inbox it feeds
  5. Read the first month's denials before changing anything at all

Worked example: a policy with every mistake, rebuilt

A team's policy, set in a hurry: per-task cap 500 dollars (no daily cap), allowlist of fourteen vendors (everything the agent had ever touched), approve-over at 500 dollars (equal to the cap, so it never fired), one approver (on leave for two of the first six weeks). In six weeks the agent spends 4,120 dollars — every dollar authorised. The receipts show the shape: many small purchases summing past any sane day, two subscriptions started from vendors nobody can justify, and zero approvals ever requested.

The rebuild applies this guide line by line: daily cap added at 220 dollars, cap refitted to 120 dollars from real prices, allowlist cut to four vendors with named needs, threshold placed at 60 dollars, backup approver named. The next six weeks spend 1,860 dollars with nine approvals and four denials — and, not coincidentally, the team's first real picture of what its agent actually does all day.

Frequently asked questions

Which mistake is most damaging?

The unstaffed inbox, because it silently converts a controlled system into either blocked work or pressure to bypass. The others misprice risk; that one dissolves the control socially.

How do I spot a bloated allowlist?

Entries with no receipts in a quarter, or entries nobody can attach to a current job need. Both are findings at the monthly review.

Is a high cap with a low threshold a valid design?

It can be: the cap as catastrophe backstop, the threshold as the real governor. It works only if approvals are genuinely staffed — otherwise the low threshold just manufactures queue pressure.

What if our job's prices genuinely vary wildly?

Use the threshold and envelopes for variance instead of a giant cap: routine band under a fitted cap, exceptional jobs in task envelopes with their own totals.

How often do these mistakes recur after fixing?

Each fix holds if the monthly review watches for its signature. Policies rot toward convenience without a rhythm; the review is the anti-rot mechanism.

Keep reading

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