← Go back

Guide // MandateGuard

How does a vendor allowlist work for 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

A vendor allowlist is the short list of sellers an AI agent is permitted to pay. In MandateGuard, every authorize call names a vendor; if that vendor is not on the policy allowlist, the answer is DENY with the reason recorded — regardless of amount. Allowlists work because they invert the default: instead of trying to enumerate every bad seller, you enumerate the few good ones the job needs. Keep the list exact and short, treat every denied unknown vendor as information, and change the list deliberately when the job changes, never in the middle of an incident.

What is a vendor allowlist?

An allowlist (sometimes called a whitelist) is a policy list of approved vendor identifiers. MandateGuard compares the vendor named in each authorize request with the policy allowlist before allowing any spend. Membership is binary and auditable: a vendor is on the list, with every allowed and denied decision receipted, or it is not, and every attempt to use it is denied and receipted too.

Why allowlists beat blocklists for agents

A blocklist assumes you can predict the sellers you must avoid. With agents, you cannot: the open web generates lookalike stores, resold storefronts and sponsored placements faster than any list can be maintained, and an agent is far more literal than a human shopper about names that merely resemble the right one. An allowlist needs no prediction. It says the job requires these three vendors, and everything else — the almost-right name, the cheaper unknown broker, the marketplace seller with perfect reviews — is out of scope until a human decides otherwise. The failure mode also flips in your favour: a blocklist fails open when it misses a bad actor; an allowlist fails closed when it misses a good one, producing a denial and a conversation instead of a charge.

How enforcement works in MandateGuard

The vendor string travels with every authorize request, alongside the amount and task reference. The engine checks allowlist membership as part of the same evaluation that checks the caps: a non-member is denied even for one dollar, and the receipt records the vendor that was attempted. That recorded attempt is operationally valuable. One attempt usually means the agent found a new source worth reviewing. Repeated attempts after a denial can mean the agent is following instructions from tool output or a web page rather than from its mandate — a steering signal worth investigating quickly. Either way, you learn about it from evidence, on the day it happens.

Building the first list

Derive the list from the job, not from the agent's browsing history. Write down the purchases the job makes, name the seller of record for each — the entity that actually charges, which is sometimes the platform rather than the brand — and add exactly those. In the MandateGuard demo policy, the allowlist is openai and aws: two entries, both load-bearing. Resist adding a vendor just in case; every entry is standing permission for the agent to move money to that counterparty at any hour. If a job genuinely needs an open-ended category of sellers, that job is a candidate for human purchasing or for a spend envelope with explicit approval conditions rather than a standing allowlist entry.

Keeping the list alive without letting it rot

Allowlists rot in two directions. They accumulate entries added during incidents and never removed, until the list quietly describes every vendor the agent has ever touched. Or they freeze while the business moves on, generating denial friction that tempts someone to disable the check. The maintenance rhythm that prevents both is simple: review denied-vendor receipts weekly, review the list itself monthly. Additions should name the job need they serve; removals should follow job changes and vendor replacements. Because every decision is receipted, you can also see which entries are never used — dormant permission is a candidate for removal, shrinking the surface the agent can reach.

Allowlists, trust scores and duplicates

Membership answers whether a vendor may be paid at all. Paid plans add two questions on top. Vendor trust, computed only from your own workspace's receipt history, shows how a listed vendor has actually behaved for you: allowed and denied counts and totals beside every score. Duplicate-commitment detection watches for the same vendor and amount being authorised again inside your window, which catches the realistic accident of an agent re-subscribing or re-ordering from an approved vendor. The allowlist remains the gate; trust and duplicates are the judgement layered over the vendors that pass it.

Edge cases: sub-processors, marketplaces and renamed vendors

Real billing is messy. A vendor may charge through a payment descriptor that differs from its product name, a marketplace purchase may settle under the marketplace's identity, and companies rebrand. Handle these as allowlist design decisions, recorded when you make them: allowlist the charging identity you can verify, keep a note of which job each entry serves, and when a vendor renames, replace the entry deliberately rather than adding the new name alongside the old forever. MandateGuard's denial receipts will show you the exact vendor string the agent presented, which is the ground truth you need to make those calls.

Five actions to take now

  1. List the seller of record for every purchase the job makes
  2. Add only those vendors to the policy allowlist
  3. Read denied-vendor receipts weekly — attempts are information
  4. Remove entries the job no longer uses at a monthly review
  5. Never widen the list mid-incident; review the denial first

Worked example: the lookalike vendor that never got paid

A purchasing agent may pay two vendors: its model API provider and its cloud provider. During a research task it encounters a sponsored page for a data broker whose name differs from an approved analytics vendor by one word, offering the dataset it needs at 45 dollars — inside its per-task cap of 50 dollars. It calls authorize with the broker's name. The amount is fine, the day has room, and still the answer is DENY: vendor not allowlisted.

The owner sees the denial in the receipts that evening, checks the broker, and finds a storefront with no company registration and prices scraped from the real vendor. The allowlist turned a plausible 45-dollar loss — and whatever follow-on charges the broker intended — into a one-line receipt. The agent completed its task from the approved analytics vendor at 52 dollars, routed through a human approval because it exceeded the cap slightly and the job genuinely needed it.

Frequently asked questions

What vendor name should I put on the allowlist?

The identity the agent will present at authorize time — the seller of record. Denial receipts show the exact string attempted, so you can align entries with reality.

Can an agent pay a marketplace with many sellers?

The allowlist checks the vendor named in the request. If you allowlist a marketplace, you are trusting its seller pool; many teams instead keep marketplace buying human-led.

What happens if I leave the allowlist empty in an envelope?

In spend envelopes, an empty approved-vendor list means no vendor restriction for that envelope; the policy allowlist still governs ordinary authorize calls. Know which surface you are configuring.

Does the allowlist block free trials that later convert?

Any charge the agent initiates is an authorize call. A converting trial presents a vendor and amount like any purchase, and an unlisted vendor is denied.

How many vendors should be on a list?

As few as the job needs. Two to six is typical for a focused agent. A list of forty is usually a job description that needs splitting.

Keep reading

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