An agency should let an AI buying agent go live only when it can show, in writing, what the agent may commit, what it has been tested against, who it will trade with, how each deal is recorded and which named person answers for it. If any of those five answers is missing, the agent is not ready to spend client money on its own. It may still be ready to observe, recommend or draft.

TL;DR. This printable checklist holds 25 yes or no checks, five in each of five areas: mandate and authority, testing and benchmarking, counterparties and identity, records and reconciliation, and people and escalation. Each check has one line to confirm and one sentence on why it matters. A failed check should narrow what the agent is allowed to do until the gap is closed, rather than end the project.

Why do agencies need a go-live checklist for AI buying agents?

Expectation is running ahead of written policy. In IAB Europe's 2026 Impact of AI on Digital Advertising Report, 58% of respondents expected agentic ad buying to reach operational use or scale within a year, yet only 48% had AI guidelines specific to advertising and marketing. Dimitris Beis of IAB Europe summed up the practical question as "who is accountable when something goes wrong?"

This checklist is built around that question. It assumes a staged approach known as earned autonomy: observe, recommend, draft, human-approved action, then bounded automatic action. The step-by-step mechanics of running a buy are covered in how to use AI agents for media buying, so this page deals only with the go or no-go decision.

Mandate and authority: what exactly is the agent allowed to do?

1. The mandate is written down, signed off by the client and versioned. An agent cannot be held to instructions that exist only in a briefing call or a planner's head.

2. Spend limits are set per deal, per day and per campaign. A single cap on the total budget does not stop an agent committing too much of it to one seller in one afternoon.

3. The autonomy stage is named for every type of action. Researching inventory, drafting a proposal and committing money carry different levels of risk, so each needs its own stage on the earned autonomy scale.

4. Hard limits are written separately from preferences. Melissa Bonnick, head of programmatic at JPMorgan Chase, has described clear restrictions as a way of telling the machine "Here is your box, stay in [the] box", and that only works when the agent can tell a wall from a suggestion.

5. The mandate is configured per client, not per agency. Scott Ensign, chief strategy officer at Butler/Till, told Digiday that agentic environments "are going to have to be controlled on a client by client basis", because clients in sectors such as pharmaceuticals and alcohol carry different levels of risk aversion.

Testing and benchmarking: has the agent proved it can do the job?

6. The agent has rehearsed against counterparty agents with no real money at stake. Sparring shows how an agent behaves under a hard counter-offer, and Alkimi's Gym is one worked example of the approach: buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves.

7. Results are graded on deal quality, not on completion. In the Concourse Bench v1 benchmark, 55 of 96 runs completed a buy, yet 95 of 110 signed contracts promised a viewability floor below 70%, so a finished workflow can still leave the buyer weakly protected.

8. The test briefs look like the client's real briefs. An agent that handles a generous display brief may struggle with a compressed video budget, and only realistic briefs reveal that before launch, as set out in how to test an AI media buyer before you deploy one.

9. The agent is tested as a system, with its harness. The software that supplies the brief, checks each action and records each deal shapes the outcome alongside the model, so changing either one means testing again.

10. The pass mark is agreed before the test, not after it. A threshold fixed in advance, as in a structured 30-day agentic advertising pilot, stops a team rationalising a weak result once it has invested time in the agent.

Counterparties and identity: who is the agent actually dealing with?

11. Every seller agent's identity can be checked against a public source. The IAB Tech Lab Agent Registry associates each registering company with its GPP ID and IAB Tech Lab reviews each entry, which gives buyers an independent reference point.

12. The agent confirms the seller is authorised to sell the inventory on offer. ads.txt and sellers.json already let buyers see who sells inventory directly and who acts as an intermediary, so an agent should check them rather than take a seller agent's word.

13. The protocols in use are named and versioned. AAMP, IAB Tech Lab's Agentic Advertising Management Protocols, and AdCP, the Ad Context Protocol published by AgenticAdvertising.org, both continue to develop, so a record of which version each agent speaks keeps later questions answerable.

14. The agent deals only with counterparties on an approved list. An approved list turns open-ended discovery into a bounded problem while agencies build confidence in agent identity.

15. A retried message cannot book the same deal twice. IAB Tech Lab says its AAMP 3.0 release strengthens idempotency across transactions "to help ensure consistent responses and data integrity", and an agency should confirm its own set-up does the same.

Records and reconciliation: can both sides prove what was agreed?

16. A deal record is written at the moment of agreement. A record assembled later from logs and invoices invites argument over what each side thought it had agreed.

17. Both parties hold the same record and either can retrieve it independently. When buyer and seller each keep their own version, reconciliation becomes a comparison of two accounts rather than a check against one.

18. Every agent action is time-stamped and explained. Rise, part of the Quad agency group, uses an audit log feature developed by PubMatic to spot agents drifting outside pre-set parameters, and group director Klaudia Smykowska told Digiday the aim is to "reconstruct what happened and why".

19. The terms the agent signed are compared with the mandate after every deal. Viewability floors, make-goods and flight dates can drift from the brief even when price and volume look right.

20. Finance can reconcile agent-made deals without asking the agent. If the only explanation of a commitment sits inside the agent's own session, the record is not fit for a client audit.

People and escalation: who answers when something goes wrong?

21. Every agent has a named, accountable human owner. Jonathan Whiteside, global EVP of technology at Dept, told Digiday that "Every deliverable has an accountable human", which rules out "the AI did it" as an explanation to a client.

22. Financial commitments need explicit human sign-off until the agent has earned more. WPP Media requires explicit human sign-off for financial commitments and campaign activation in its Buyer Agent for Video, according to Digiday, and that is a sound default for any first deployment.

23. Escalation triggers are defined in advance. A price outside the expected range, an unfamiliar seller or a term the mandate does not cover should reach a person straight away, not at the weekly review.

24. Someone has rehearsed pausing the agent mid-campaign. A stop control is only useful if a person has used it before the day it matters.

25. The client knows the agent exists and what it is allowed to do. David Dworin, chief product officer at FreeWheel, said at Advertising Week that most agentic systems are designed either with a human who approves everything or with very specific controls (Digiday), and a client is entitled to know which design is buying its media.

What should an agency do when some checks fail?

A failed check is a limit on autonomy, not a verdict on the project. An agent that passes the mandate and testing checks but not the records checks can still observe, recommend or draft while the gap is closed. The standards bodies point the same way: IAB Tech Lab builds AAMP on existing industry standards, and AgenticAdvertising.org governs AdCP through what it calls Embedded Human Judgment, keeping humans in the loop on decisions with real consequences.

Alkimi, which publishes the Concourse benchmark cited in check 7, approaches the checklist in three parts: Concourse for benchmarking buying agents, the Gym for sparring before money moves, and a neutral Marketplace that works within IAB Tech Lab's AAMP framework today and is being built to be AdCP compatible. In the Marketplace, humans keep approval rights at the deal level, and the DealSheet covers checks 16 and 17: it is written at agreement, timestamped, held by both parties and retrievable by either side independently.

Frequently asked questions

Can AI agents be trusted to manage advertising budgets?

Trust is best earned in stages, with an agent moving from observing and recommending towards bounded automatic action only as it passes checks like the 25 above. Until then, every financial commitment should need explicit approval from a named person.

How do agencies use AI for programmatic buying?

Agencies are running controlled tests by channel and by client, and Butler/Till's tests have extended from CTV into online video, display and streaming audio, according to Digiday. Agency leaders quoted in trade press, including at WPP Media and Dept, keep a named human accountable for what the agent does.

How do marketing managers evaluate agentic advertising platforms?

A practical starting point is the five areas in this checklist: mandate controls, testing evidence, counterparty identity, shared deal records and escalation paths. Each platform should be asked to show, not describe, how it handles all five.

What is earned autonomy in AI media buying?

Earned autonomy is a staged model in which an agent moves through observe, recommend, draft, human-approved action and bounded automatic action. Each step up is granted on evidence from the stage below, not on a promise.

Does an agency need AAMP or AdCP before an agent goes live?

Neither is a precondition, but knowing which protocol and version each agent uses makes deals easier to verify. AAMP and AdCP both publish their documentation openly, so an agency can check the specification its agents claim to follow.

25 Checks Before an Agency Lets an AI Buying Agent Go Live