Most first attempts at agentic advertising go wrong for organisational reasons rather than technical ones: brands hand an AI agent a budget before they have written down what it may do, who answers for it and how its work will be checked. The ten mistakes below are the ones that recur, each paired with the fix that prevents it.

TL;DR: AI agents can be trusted with advertising budgets only inside limits a person has written, tested and can enforce. The commonest mistakes are skipping the written mandate, skipping a benchmark, judging an agent on whether it finished rather than on what it agreed, and granting full autonomy on day one. Each has a known fix that costs little before the first live deal.

Why do brands get their first AI agent campaign wrong?

Brands get their first agentic campaign wrong because the agent is new but the habits around it are not. Teams carry over the set-up they use for a DSP, where software executes and a person decides, and then let the agent decide as well.

Caution is already the norm. A Digiday+ Research survey of more than 100 marketing professionals in the first quarter of 2026 found that, among marketers using AI in their campaigns, only 32% used it to buy social or retail media inventory. Henry Webster of KSM Media put the underlying fear plainly at Digiday's Programmatic Marketing Summit: "Would it blow a quarter's worth of budget in a weekend?"

The full sequence for a first deployment is in the guide to using AI agents for media buying.

Mistake 1: Handing over a budget without a written mandate

Why it happens: Briefs written for people rely on judgement the reader already has. An agent reads a brief literally, so anything left unsaid (excluded categories, a maximum price, the contract terms that matter) is open to interpretation.

What it costs: An unwritten limit cannot be enforced or audited. When a deal goes wrong, there is no document against which to show that the agent stepped outside its authority.

The fix: Write a mandate before the agent touches money, covering objectives, budget ceiling, price limits, permitted inventory, required contract terms and what must be escalated. The guide to writing a media buying mandate for an AI agent sets out the clauses.

Mistake 2: Skipping a benchmark before the first live deal

Why it happens: A demo shows an agent at its best on one brief. It says nothing about how often the agent finishes, or how it behaves when a seller holds firm on price.

What it costs: Reliability varies widely between agents. In Concourse Bench v1, a controlled simulated market published by Alkimi in which eight AI buyers negotiated with three software-controlled sellers on one budget, 55 of 96 registered attempts ended in a completed buy. Two models completed every attempt, while others never completed one.

The fix: Test the agent on realistic briefs against counterparty agents before real money moves, and record completion, terms secured and cost per run. Repeat the test whenever the model, the prompt or the mandate changes.

Mistake 3: Judging the agent on completion instead of deal quality

Why it happens: Completion is easy to count and looks like success on a dashboard. Checking whether the agent secured the protections the brand needed means reading the contract.

What it costs: The same benchmark shows the gap: across 110 contracts its agents signed, 95 promised a viewability floor below 70%. The buys closed, but the buyer was left weakly protected.

The fix: Score every agent deal against the mandate's required terms: price, viewability floor, make-goods, brand safety settings and flight dates. A completed deal that misses a required term counts as a failure.

Mistake 4: Setting approval thresholds so high that nothing reaches a person

Why it happens: Teams want the efficiency they were promised, so they set the spend level that triggers human sign-off above most real deals. The approval step then exists on paper only.

What it costs: Errors surface after delivery rather than before commitment, when they are far harder to correct.

The fix: Trigger approval on more than spend, so that new sellers, new formats, terms outside precedent and anything close to the mandate's limits all route to a person. Lauren Wetzel of WPP Media told Digiday that where there is "a financial or activation impact", the group requires "explicit controls, approval thresholds, accountable human oversight". The guide to designing approval thresholds for agentic campaigns covers how to tier them.

Mistake 5: Giving the agent every task at once

Why it happens: Once an agent handles one task well, it is tempting to hand it planning, negotiation, booking and optimisation at once.

What it costs: With several tasks changing at once, nobody can tell which one caused a problem, and one error spreads across all of them. David Dworin of FreeWheel described the industry to Digiday as "working up the comfort and maturity curve right now".

The fix: Use earned autonomy, in which the agent moves through five stages (Observe, Recommend, Draft, Human-approved action and Bounded automatic action) and advances only when its record at the current stage justifies it. The explainer on what earned autonomy means sets out the evidence each step needs.

Mistake 6: Skipping identity checks on counterparty agents

Why it happens: In agent-to-agent trading the counterparty is software, and an agent's claim to represent a publisher or a buyer is only a claim until someone checks it.

What it costs: A deal agreed with an agent that does not represent the inventory it sells leaves nobody accountable on the other side, and the problem shows only when delivery fails.

The fix: Check who operates every counterparty agent before negotiating. IAB Tech Lab's AAMP includes an Agent Registry in which each registering company is tied to its GPP ID, entries are reviewed by IAB Tech Lab and validation tools let users run their own checks.

Mistake 7: Leaving no named human owner

Why it happens: Agentic projects often sit between media, data and procurement teams, so each assumes another is accountable.

What it costs: Digiday reported that a single incorrect CPM could make or break an ad buy, leaving the agency rather than the agent answerable. Without a named owner, nobody inside the brand is positioned to catch it.

The fix: Name one person who owns the mandate, approves changes to it and can stop the agent. AgenticAdvertising.org, which publishes AdCP, governs through a principle it calls Embedded Human Judgment, under which humans stay in the loop on decisions with real consequences. Glenniss Richards of Bayer put it more briefly: "I want a person overseeing the bot."

Mistake 8: Letting each side keep its own deal record

Why it happens: Each agent logs what it believes was agreed, in its own format, on its own systems.

What it costs: Reconciliation then compares two accounts that may never have matched, and small differences in impressions, flight dates or targeting accumulate across many deals.

The fix: Agree one shared, timestamped record of the terms at the moment of agreement, held by both parties and retrievable by either. Delivery is then checked against one agreed version rather than two.

Mistake 9: Measuring with the platform's own reporting only

Why it happens: The agent and its platform report their own results quickly and in detail, so teams accept them by default.

What it costs: Self-reported results are not independent evidence. Evgeny Popov wrote in AdExchanger that "the agent's own confidence score is not independent validation", and that AI "can interpret data, it cannot vouch for it".

The fix: Agree the evidence standard and the measurement source before the first bid, and use independent verification or a causal test where the spend justifies it. Amy Porter of RPA told Digiday that relying too heavily on AI for performance assessment without human oversight risks obscuring critical decisions.

Mistake 10: Launching without an exit or rollback plan

Why it happens: Pilots are planned for success, so nobody writes down what happens when the agent misbehaves mid-flight.

What it costs: In the same AdExchanger piece, Popov notes that a brand can halt new spend but cannot recover money already committed.

The fix: Before launch, decide who can pause the agent, what triggers a pause, how live deals pass back to a human trader and which terms survive a cancellation. Popov's suggested default when evidence arrives late is to pause rather than grant fresh funds.

Can AI agents be trusted to manage advertising budgets?

AI agents can be trusted with advertising budgets to the extent that their authority is written down, tested, checked and reversible. Each of the ten fixes above adds one of those properties, and none depends on a particular vendor or model.

The products Alkimi builds follow the same sequence. Its Gym lets buyers and sellers build their own agents, set the rules those agents work to and spar against counterparty agents before any real money moves. Its Marketplace is neutral, works within IAB Tech Lab's AAMP framework today and is being built to be AdCP compatible. It runs AAMP identity checks, keeps human approval at the deal level and writes a DealSheet at agreement that both parties hold and either can retrieve independently.

Frequently asked questions

Can AI agents be trusted to manage advertising budgets?

Yes, within a written mandate, tested limits and approval thresholds that a named person enforces. Trust should grow with the agent's record rather than be granted in full on day one.

What are the risks of using AI agents for media buying?

The main risks are agents agreeing weak contract terms, spending outside the brand's intent, trading with unverified counterparties and keeping records that do not match the other side's. Most are organisational risks that a mandate, a benchmark and a shared deal record reduce.

How are brands using AI agents in their advertising?

Most brands use AI for analysis and content before they let it buy: the Digiday+ Research survey found 67% of marketers using AI in social media campaigns used it to analyse results and data. Where agents do buy, brands such as Bayer describe adding spend caps and human oversight.

What is the most common mistake brands make with AI agents?

Giving an agent a budget before writing down what it may do. Without a written mandate, every other control has nothing to enforce.

10 Mistakes Brands Make When They First Use AI Agents for Advertising