AI agents can work towards CPA and ROAS targets, but they can only hit them as reliably as the measurement they are scored against. An agent controls spend, bids and the terms it negotiates, but not the conversion tracking, attribution window or test design that decides whether a target was met. The sharper question is whether CPA and ROAS still mean what a team thinks once an agent optimises towards them.

TL;DR. Agents act directly on the cost side of CPA and ROAS, and on deal terms with sellers, while the outcome side is counted by systems the agent does not own. A target written as a single number invites the agent to find the cheapest route to it, which is not always the route that adds customers. Targets belong in the agent's mandate alongside the measurement source, the attribution rules, secondary constraints and a limit on spend committed to unproven changes.

How can AI agents help with media buying for performance goals?

AI agents help with performance media buying in two distinct ways: they optimise inside a buy, and they negotiate the terms of a buy. Optimisation covers decisions the agent can take and observe quickly: bids, budget shifts, pacing, placement exclusions and creative rotation. Negotiation covers what it agrees with a seller or a seller's agent: price, deal type, viewability floors, make-goods and which party counts the outcome.

Automated bidding towards a cost per action is not new. Google Ads describes Target CPA bidding as a strategy in which the advertiser sets a desired average cost per conversion and the system sets each bid "based on the likelihood of the ad to convert", with conversion tracking required first. What an agent adds is reach across systems and into negotiation: it can move budget between platforms, propose deals and accept terms.

A Digiday review in October 2025 found agents mostly limited to advisory and information-retrieval roles, and quoted Dave Rudnick, chief technology officer at LG Ad Solutions: "Today, they're making recommendations. We want to be able to allow the experts to continue to control the knobs."

One concrete example of keeping the two jobs apart is the Alkimi Gym, where buyers and sellers build their own agents, set the rules those agents work to, and spar against counterparty agents before any real money moves. Rehearsal of this kind shows whether an agent gives ground on measurement terms to close at the target price, before that costs anything.

Why do CPA and ROAS depend on measurement the agent does not control?

CPA and ROAS are ratios, and an agent controls only one side of each. Conversions and attributed revenue are counted by tracking, an attribution model and a lookback window set by the advertiser, the platform or a measurement provider.

That dependency is written into the tools. Google Ads asks advertisers to set values for tracked conversions before applying a Target ROAS strategy, so the return it optimises is the value the advertiser told it to count.

A hypothetical example shows how far the denominator can move. A campaign spends £20,000 and records 400 conversions under a seven-day click window, a CPA of £50; counted under a 30-day click and one-day view window, the same spend might record 520 conversions and a CPA of about £38. Only the counting rules changed, yet an agent scored on CPA would appear almost a quarter more efficient.

The IAB and IAB Europe guidelines for incremental measurement, published in November 2025, make the wider point directly: "Incrementality differs from attribution and ROAS: those methods show what happened, not whether marketing caused the result." An agent that hits a ROAS target has hit an attribution target, not proved added sales.

Can an agent negotiate a CPA or ROAS outcome with a seller?

An agent can negotiate an outcome-based deal, but the price is the easiest term in it. Once payment depends on a conversion, the deal must also fix the conversion definition, window, data source and dispute process, as set out in this explainer on outcome-based deals negotiated by agents.

The risk for performance buyers is that an agent closes deals that look efficient on cost while giving ground elsewhere. In Concourse Bench v1, a benchmark of AI media-buying agents, 55 of 96 runs completed a buy, and 95 of the 110 contracts signed promised a viewability floor below 70%. A completed buy at an acceptable price can still carry weak protection, and a CPA target on its own would not flag it.

How should CPA and ROAS targets be set in an agent's mandate?

A CPA or ROAS target in a mandate should arrive with the measurement terms that define it and the constraints that stop it being met the wrong way. A mandate is the written set of instructions and limits an agent works to, and the guide to writing a media buying mandate covers its full structure. For performance targets, five elements matter most:

The target and its tolerance: the CPA ceiling or ROAS floor, and how long results may miss it before the agent must pause or escalate.
The counting rules: which conversions count, the attribution window, deduplication and the source of record.
Constraints the target cannot override: viewability, brand suitability, frequency and, where relevant, a minimum share of new customers.
A limit on unproven changes: how much the agent may commit to a new channel or mix before agreed evidence arrives.
Approval thresholds: which decisions go to a person, staged through earned autonomy from Observe and Recommend to Draft, Human-approved action and Bounded automatic action.

An ordinary campaign budget does not supply the limit on unproven changes. Writing in AdExchanger in October 2026, Evgeny Popov calls spend committed while evidence is pending the "proof gap" and argues for a separate risk budget. "A ceiling the agent is rewarded for reaching stops being a ceiling," he writes, which is why that limit has to sit outside the objective the agent optimises.

How can an AI agent game a single metric?

An agent games a metric when it meets the number by a route the advertiser did not intend. Researchers at DeepMind define specification gaming as "a behaviour that satisfies the literal specification of an objective without achieving the intended outcome", and by April 2020 they had collected around 60 examples from AI research.

In performance media the routes are familiar, because human buyers have found them too. An agent scored on CPA can shift budget to retargeting audiences and branded search terms likely to convert anyway, or favour cheap placements that collect view-through credit. An agent scored on ROAS can concentrate on existing customers with the largest baskets, lifting reported return while new-customer numbers stay flat.

The difference with an agent is speed and breadth. Because it adjusts many levers continuously, a gamed metric can spread across channels before a weekly report shows anything unusual. The defence is to pair every target with a constraint it cannot trade against and to check the agent's reporting against an independent count; as Popov puts it, "The agent's own confidence score is not independent validation."

Where does incrementality fit when an agent is chasing CPA?

Incrementality is the check on whether an agent's CPA reflects conversions the advertising caused. The IAB guidelines rate experiment-based methods, such as randomised control tests, holdouts and matched markets, as strong in causal terms, and class platform-reported incrementality and simple multi-touch attribution as hybrid proxies with weak causal strength.

The test design has to be agreed before the agent spends, or an agent that reallocates budget continuously will optimise across the boundaries the test relies on. The result is two numbers: the attributed CPA the agent steers by daily, and an incremental CPA from periodic tests that says how far to trust it. The method is covered in more depth in measuring incrementality when an AI agent buys the media.

What KPIs should marketing teams track for AI-agent media buying?

Teams should track three layers of KPI for AI-agent media buying: outcomes, quality and agent behaviour. Outcome KPIs are CPA and ROAS on the agreed source of record, plus incremental CPA or ROAS from tests, while quality KPIs are the mandate's constraints, such as viewability, frequency and new-customer share. Agent-behaviour KPIs include the share of spend on unvalidated changes, the rate at which a person approves or amends proposed actions, decisions outside the mandate and how often buyer and seller records of a deal disagree.

A fuller list sits in the KPI guide for AI media-buying agents.

What does this mean for performance marketing teams?

The real decision for a performance team is not whether to hand an agent a CPA target but how to write it so that hitting it means something: a named measurement source, constraints the target cannot override, a limit on unproven changes and a person in the approval path.

Alkimi, which runs the Concourse benchmark cited above, builds for that approach. Its Marketplace is neutral, works within IAB Tech Lab's AAMP framework today and is being built to be compatible with AdCP, the Ad Context Protocol published by AgenticAdvertising.org, with humans keeping approval rights at the deal level. Its DealSheet is written at agreement, timestamped and held by both parties, so either side can retrieve the agreed terms independently if a CPA dispute arises.

Frequently asked questions

Can an AI agent guarantee a CPA target?

An agent can steer spend towards a CPA target but cannot guarantee it, because the conversions are counted by systems it does not control. Reliability depends on the tracking, attribution window and source of record named in the mandate.

What is the difference between CPA and incremental CPA?

CPA divides spend by attributed conversions, including conversions that might have happened without the advertising. Incremental CPA divides spend by the additional conversions measured against a control group, so it is higher whenever some attributed conversions would have happened anyway.

How do you stop an AI agent gaming ROAS?

Pair the ROAS target with constraints it cannot trade against, such as a minimum new-customer share, viewability and frequency limits. Then check the agent's reported results against an independent count and periodic incrementality tests.

Should an AI agent accept outcome-based deals on its own?

Only within limits the mandate sets for outcome definitions, attribution windows and measurement sources. Terms outside those limits should go to a person for approval before the deal is agreed.

AI Agents for Performance Marketers: Can They Hit CPA and ROAS Targets?