For several years, AI in advertising has followed a familiar model: it reviews campaign data, identifies weak performance and recommends a course of action. Yet the practical steps still fall to the advertiser, who must open the platform, find the relevant campaign and make the adjustment.
That model is beginning to evolve as more sophisticated platforms allow AI agents to make campaign changes themselves. An advertiser might instruct an agent to stop every push campaign with a click-through rate below 0.3% and increase bids by 15% for the three strongest performers. The system can then carry out both actions without a dashboard or a sequence of manual inputs.
By taking on this operational work, agents can free advertisers to concentrate on strategy, experimentation and growth. Leading AdTech providers are consequently enhancing their AI products with agents, or “campaign co-pilots”, capable of supporting campaign creation, editing, targeting, budgeting, scheduling, creative management and reporting through one conversational interface instead of multiple setup screens.
These systems are most effective, however, when advertisers equip them with high-quality data, sufficient context and a clear strategy from which to learn and operate. Without meaningful human direction and supervision, complications can arise, undermining the purpose of using AI and creating tangible consequences for the wider industry.
The key questions, then, are how advertisers can collaborate with AI effectively and where autonomous action should give way to human judgement.
What changes when AI moves from advice to action
The fundamental distinction between AI that advises and AI that acts lies in where responsibility ultimately rests.
When a person assesses a recommendation and chooses to implement it, ownership of the decision is clear, and the reasoning can usually be reconstructed if the outcome is poor. Once an agent acts independently, that line of accountability becomes harder to follow. A broad bid increase may appear justified by the available figures, but those figures cannot account for a competitor announcement released that morning, an internal brief that has just shifted campaign priorities or a brand issue unfolding behind the scenes. A human operator could recognise those developments; an agent working from the previous night’s dataset could not.
This does not mean execution-level AI should be rejected. Rather, it highlights the need to define precisely where autonomous action is appropriate and where it is not.
Where autonomy adds value
Some campaign activities are natural candidates for autonomous execution, including stopping campaigns once they reach their spending limits and producing performance summaries. These are operational functions: the strategic choice has already been made, leaving the agent to implement it.
Evidence also indicates that agents deliver markedly stronger results when they receive adequate context. Tests of agentic campaign configurations have found that advertisers who provide detailed information about objectives, funnel design and target cost per acquisition achieve substantially better outcomes than those who offer only brief instructions. In some instances, conversion performance differs by more than 100%. Put simply, the more clearly an agent understands the intended result, the more effectively it can pursue it.
Rethinking access and permissions
Deciding what agents may do is only part of the equation; organisations must also determine how those systems obtain access.
The sector is increasingly adopting integrations based on the Model Context Protocol (MCP), which enables external AI agents to connect directly with advertising platform APIs. Instead of entering a platform’s proprietary interface, an advertiser can remain within their preferred AI environment while the agent calls on the platform whenever an action is required.
In these arrangements, access is generally granted through API tokens rather than the advertiser’s account credentials. A token is distinct from the user’s login, can be restricted to defined permissions and can be withdrawn immediately when necessary. That is a sensible approach, but anyone who possesses the token also gains the access attached to it. Its scope therefore needs careful consideration before the token is distributed across a team or provided to an external party.
Greater interoperability is likely the right direction for the industry. Even so, there is an important distinction between an agent that can inspect campaign information and one empowered to alter it. Governance for execution rights must be more rigorous than governance for reporting access, because the concern is no longer confined to what the system can view; it extends to what the system can change.
Why human oversight remains essential
In practice, the boundary between autonomous execution and human supervision depends on how clearly the strategy governing the automation has been articulated.
Most campaigns still depend on a person to resolve ambiguities. For example, they may recognise that a threshold no longer serves the campaign’s true objective or that an external development should alter activity within the platform. If the human is removed without those gaps being addressed, ambiguity quickly becomes error.
Before delegating execution to an agent, advertisers should work through several fundamental questions. Which conversion events carry genuine business value, rather than simply being convenient to measure? Which adjustments are routine, and which require human approval? Under what circumstances should the agent escalate an issue instead of taking action?
None of these questions is especially complex, but traditional campaign setups have rarely required explicit answers because a person has always been available to exercise judgement at the point of decision. Agentic AI turns that tacit judgement into defined rules. Establishing those rules before automation begins is considerably easier than correcting the consequences afterwards.
Defining the next era of AI-powered advertising
The industry has already established that AI agents are capable of taking action. The more difficult task is deciding whether advertisers, platforms and partners are prepared to specify the boundaries of that authority. Execution-level AI has the potential to make campaign management quicker and reduce manual effort, but its success depends on clear rules from the outset.
Advertisers must set firm limits, supply agents with sufficient context and retain human judgement for decisions that still demand it. The objective is not to surrender control simply because the technology permits it. It is to make a deliberate choice about where automation creates real value and where a person still needs to decide.
