The rise of agentic AI in advertising marks a shift from passive recommendation engines to systems that can autonomously execute decisions based on data inputs. This evolution promises greater efficiency, especially in high-frequency, rule-driven environments like programmatic buying, where speed and consistency are prized. Yet, as the technology matures, a fundamental tension emerges: while agents can optimise toward a given key performance indicator, they lack the innate ability to judge whether that KPI truly reflects meaningful business growth. This gap raises urgent questions for marketers about where to draw the line between automation and human judgment, and how to ensure that automated actions serve broader strategic goals rather than merely chasing superficial metrics.
At BidMatrix, VP of growth Mark Nedzelskii outlines a pragmatic view of where agentic AI proves most valuable today. The technology thrives in the “unglamorous middle” of campaign management—tasks such as translating briefs into initial media plans, verifying setup configurations, monitoring delivery pacing, flagging anomalous behaviour, generating performance summaries, and suggesting routine bid adjustments. These activities share common traits: they are repetitive, governed by clear rules, and benefit from relentless consistency rather than creative inspiration. By offloading such workloads to AI, teams can accelerate operational cycles and reduce manual errors, freeing human talent to focus on higher-order challenges that demand contextual awareness and strategic foresight.
However, the industry’s enthusiasm sometimes outpaces reality when it comes to assigning strategic responsibilities to agentic systems. An agent might declare one ad placement superior simply because it yields a lower cost per acquisition, yet remain oblivious to downstream consequences such as user churn, failed identity verification, lack of monetisation, or brand safety concerns. As Nedzelskii aptly puts it, an agent can optimise the map it is given, but it cannot determine whether the business has selected the right destination. This limitation underscores the necessity of human oversight: without the capacity to interpret signals beyond immediate conversion data, agents risk delivering numerically sound outcomes that are commercially or reputationally detrimental.
The concept of a “value ladder” offers a useful framework for understanding why human strategists remain indispensable. In a finance app, for instance, the journey from a low-cost registration to a profitable customer involves multiple stages: registration, KYC verification, first deposit, day‑30 activity, and ultimately positive margin contribution. An agent focused solely on minimising registration cost may aggressively scale that top‑funnel event, inadvertently flooding the pipeline with users who never progress beyond sign‑up. Only a human strategist can define which intermediate actions constitute real value, adjust the ladder as business priorities shift, and recognise when optimising one metric undermines another further downstream.
Trust in agentic AI, according to Nedzelskii, should not be based on the advertising channel alone but on the robustness of the system’s feedback loop. Environments where media exposure can be tightly coupled to verifiable post‑install events—such as registrations, purchases, or subscriptions via a mobile measurement partner (MMP)—offer a stronger foundation for autonomous optimisation. These closed‑loop signals provide richer, more actionable data than isolated clicks, enabling the agent to learn from actual business outcomes. Nevertheless, even in‑app settings present challenges: delayed postbacks, inconsistent event definitions, fraud vulnerabilities, privacy‑driven data limitations, and opaque supply paths can erode confidence and introduce noise into the learning process.
To mitigate these risks, BidMatrix has invested in greater control over its in-app inventory through tools like the MatrixSDK layer. By owning more of the supply chain, the company gains clearer visibility into traffic sources, applies uniform quality controls, and secures event‑level measurement fidelity. Such investments tighten the feedback loop, making it safer to delegate more autonomy to agents while preserving the ability to audit and reverse decisions when needed. The underlying principle remains: autonomy should be granted only when the system’s inputs, actions, and limitations are transparent, measurable, and governable.
The dynamics differ when looking beyond in‑app environments. Open‑web optimisation often relies on a blend of contextual cues and conversion signals that may be less direct or more fragmented, requiring sophisticated modelling to connect media exposure to business results. Connected TV (CTV) introduces another layer of complexity, with performance signals frequently spanning multiple devices and maturing over longer time horizons. In these settings, the agent’s ability to infer causality is further constrained, reinforcing the need for human analysts to interpret trends, account for external influences such as seasonality or competitive moves, and decide whether observed performance warrants scaling, investigation, or restraint.
The practical question of when to grant an agent greater autonomy hinges on governance frameworks that move beyond theoretical guidelines into operational reality. An agent can successfully hit its programmed KPI while simultaneously inflicting harm—for example, by purchasing inventory that damages brand suitability, overexposing a specific audience segment, or amplifying a creative message that resonates for the wrong reasons. Consider a premium finance app that acquires its cheapest registrations from low‑quality sources exhibiting anomalously fast conversion times. While the agent celebrates a low cost per acquisition, a human strategist must interrogate whether those users clear KYC checks, go on to make deposits, remain active over time, and align with the brand’s risk tolerance. As the article warns, “Low CPA is not a moral defence.”
Effective governance requires scrutiny at three levels: the transparency of the agent’s decision‑making process, the alignment of its objectives with broader business goals, and the presence of mechanisms to detect and correct misaligned behaviour. Direct supply visibility, robust anti‑fraud controls, and post‑install optimisation tools help sharpen the feedback loop, but they do not eliminate the ultimate accountability of human stewards. Agents should operate strictly within predefined boundaries—recommending and executing actions, but never assuming the roles of judge, jury, or appellate authority over their own outcomes.
As automation handles more of the mechanical aspects of campaign management, the strategist’s role is evolving from platform operator to architect of decision‑making systems. Less time will be spent copying settings between dashboards or manually adjusting bids; instead, effort will shift toward articulating clear business objectives, structuring data feeds for AI consumption, establishing guardrails that define acceptable behaviour, interpreting exceptions that fall outside learned patterns, and connecting media performance to tangible results such as revenue, customer lifetime value, or brand equity. This transition demands a new skill set: fluency in data literacy, comfort with AI‑augmented workflows, and the ability to translate ambiguous business goals into machine‑readable instructions.
Crafting effective briefs for agentic AI is becoming a core competency for modern planners. Vague aspirations like “Find me more users” provide insufficient guidance and leave too much room for misinterpretation. A well‑structured brief, by contrast, specifies the desired outcome, applies measurable constraints, and delineates exclusions—for example, “Increase first deposits in Germany while maintaining a D30 payback period below a defined threshold and blocking traffic from these identified low‑quality sources.” Such precision gives the agent a clear optimisation target and the necessary boundaries to operate safely, reducing the likelihood of unintended consequences while still harnessing the system’s speed and scalability.
The ongoing collaboration between humans and agentic AI points to a future where the greatest competitive advantage lies not in outperforming the machine on speed, but in mastering the art of strategic direction. Marketers should invest in developing three key capabilities: first, the ability to design and refine value ladders that capture the full customer journey; second, expertise in building and auditing transparent, reversible AI workflows; and third, fluency in translating business intent into precise, constraint‑driven briefs that guide autonomous systems toward meaningful outcomes. By embracing this complementary partnership—letting AI handle the «how» while humans steward the «why»—organisations can move beyond superficial metric chasing and drive growth that is both efficient and genuinely valuable.