Mark Nedzelskii, VP of growth at BidMatrix, shares where he trusts AI agents today, where he does not, and why better automation may actually make human strategy more important

Ad tech has spent years automating media buying. Agentic AI moves the line further: the software no longer just surfaces a recommendation. It can act on it.

That makes the practical questions more urgent. What can safely run without human approval? What still needs judgement? And who is accountable when an agent hits the KPI but gets the business outcome wrong?

We spoke to Mark Nedzelskii, VP of growth at BidMatrix. He chats to us about trust in AI, and why human strategy remains key.

Where has agentic AI already earned its place in programmatic, and where is the industry still overselling it?

Agentic AI has earned its place in the unglamorous middle of programmatic: turning briefs into first-pass plans, checking campaign setup, monitoring pacing, detecting anomalies, summarising performance, and recommending routine optimisations.

These are high-volume, rules-heavy tasks where consistency matters more than inspiration. At BidMatrix, we use AI in that spirit: to accelerate analysis and operational workflows, not to cosplay as an all-knowing media director.

Where the industry oversells agentic AI is strategy without context.

An agent might decide that one placement is "better" because it generates cheaper conversions. But it may not be understood that those users churn after three days, fail KYC, never deposit or come from inventory the brand would prefer to avoid.

"An agent can optimise the map it is given. It cannot decide whether the business has chosen the right destination."

The practical test is simple: can the agent show its inputs, actions and limitations? If not, it is a black box with a confident tone.

I trust agents most when the workflow is measurable, reversible and governed. The closer a decision gets to brand, budget or business strategy, the more human oversight it needs.

If an AI agent can plan, bid and optimise a campaign with minimal input, what is left for a human strategist to do?

Plenty. Planning, bidding, and optimisation are only the mechanical layer of strategy.

A human still needs to determine what the campaign is trying to achieve, which signals represent value, which trade-offs are acceptable and when the numbers are technically correct but commercially misleading.

Consider a finance app. An agent may learn that registrations are inexpensive and scale them aggressively.

But registration is only the beginning of the value ladder:

Registration → KYC → First deposit → D30 activity → Positive margin

The strategist defines this value ladder and changes it when the business needs change.

Humans also bring causal judgement. Did performance improve because of the campaign, a promotion, seasonality or a product update? Should an unexpected pocket of performance be scaled, investigated or blocked?

The agent can become an excellent pilot, but the strategist still chooses the destination, the fuel budget and the emergency procedure.

In-app environments run on different signals from the web. Does that change how much you trust an agent to optimise autonomously?

I do not trust an agent based on the channel. I trust it based on the quality of its feedback loop.

In-app can provide a strong environment for autonomous optimisation because media exposure can be connected through an MMP integration to post-install events such as registration, purchase, deposit, subscription or retention. These signals are far more useful than an isolated click.

However, in-app also has its own traps: delayed postbacks, inconsistent event definitions, fraud, privacy restrictions and opaque supply paths.

This is one reason we've been investing in more direct in-app inventory, including our MatrixSDK layer. Clearer source visibility, quality controls and event-level measurement make it safer to give an agent more autonomy.

Open-web optimisation may depend on broader contextual and conversion signals. CTV is different again: its signals are often cross-device and slower to mature.

When should an agent get more autonomy?

• Strong signals
• Clear data provenance
• Reversible actions
• Transparent decision logs

When should a human stay close?

• Weak or delayed signals
• High-cost decisions
• Ambiguous causality
• Brand or compliance risks

What happens when an agent optimises correctly by the numbers but incorrectly for
the brand?

This is where governance stops being a slide and becomes an actual job.

An agent can hit its KPI while damaging the brand. It might buy unsuitable contexts, overexpose an audience, chase misleading conversions or push a creative angle that performs well for the wrong reasons.

Imagine a premium finance app receiving its cheapest registrations from low-quality inventory with suspiciously fast conversion times. Numerically, the agent may be delighted. A strategist should ask whether those users pass KYC, make deposits, remain active, and fit the brand’s risk profile.

"Low CPA is not a moral defence."

The problem needs to be caught at three levels:

• Before launch: Define suitability rules, event-quality standards, source exclusions,
  frequency caps and stop-loss thresholds.
• During delivery: Use anomaly detection, transparent logs and automatic escalation
  when activity moves outside the guardrails.
• After delivery: Evaluate performance through quality, retention, incrementality, fraud
  and brand context, not one heroic metric.

Direct supply visibility, anti-fraud controls and post-install optimisation help tighten that loop, but they do not remove human responsibility.

Agents should recommend and execute within defined boundaries. They should never be judge, jury, and appeals court.

As agentic tools become more capable, how will the strategist’s job change?

The strategist’s role will move from operating platforms to designing and supervising
decision systems.

Less time will be spent copying settings between dashboards. More will go into defining objectives, structuring data, setting guardrails, interpreting exceptions and connecting media performance to business results.

What should today’s media planners learn?

• Measurement: Attribution, incrementality, cohort analysis and LTV.
• Data literacy: Enough SQL or spreadsheet logic to challenge an answer that looks
  convincing.
• Agent architecture: How agents use data, tools, and permissions, and the difference
  between generating a summary and moving budget.
• Governance: Audit trails, approval thresholds, brand-safety controls, and rollback
  rules.
• Creative and commercial judgement: The areas that remain difficult to encode.

Planners should also learn to write machine-readable briefs.

"Find me more users" is not a strategy.

"Increase first deposits in Germany while keeping D30 payback below a defined threshold and excluding these sources" gives an agent a clear objective and constraints.

We can already see this shift: agents can prepare analysis, surface patterns and remove repetitive work, while people handle client context, quality calls, and uncomfortable questions.

Do not compete with the agent on speed. Become the person who knows what it should do, what it must never do, and whether it worked.