Serhii Shchelkov, ad tech expert, looks at why buyers should  beware of paying a premium for the AI label, without getting genuine innovation…

"Any sufficiently advanced technology is indistinguishable from magic," - Arthur C. Clarke 

The ad tech industry took that as a product roadmap. 

And the numbers make the problem concrete. In the last three years, "AI-powered" became the default label for any platform feature that operates without a human clicking a button. According to Gartner's 2025 Hype Cycle, AI remains one of the most over-marketed terms across enterprise software, and ad tech is no exception. Bid adjustments. Creative rotation. Pacing rules. Frequency caps. Vendor marketing says that all of it is now artificial intelligence. But the question is – is it really?

A February 2025 MMC Ventures survey of 1,200 software companies found 40% of self-described "AI-first" vendors had no machine-learning code running in production. 

To be clear: Some of it genuinely is AI/ML. The problem is that much of it is not, and the two look identical from the outside.

The AI label is a pricing mechanism

The industry's obsession with the AI label is not primarily a technical phenomenon. In most cases, it is a commercial one.

Simply put, AI is trendy. And in a market where buyers associate the label with sophistication and premium value, vendors have strong commercial incentives to apply it broadly. Even without genuine AI/ML development behind the scenes, the term gets used freely. Opaque optimisation algorithms get packaged as "AI-based," black-box logic gets marketed as machine learning, and the buyer has no easy way to tell the difference.

Behind most of these labels sits deterministic automation: 

• Rules fire when conditions are met. 
• Budgets shift when performance crosses a threshold. 
• Caps apply when a counter hits a limit. None of this learns from data. 

These functions do not learn from data. They follow predefined instructions.

The distinction matters because you are often paying a premium tier for the AI label, not for a genuinely different capability. Two ad servers can run on functionally identical optimisation logic. One calls it "AI-powered." The other just calls it what it is. The first one charges more.

There is also a deeper problem with genuine black-box AI/ML, where it does appear. The issue is not that it performs poorly. The issue is that it removes your ability to understand what is happening and why.

A campaign underperforms.

• Does the problem lie in the creative?
• The targeting criteria?
• The floor price?
• The algorithm's behaviour in a specific inventory segment?

In a transparent, rule-based system, finding the cause is faster and testing a fix is straightforward. You isolate the variable, make the change, and verify the result. In a black-box AI/ML system, the model is the only party that knows what happened, and it offers no explanation. In ad serving, losing sight of why your campaigns behave the way they do is an operational risk with a direct revenue cost.

When you don't need ML

Let’s try a thought experiment. If a human AdOps manager can write the steps on a whiteboard, you probably don’t need machine learning to carry them out.

Campaign even pacing is one example. Distributing a budget evenly across a flight is a deterministic problem with a clear solution. A well-designed rule handles it reliably and cheaply. A learning model adds complexity, reduces auditability, and changes its behaviour as it accumulates data, which creates unpredictability in short-flight direct campaigns where consistency matters.

Campaign priorities and banner weights give publishers direct, auditable control over yield. At scale, ML can outperform static weight-setting in terms of raw yield. Whether that matters depends on one thing: whether you can see what the model decided and why.  

These tools let publishers decide what runs in which placements, how much direct demand is prioritised versus programmatic backfill, and how inventory is allocated among advertisers. That is the foundational layer of publisher control in an ad server. It is manual by design, and that is a feature, not a limitation.

For filtering underperforming placements and blocking by geography, device type, connection type, or domain, traditional rule-based approaches have real advantages. You can review the rules, confirm they fired, and adjust them immediately when requirements change. The trade-off with AI/ML filtering is real: It surfaces patterns a human would miss. The operational question is whether your vendor actually logs what the model changed, why it changed, and what the performance result was. AI filtering that writes structured decision logs is auditable. AI filtering that operates as a background process with no output record is not. The capability exists. The implementation varies significantly by platform. 

For publishers who need to be accountable to advertiser clients for specific delivery parameters, the ability to explain what happened and why is not optional.

NB: Wherever you need to be accountable for specific delivery parameters to an advertiser client, transparent rules are more valuable than learned behaviour. You can explain a rule. You cannot always explain what a model decided and why.

Where AI actually earns its place

There is a category of problems in ad serving where machine learning provides a genuine competitive edge. In other words, there are situations in which the signal space is too large, too dynamic, and too contextually variable for human analysis or for static rules to handle accurately.

Traffic quality and fraud detection are the best examples. Finding fake traffic among millions of daily views means checking bot traffic levels, groups of IP addresses, fake-click timing, use of emulator devices, data centre proxy use, unusual video-viewing patterns, and daily traffic cycles, while remembering that the same sign can mean different things depending on the situation. A steady 24-hour traffic pattern is suspicious for display ads. High IP-to-user ratios are normal for mobile CGNAT networks. Missing screen resolutions are expected in VAST video ads. No set of rules can handle all of this smoothly on a large scale.

This type of multi-signal, format-aware fraud analysis is where AI/ML genuinely outperforms alternatives, and it is increasingly available as a built-in capability in modern ad servers. Epom Ad Server, for instance, includes a Traffic Quality Fraud Analyser that runs daily automated scoring across placements, using AI-powered deep analysis to investigate flagged traffic and produce structured investigation reports with risk assessments and prioritised recommendations. The AI does the pattern recognition. The operator decides what action to take.

That balance is the key. AI manages tasks that are too complex for humans to handle on a large scale. Humans stay in control of the important decisions.

The practical question(s)

Before accepting any AI/ML claim from a vendor, two questions cut through the marketing quickly.

1. Can you explain what the AI/ML actually does? If the answer involves trained models, pattern recognition across large datasets, or the interpretation of contextual signals, then it is plausibly genuine. If it describes rules, thresholds, or automated triggers, that is automation with a more expensive name.
2. Can you audit the output? Can you see why a decision was made, test a change, and verify the result? If yes, you have a tool you can manage. If not, you have a system you hope performs correctly but can't confirm it.

The best ad servers in 2026 offer both: transparent optimisation controls for the decisions your team needs to own, and genuine AI/ML assistance for the problems that actually benefit from it. The risk is not in platforms that use AI/ML; it is in platforms where the AI/ML features being marketed are primarily a positioning tool rather than a functional differentiator. And where the opacity of those features makes it impossible to hold the system accountable when results fall short.

In ad serving, knowing what your stack is doing and why is not a preference. It is a business requirement.

Serhii Shchelkov is an ad tech expert with deep experience in programmatic infrastructure, ad server strategy, and traffic quality operations. He works with Epom Ad Server, a white-label ad-serving platform built for publishers, networks, and agencies, combining transparent, rule-based optimisation controls with AI/ML-powered fraud detection in a single stack.