AI is now the dominant narrative across advertising. But long before generative AI entered every conversation, machine learning was already a cornerstone of programmatic infrastructure.
As Tim Koschella, who spent more than a decade building and scaling ad tech companies and now leads Kayzen, he has witnessed programmatic evolution from the earliest automation systems to today’s AI-powered ecosystems.
In this conversation, he shares his perspective on where AI really adds value, where it’s overrated, and why human responsibility must remain central as systems become more intelligent.
What is actually new and what is not?
While AI dominates the headlines today, Tim grounds the conversation in history.
“Let’s look at the facts first: AI has been at the heart of ad technology for over a decade already. Advertising was one of the earliest markets where AI was applied at scale.”
What most people now refer to as AI has long been part of programmatic advertising, where machine learning systems operate at scale. Based on structured data, these systems predict whether a user is likely to click, convert, or buy, and return those predictions within milliseconds.
Generative AI changes the interface layer. “It’s more flexible. It handles unstructured data. It tolerates incomplete input. It’s a much better partner for humans.” But that change doesn’t make it universally better. “Speed is critical in RTB. Inference calls must run within milliseconds. Generative AI often takes seconds. That time lag makes machine-to-machine communication less useful.”
In his view, utility is practical rather than philosophical. “Anything that improves measurable outcomes for advertisers or helps humans do their jobs faster and better is useful. As long as you put the customer’s problem first and find ways that AI can solve it, it’s a path worth pursuing.”
Why is performance still difficult to explain?
Even though AI has been integrated into advertising infrastructure for years, many teams still struggle to clearly explain performance outcomes. The problem, he argues, is not the sophistication of the model, but the definition. He emphasizes that performance must start with a deep understanding of the advertiser’s business and the goals the company is trying to achieve with its marketing investments.
“Inherently, performance remains vague until it is clearly defined in context because it is used in so many different ways. A brand agency managing Coca-Cola’s mobile advertising budget in Europe and a mobile RPG game looking to grow in Japan will have very different perspectives on how performance is defined and measured.”
Conversations about optimization often stop at metric labels rather than the business realities behind them. “Take ROAS as an example. When someone says they are optimizing ROAS, I immediately ask: What ROAS? D7? D30? Lifetime? How is your product generating revenue? Is your revenue model driven by outliers? How do you measure incrementality and attribution? What if they point in different directions?”
This complexity is not purely technical. It’s strategic. “We need smart experts to understand this complexity. AI can help, but humans must remain at the center of decision-making.”
Pressure to have an “AI story”
If AI has been integrated into advertising systems for years, why does the current moment feel so amplified? Much of this, Tim suggests, is due to narrative pressure, particularly from investors and capital markets.
“The pressure to deploy AI everywhere is huge, especially for companies backed by large investors or publicly traded companies. Every CEO is expected to have an AI equity story.”
He doesn’t dispute that AI can improve most businesses. What concerns him is how it is structured. “AI doesn’t have to be the fabric that connects everything in your business, like the tissues in your body. In many cases, AI is just a useful tool to improve efficiency and quality.”
When AI becomes an identity rather than a feature, expectations can quickly outstrip reality. “If intentionally created expectations are not met in the future, a feeling of paranoia may emerge, and companies developing applications of AI that actually create value may need to make an extra effort to explain themselves.”
He is not against the introduction of AI. What he is concerned about is the expansion of the story beyond its practical value. He points out that AI is often a tool to improve efficiency and quality, not reinvent the way businesses operate. Therefore, the differentiator is not how aggressively you deploy AI, but whether it solves real problems for your customers and delivers measurable value.
Pre-technical principles
This concern leads to a deeper question. As AI becomes more integrated into the way businesses operate, what should guide its use?
“Programmatic advertising has many data signals that AI can process, interpret, and act on. But the signals also contain noise that can lead to false or misleading interpretations. AI can be easily misunderstood.”
In the global programmatic advertising market, the consequences of miscommunication can be severe. “If you make the wrong move, you can incur enough losses to bankrupt an entire company overnight. It’s similar to financial markets. If you can automate financial transactions without imposing limits or systems of checks and balances, you can blow everything up in one day.”
Therefore, human surveillance is structural rather than optional. “AI will never replace the critical thinking and decision-making that humans bring. Ultimately, we humans are responsible and accountable for the actions of the AI we use.”
However, there is a difference between interpretation and execution. “In fact, very few people are good at interpreting large, complex datasets, so the chances of a human being ‘wrong’ are much higher than those of an AI.” But when automation takes direct control of capital without guardrails, the risk profile changes. “If AI gets it wrong, it can spend an unapproved $1 million budget in the wrong place faster than you can blink. Humans are unlikely to make such high-impact, critical mistakes.”
When AI is underutilized
Where should AI create the most value before spending money?
“In programmatic, it’s obviously the campaign planning stage. AI is very active right now in the optimization stage. But AI is rarely used in the pre-campaign planning stage, where strategy is decided and expectations are set.”
Intelligence has focused disproportionately on the actual campaign execution, but the strategy stage, where hypotheses are formed and expectations are set, remains relatively undeveloped. Before execution begins, AI helps teams model scenarios and systematically leverage structured first-party signals. “Before the initial funds are spent, the most valuable assets are the advertiser’s first-party data, data about existing users, funnel metrics, seed audiences of lookalike users, and relevant segmentation inputs. Here, using AI systematically can add a lot of value without spending money on campaigns.”
Once a campaign actually launches, intelligence must extend beyond acceleration and efficiency to support understanding. “AI should help teams learn rather than just react. It should surface anomalies and expose trade-offs.” As systems improve in functionality, clearly defined guardrails become even more important. “Humans must act as the final decision layer for important decisions made outside of defined boundaries.”
He illustrates this with a practical scenario: a $20,000 per day campaign with a $100 CPA goal. If a piece of inventory suddenly shows a CPA of $10, the system will aggressively scale in to it. However, if defective conversion data feeds that signal, the efficiency logic can amplify the mistake before it is detected. “Humans, with good judgment, may understand that something is wrong and decide to pause. Boundaries need to be carefully defined. That is very difficult, but essential.”
The difference is not between human and machine capabilities, but between automation and accountability.
Result: Empowerment
The conversation eventually moves from infrastructure to mindset. Tim challenges one of the industry’s core assumptions. “One of the common myths is that you need a large team of data scientists to compete with the established incumbents. Even more important is the quality of your team members and how they work together to establish a deep understanding of the problem you need to solve.” Collaboration between disciplines such as data science, engineering, and business stakeholders is more important than size alone.
He is equally wary of exaggerated technical mystique. “Beware of small businesses that claim that a clever team of scientists and former NASA engineers has built a solution that no one else has thought of. This is probably a great sales pitch.”
If AI is implemented carefully, marketers shouldn’t feel like they’ve been replaced. They should feel able to do more with clarity and control. “Now, our team of five can easily manage a $500 million annual marketing budget while managing real-time campaign flight adjustments and optimization decisions every day.”
For Tim, intelligence must expand capabilities without compromising responsibility. “AI should enable decision-making, never replace judgment.”
