I’ve started a project to talk with the companies driving AI change in digital media and ad tech, especially the ones building for publishers and other ad tech providers.
This time I spoke with Nicolas “Nico” Schueller, CEO and co-founder of Adomik. Nico and I first met more than a decade ago, when I was at Yahoo and Adomik was helping publishers optimize price floors on AppNexus. Since then, header bidding pushed Adomik into revenue management, the company built a market data set, and in January it launched JP, its AI platform. Today about 95% of Adomik’s development bandwidth goes to JP.
1. The Market Data Set: Accuracy Over Completeness
My first question was whether the market data was directional or credible. Too often, I see people take a small slice of the market and extrapolate.
Nico was direct about the limits. In the US, the data covers roughly $5.5 billion in annual advertiser spend, about 15 to 17% of the open web. He would never encourage you to build strong analysis of the whole market from it.
What he does claim is accuracy. The data comes straight from SSP reporting. It isn’t extrapolated or calculated. If you want to know what Geico pays for in-stream video in the US when it buys through Google, that’s the actual number. That’s useful when you’re preparing for a sales meeting, studying a brand’s buying strategy, or making the case to keep a client on PMPs rather than curation.
Privacy protections are built in: no individual business relationship can be reverse-engineered, no publisher can make up more than 15% of the revenue in any view, and nothing on the sell side is visible beyond the SSP.
As Nico put it, “It’s a whole trade-off between completeness and accuracy.”
2. JP Is Infrastructure, Not an Agent
When most people hear “AI agent,” they think chatbot. Nico’s answer: JP is not an agent, and it is not a chatbot. It’s infrastructure.
The heavy lifting was the data. LLMs cope well with unstructured information, but transactional data is different. Adomik spent almost nine months building JP’s semantic model so an LLM could interpret ad revenue data correctly.
His example: give a general-purpose model a few SSP reports and ask for your revenue with one of the holdcos. The model knows what a holdco is. Mapping that to your actual data is another matter. “At the beginning we tried this,” he said. “This does not end well.” The same goes for transaction types, CTV versus display, and app versus web.
That resonated with me. Whether you’re a vendor or a consultant, the expectation is to be accurate and reliable. Leave room for interpretation and the client eventually comes back with, “Your data isn’t saying the same thing.”
On top of the semantic model, Adomik has built Revenue Analyst, Market Analyst, and agents that combine Adomik data with Google Analytics. It also offers JP as an MCP server, so clients can build their own agents.
Two changes are coming. In November 2026, Adomik expects AI-assisted onboarding to cut onboarding time by about 30%. In December, a new version of JP adds role templates such as JP for AdOps, JP for Sales and JP for Yield, which start by asking each user about their objectives and targets.
3. Can Publishers Trust Their Data Again?
I asked whether we’re getting back to the confidence publishers had when they ran one ad server and sold direct.
Nico’s honest answer: it’s still a day-to-day challenge. Curation, multiple flavors of PMPs, header bidding, server-side and wrappers have all added complexity. Add a new SSP for testing with three seats, miss sharing one with Adomik, and you have a gap. Adomik checks configurations against ads.txt, but not every ads.txt file is maintained with the same rigor.
His conclusion: it’s teamwork between the client and the vendor.
4. Guardrails Start With What You Feed the Agent
When I asked about keeping a human in the loop, Nico started somewhere I didn’t expect: how the agent gets its data.
Should it call another agent, run a SQL query, or execute a Python routine? If you know up front whether the agent is reasoning on deterministic or probabilistic data, you avoid much of the checking you’d otherwise need.
For changes to an external system, such as an SSP setting, Adomik recommends a human in the loop 100% of the time. But it has built change logs that track what was changed by a person and what was changed by AI with human approval, which sets the stage for more autonomy later.
5. The Contrarian View: Enterprise AI Is Enterprise Software
I close each interview by asking what the guest believes that most people don’t.
Nico’s answer: enterprise AI should remain in enterprise software companies. He acknowledged his bias given where Adomik sits in the value chain, and he praises people who build their own tools. But developing AI software is still developing software. Who maintains it? What happens when the person who built it leaves? Where is the output stored? If the answer is a DuckDB file on someone’s Mac, that won’t fly very far.
The Takeaway: Four Questions Before You Trust AI With Your Revenue Data
Whether you buy it or build it, how do you know an AI tool will hold up?
- What is it reasoning on? Know whether the agent works from deterministic data (SQL, Python) or probabilistic output.
- Does it speak your language? A generic model knows what a holdco is. It doesn’t know how your data represents one.
- Who approves changes? For anything touching an external system, a human should sign off, and there should be a log.
- Who owns it next year? Maintenance, security and continuity don’t disappear because the code was easy to write.
If you can’t answer all four, you have a demo, not a tool.