By Mike Conway, Field CTO, AdTech
A few weeks ago I had dinner with a small group of AdTech customers and prospects. While it was off the record and I can’t share company names or direct quotes here, some of what came up is worth passing along, especially with our team heading into Programmatic I/O, DMEXCO and Advertising Week this fall, where I expect a lot of these same conversations to keep going.
Change is the one constant
I opened the dinner talking about how the industry seems to evolve every three years. In 2018, it was GDPR. Then cookie deprecation. Now it’s agentic AI. Different problem every time, same underlying lesson: sooner or later, the architecture underneath has to change too.
Most of the industry conversation right now is about adding AI to what already exists. In some cases, people are talking about what it looks like to build with AI in mind from day one.
Two stories worth repeating
One customer at the table told us about moving her file and data delivery processing onto OcientAIQ™. Faster, cheaper, and different enough architecturally that she’d already redirected the cloud spend she’d have otherwise committed to scaling the legacy system. So the win landed twice for her: once on performance, once on budget.
Then there was a CTO I’ve now worked with across two different companies. He was pretty direct that the tech wasn’t really what sold him. It was that we’d sit down and work through the actual use cases together, and teams could experiment before anything went near production. Almost as an aside, he mentioned OcientAIQ could probably replace a handful of point solutions he was running like Pinot, a vector database and a graph database.
Full stack, live data, real numbers
We’re at the point where we’d rather show this than describe it, so we ran a live demo during dinner.
The platform inferred schema automatically as data came in, typing IP addresses natively instead of storing them as strings, which helps both compression and query speed. A “Data Advisor” agent watched the incoming data (cardinality, nulls, ranges) and suggested schema changes on the fly. Every SQL statement an agent ran got logged and was fully traceable.
Then there’s the scale piece. The demo environment held 50.6 billion records across 4.41 petabytes. Queries came back in about 117 milliseconds. Joins that would normally mean an overnight batch job ran in seconds instead, and enrichment happened as data landed rather than waiting for someone to write a query against it. Agents can also run on a recurring schedule, similar to how you’d set up detection jobs.
The reaction that really landed, though, was around cost. Because the heavy lifting happens inside the engine rather than an LLM stitching results together across a semantic layer, token costs come down a lot. In one benchmark we saw roughly 100x lower cost than a competitor we’re not naming doing comparable work, with 10x being more typical. Permissions work the way you’d expect too: a campaign manager and a trafficking analyst get different levels of access, matched to how AdTech teams are already organized around buying, planning, and measurement.
What people kept coming back to
Once the demo wrapped and the table conversation opened back up, a few things kept resurfacing. Getting an answer in 117 milliseconds without sampling or an overnight wait changes what you can actually build. Enrichment at ingestion came up a lot too, since it removes a pipeline most teams currently have to build and maintain themselves.
Cost was probably the biggest one. “AI gets expensive at scale” is an objection almost everyone at the table said they’d heard internally, and it’s a fair one, until the database does the work instead of the LLM. And then there’s auditability: being able to see exactly what an agent did and why.
Worth noting the savings aren’t just an AI story either. At multi-petabyte scale with complex joins, we’ve seen customers save 60 to 70 percent compared to cloud data platforms like Databricks or Snowflake. And none of this requires ripping anything out — OcientAIQ sits alongside what you already have.
The thread through it all
Every story that night, the migration, the stack consolidation, the demo, the cost conversation, pointed back to the same thing I’ve written about before. The infrastructure tax on agentic AI in AdTech is real, and it shows up the moment an agent stops being a demo and starts making live, autonomous decisions.
The teams who’ve already crossed that line aren’t spending much time debating which model to use. They’re talking about what they had to build underneath it.
I’ll be at Programmatic I/O and Advertising Week this fall. If you’re there, come find me, or reach out beforehand. And honestly, I’m just as curious to hear where this doesn’t match what you’re seeing.