AI Agents and Analytics: Lessons from Our New York Roadshow

At our recent Everywhere Roadshow in New York, we asked a room of product and technology leaders how many analytics queries they expected to run through AI agents in the future. Almost everyone thought the answer would eventually be close to 100%.

I found that quite surprising. We've been collecting responses for our Everywhere Index, and the broader research paints a more cautious picture. Even so, the people around the table in New York seemed convinced that agents would eventually handle almost everything.

Looking back, it probably shouldn't have surprised me. We had brought together people who were already experimenting with agentic analytics, and the conversations quickly moved into some fairly technical territory. I particularly enjoyed hearing how differently people were thinking about the same problems.

MCP, in-app analytics and why I'd choose both

One discussion I remember particularly well was between Alexander Sabel from Perchwell and Jinbae Kim from Zipsec. Alexander was very much in favor of MCP, whereas Jinbae was considering whether certain capabilities should remain inside the application.

My reaction was: why choose?

I can see plenty of reasons to expose analytics through MCP. People increasingly want to access information through their preferred AI assistants, and there's little reason to force them back into an application for every question. But I also expect people to continue using analytics within the products they work with every day, and use the results in the tooling.

The important thing is that both experiences can access the same underlying analytics. Otherwise, we're creating a situation where someone asks an agent a question, opens a dashboard to check the answer and potentially gets a different result.

This came up in our Everywhere Index research, too. Only 11% of surveyed product leaders are confident that the same question would produce the same answer across their dashboard, chat interface and AI agent.

That's a fairly uncomfortable number when you consider how quickly companies are adding AI functionality to their products.

We also talked about generative UI, which I find particularly exciting. If interfaces can be generated around what someone is trying to do, many of the barriers associated with traditional analytics tools could disappear. Users wouldn't necessarily need to understand a reporting interface before they could start working with their data.

Of course, that creates some interesting design questions. One we discussed was how companies maintain their brand identity when analytics can appear in so many different places, including interfaces they haven't designed themselves. I suspect we'll be hearing much more about that as generative UI develops.

The early adopters are already thinking several steps ahead

The expectations around AI agents were probably the clearest indication of the kind of group we had in New York.

Our early Everywhere Index results show that 64% of product leaders expect more than a quarter of queries on their data to come from AI agents within two years. That's already quite a substantial change. The people in our New York discussion were thinking much further ahead, with almost everyone expecting agents to eventually handle nearly all queries.

I wouldn't read too much into the difference. This was a small group of people with a particular interest in the topic, many of whom were already familiar with the technologies we're discussing. Still, it's interesting to hear what people expect when they've spent time experimenting with these capabilities themselves.

I also enjoyed the conversations happening around the room. Prashant Nandavanam from UiPath was already introducing us to other people interested in agentic analytics, and Arlando Battle from Odie mentioned technologies he wanted to investigate after the event.

We had a great mix of people, including Yarne Hermann from Socure, Shubham Upasani from Octus, Hakim Elakhrass from Soda and Santosh Nayak from Zeta. It was particularly nice to meet people from different backgrounds who had clearly been thinking about similar questions.

Model agnosticism deserves a bigger conversation

Another topic we spent time on was the choice of LLMs.

I think companies are going to become more selective about which models they use. Some analytics tasks need sophisticated reasoning, but others can be handled perfectly well without using the most expensive frontier model available. Once AI becomes a regular part of a product, the economics of those decisions become harder to ignore.

At Luzmo, we're taking a flexible approach to AI architecture. We currently use OpenAI, but we're not tied to a single provider or even a single type of model. We're also experimenting with combining general-purpose LLMs with more specialized decision models that work with deterministic options rather than generating content. The interesting part is finding the right combination of models, reasoning capabilities and thinking depth for each task. As speed, performance and costs continue to change, I believe this kind of setup will become increasingly important. The future probably isn't about choosing one model, but finding the right mix for the job.

There's also the question of sovereignty, which I expect will be especially relevant during our upcoming Paris event. Inference in your region and providers in your region are giving companies more options, and I'm interested to hear whether product leaders are already taking those options into account when designing their applications.

It's one of the topics I'd like to spend more time discussing as the roadshow continues.

Taking the conversation to Europe

What I enjoyed most about New York was how much we learned from the people attending. We deliberately kept a small group, and that gave everyone an opportunity to get involved. A few conversations continued well beyond the topics we'd originally planned to cover, which was exactly what I'd hoped for.

We're continuing the Everywhere Roadshow over the coming weeks, with three European events:

• Amsterdam, November 12: Hosted together with ClickHouse, with conversations around analytics infrastructure and the growing demands of AI agents.
• Paris, November 26: Jonathan Wuurman and I will host an evening where I'm particularly keen to discuss European LLMs and model agnosticism.
• London, December 2: Another opportunity to compare experiences with product leaders working on AI-powered analytics.

We'll share some early Everywhere Index findings at each event, but the format is mostly about getting people together to discuss what they're seeing in their own work. Places are free and limited to 30 participants per city.

After New York, I'm curious whether we'll hear the same predictions about AI agents in Amsterdam, Paris and London. I have a feeling the European conversations around model selection might be particularly interesting.

If you're working on these technologies, I'd be happy to meet you at one of the upcoming events.