I have been in the data business for over thirty years. Longer than I planned, if I'm honest. And the whole time, the goal has been the same.
A business person asks a question about the business. The data gives a trusted answer. That answer triggers the next question, and the next, and you get answers to those too. At the end of the chain, you know something you didn't know before, and you can act on it.
We have chased this with every generation of technology. In the OLAP days, Cognos PowerPlay got remarkably close: answers in seconds, drill down, follow-up questions. But only inside the cube someone had built for you. Step outside it and you were back in the queue for the data team.
Then data platforms moved to the cloud, and a curious idea took hold: that structured, modelled data was optional. Storage was cheap, compute was elastic, and you could load everything as it was and work it out later. Careful data modelling started to look like the thing that slowed engineering down. Teams moved fast, and the structure quietly eroded along the way. Not because anyone decided it didn't matter, but because it felt slow next to everything else.
AI has turned that around. A model working on unmodelled, undocumented data doesn't produce insight. It produces confident answers that nobody can trust. AI needs exactly what good data engineering always provided: structure, clear definitions and context. The foundations that felt strict a few years ago are what pay off now. Discipline isn't the opposite of speed. It's what makes speed safe.
At Solita Summit last week, I listened to Neste and Nordic Investment Bank (NIB) describe a year of agentic data work. For the first time, I feel we are genuinely close to that goal, and at enterprise scale. Here's why.
When generative AI arrived, many of us thought the problem was solved. Put a chat box on top of the data warehouse and let people talk to their data. The pilots looked great. Then they stalled.
Aarni Kekkonen, Lead Data Architect at Neste, explained why better than any analyst report I've read. A language model has been trained on more or less everything public on the internet. It may know the basics of how oil is refined. It does not know how Neste's refinery is set up, which devices sit where, or what moves through them.
Give it that knowledge, and a pressure reading stops being just a number. The model knows which valve it comes from, how that valve is built, what flows in and where it goes next. Now it can answer a business question, not just run a query.
That is the missing piece. Not a better model, not a bigger platform. Business knowledge, captured in a form that AI can use.
Neste runs a large platform: around 30 developers, more than 70 source systems, hundreds of self-service Power BI users. A year ago, Aarni's verdict was that the tech wasn't ready. The models were good, but you had to feed them context by hand.
One year later, everyone works with agents in some form. Piece by piece, the team connected the agent to what developers already rely on:
The agent now also writes the documentation back into the wiki. Documentation, the eternal bottleneck, became the first clear win.
The developer's job has changed with it. In Aarni's words: less writing SQL, more reading metadata which describes the intent. The work is now finding the hard context, like an undocumented source system, and keeping the agent on track. His lesson is simple: the more context you give the models, the better the results.
Where Neste grew its way bottom-up, NIB designed it top-down. Karri Linnoinen, NIB's Chief Data Officer, put it plainly: AI agents are only as good as their context, and model capability alone is not enough.
NIB gives its agents four layers of context:
On top of this, NIB is building three stages into one operating model:
The second stage is the one that matters most to me. It turns context from a one-off clean-up project into a loop that keeps getting better on its own.
What I appreciated most was that neither speaker oversold it.
Both also agree on where the human belongs. Promoting a change to production is always a human action. Karri's line stayed with me: human in the loop is not a temporary compromise, it is a core architectural feature. Strong context plus approvals keeps the blast radius small when a model gets something wrong.
And the old rule still holds, perhaps more than ever. Quality in, quality out. Messy, conflicting documentation from years ago won't fix itself because an agent reads it.
Put the two stories together and the pieces of my thirty-year goal are finally on the table:
The PowerPlay cube was fast because someone had captured the missing business knowledge into it by hand: the right names, the hierarchies, the groupings. What has changed is that we no longer need to do this one cube at a time. Business knowledge can now be captured across all the relevant sources and kept up to date as work happens. That is what makes the chain of questions possible at enterprise scale, not just in a carefully prepared pilot.
Or, in one line: give your AI what your people know, and you can trust the answer on all your data.
Neste and NIB are showing what the journey looks like. Not every organisation needs a year to find out whether it works for them, though. Together with Solita, we have put together a way to test it in five days: your business questions, your data, and an honest answer at the end on whether the AI gets them right.
If that sounds like something you'd like to try, drop me a message.
Matti Karell
Chief Executive Officer
Agile Data Engine