Enterprise AI in 2026 (and Beyond)

There's probably one simple term, and that's agents.

We've talked about agents for quite a long time, but it's that pivot from leveraging large language models for data retrieval, getting answers to questions, writing reports and getting insight, to actually moving into action.

The ability for organizations to start leveraging agents to do tasks and connecting agents to other parts of their core business so they can actually implement things.

So it's gone from a research-driven engagement to an action-driven engagement.

I think it's getting easier. Companies are starting to realize they need to focus on things that have a material impact on their business.

A lot of people initially thought, we'll use that edge case because it's safe and contained. But they see little value out of that, because even if it wasn't running on AI, it probably wasn't that valuable to the company.

You see the most value where people build new products and services, where they can drive transformation in their own business. Where agents move into action and transform workflows, that's where you start to see the value.

You can still work on a grand challenge in a step fashion. You can add in more capability, give the agent more permissions. But you start to realize value very quickly.

There's always this fixation about cost reduction in AI, but actually the most value isn't about cost reduction. It's being able to deliver new products, new services, to leverage the value of the data that you have.

In Europe in particular, we are concerned about security, safety. There’s strong regulation. There's always been something that said “Oh, we can't quite do that yet because we can't trust, or we don't have transparency.”

So there's three things that have really come together in the last 12 months.

It's those three things—compute, open models, and trusted agentic platforms—that allow people to say: “We know where the data resides, how the agent's going to act, and now we can move forward.”

I think it needs to be a little bit of both.

There needs to be board-level support. Often we would want sponsorship from the CEO, to allow the project to have access to the right tools, capability, data, etc. and to take on the grand challenge, the big task.

But at the same time, you need that knowledge and skills from the people doing the work.

Without that executive sponsorship, though, it's a challenge.

First of all, I'd try and understand where the company derives its value from. What is the secret sauce of that company? Where does it differentiate in the market? How can you empower that in this new AI world?

And then the question is, where does that value reside in data? Where is that data, and can you access it? Most corporates have lots of different systems that often don't speak to each other, and data lakes that no one quite knows how they work.

Then work out where you want to start deploying AI. Perhaps initially internally, to build faith and understanding and uplift people's skills, so they get comfortable using the technology. And then look at opportunities: how would we position that to the market?

You have to take people on a journey.

If you ask for a definition of sovereign AI, as many people as you speak to, you'll get the same number of definitions. What I think it is, is about choice.

You want to ensure where your data is and that it does not leak outside of the organization. But at the same time, you want to leverage that to build new products and services, or serve citizens.

You need to understand the whole stack. We talk about a five-layer cake. Right from where's the energy, where's the data center, what is the compute platform, to what runs on top of that—the models—and then the application.

If you want to build a customer service AI platform, are you going to give agents access to your billing system? Maybe you want to enable it to give refunds. How do you do that in a safe manner?

We'll get to the same place as with cloud in general. It will be hybrid.

It depends on the workload. You don't restrict certain workloads and add complications that aren't necessary. Agents serving information about your products and services? That's probably fine. That's not particularly a secure workload.

But then there's other elements which contain your core IP. Where should they sit? It's about orchestration and choosing the right AI model for the task.

For certain tasks, you don't need a frontier model. Why use a lot of tokens and expense for a simple task? You could have your own internal model, or take open-source models and train them on your data.

At the same time, for deep research you might leverage more complex models. Or you might need an air-gapped solution that's very, very secure indeed.

It's a continuum between the most secure, right the way through to things you're more relaxed about.

A lot of people start with foundational models because they're there. They're easily available. You start to learn what that could do.

But you need to be clearer where you want to leverage those models, depending on your view on sovereignty and IP. Make sure the right guardrails are in place.

And then you look at the economics. For a lot of things you don't need a frontier model. You could have an open-source model, a smaller model. Or you could build your own.

I think the next wave—and we're seeing startups in this space—is building orchestration layers that build low-cost routing. Based on the ask, it routes to the right model, which is most cost-effective but gives the best result. And therefore the enterprise has control.

These were just some topics we covered. You can watch the full interview here.

One common thread: Enterprise AI is moving from experimentation into action. Making AI work in a real business takes more than a good model. Data, workflows, security and infrastructure all matter.

And economics, of course, matters above all.

Thanks to David for the conversation and to Nvidia for inviting me back to GTC.

See you next Saturday,
Tobias