I went to Big Data LDN and, in a completely unexpected turn of events, everyone was talking about AI.

There wasn’t a huge amount that completely changed how I think about AI, but that was useful in itself. Across the sessions, I kept hearing variations of the same challenges businesses are already working through. How do you move beyond an interesting AI experiment and create something that delivers genuine value? How do you measure that value? And how do you get people to trust and actually use it?

My favourite session, Scaling AI for Business Impact: From Experimentation to Value, brought a lot of this together.

Start with the problem, not the AI

One of the clearest messages from the session was also one of the simplest: don’t start with the technology, start with the business problem. It sounds obvious, but with all the excitement around AI, it’s easy to start with what the technology can do and then look for somewhere to use it. A related question from the session was:

Do you want to improve what you’re doing and do it better, or completely change what you’re doing?

That distinction matters. Once the problem is clear, we still need to decide what role AI should play in solving it. Should it augment an existing process, automate part of it or enable us to rethink the process altogether?

Being clear about that from the beginning shapes how we think about value, adoption and the controls we need around it.

Easier answers make judgement more important

The conversations around analytics also raised a broader point about the role people play as AI becomes more capable.

If asking a question and getting an initial answer becomes easier, the value increasingly moves towards understanding whether that answer actually makes sense. That requires good data underneath it, but also business knowledge, context and critical thinking. Someone still needs to know when to challenge an answer, what might be missing and which question is worth asking next.

The same applies more broadly to AI adoption. A technically good solution won’t get very far if people don’t understand it or trust it. We need to be clear about where AI is there to help, what remains a person’s responsibility and where human judgement is still required. That means culture and training aren’t side issues to solve after the technology is built. They’re part of whether it succeeds in the first place.

More autonomy needs stronger controls

Trust also links closely to governance, particularly as we move from AI that provides information towards AI that can take action.

One of the later sessions focused on governing AI agents and made an important distinction: telling an AI what it is allowed to do isn’t the same as actually controlling what it can do.

There’s a big difference between an assistant that can answer a question and an agent that can update a system, send a communication or trigger a workflow. As the level of autonomy increases, the controls around it need to increase too.

That doesn’t mean every use case needs a huge governance framework. The level of control should reflect the level of risk. Permissions, accountability, auditability and human approval become more important as we give AI the ability to do more.