AI agents are everywhere at the moment. Almost every conversation seems to come back to them in some form, whether that’s internal copilots, customer support, or automating parts of a workflow. And to be fair, a lot of the early results look promising. It’s now relatively easy to get something up and running that can answer questions, pull together context, and even take basic actions.
Where things start to get more interesting is what happens after that.
What we’re seeing
There’s a fairly consistent gap between what works in a demo and what actually holds up day to day. The model itself is rarely the issue. Most of the time, the limiting factor is everything around it.
An agent is only as good as the context it has access to, the systems it can interact with, and the boundaries set around how it operates. In practice, that means it sits across multiple parts of a business at once. Data, tools, processes. If those things are fragmented or inconsistent, the agent doesn’t fix it, it simply exposes it.
Where it starts to break
Early use cases tend to be quite contained, which is why they work well. Support triage, internal knowledge queries, simple workflow automation. Clear inputs, clear outputs.
As soon as you try to extend beyond that, the complexity increases quickly. Better data is needed, ownership becomes less clear, and the underlying processes need to be more thought through than they often are.
Without that foundation, outputs become inconsistent, trust drops, and the initial momentum fades.
What the better teams are doing
The teams getting further with AI agents aren’t necessarily doing anything radically different. They’re just more deliberate.
They think carefully about where the agent sits within a workflow, make sure the data it relies on is actually usable, and keep an appropriate level of oversight rather than trying to fully automate too early.
It’s not the most exciting narrative, but it is what seems to work.
Why this matters
It also ties into something we’ve been seeing more broadly. Increasingly, the real value sits a layer above the code itself. In how systems are shaped, how decisions are made, and how different parts of a business connect together.
Agents don’t replace that. If anything, they make it more important.
As more organisations move beyond experimentation, that distinction will become clearer. It won’t just be about who is using AI, but who has the foundations in place to make it genuinely useful.