Your organisation is already an agent. Your AI agents just have to fit inside it.

When I talk to engineering leaders about where their AI programmes have stalled, the story is almost always the same. They built a promising prototype. They tried to put it in production. And then it hit the organisation, and the organisation ate it.

The operating model eats the model

The organisation isn't a passive substrate. It has its own decision-making patterns, its own escalation paths, its own sense of who is accountable when something goes wrong. An AI agent dropped into the middle of that without an answer to any of those questions will be rejected - often politely, sometimes explicitly, always firmly.

So the question for any serious enterprise AI programme is not "how do we build the agent?" It's "what does this team's operating model look like with the agent in it?" Who does the agent report to? Who reviews its decisions? Who owns the incidents it causes? What can it do without approval, and what can it not?

You are not deploying a model. You are redrawing an org chart.

Start with the interface, not the model

The most productive AI programmes I've been part of spent roughly as much time on the agent's interface to the organisation as they spent on the agent itself. What does it ask for? What does it offer? How does it signal confidence, and how does it ask for help?

That's not UX, exactly. It's closer to onboarding. You are introducing a new actor into a team, and if you don't explicitly design how it fits, the team will improvise - and it usually won't improvise in a direction you like.