Exactly one year ago we published our first post on AI agents here. Back then the most common question in client conversations was: “What actually is that?” Today it is: “Why is our pilot still not live?”

The most honest number of the year

Anyone reading through the 2026 studies finds the same pattern everywhere: almost every larger company has by now experimented with agents — but only a fraction of those projects have made it into regular operation. This gap between pilot and production is not a technology gap but an organisational gap. The prototype that delights in the workshop has no error handling yet, no access control, no monitoring and nobody responsible when something goes wrong at night.

What productive agents do differently

A simple checklist has proven itself in our projects. First: a clearly defined process with a measurable outcome — no “agent for everything”. Second: defined boundaries, i.e. which systems the agent may read, which it may change and where a human approves. Third: observability from the start, because an agent whose decisions nobody can trace will be stopped by the works council or the auditor at the latest — rightly so. And fourth: an operating model. Software is maintained, agents are managed.

The culture change behind it

The most surprising insight from twelve months of agent projects: technology is rarely the bottleneck. The harder question is how teams work when part of the work becomes delegable. The companies where the leap succeeds treat agents like new team members — with onboarding, clear responsibilities and regular feedback. Those where it stalls treat them like an IT tool that is installed once and then forgotten.

One year of AI blog, twelve posts, one conclusion: the lead is not created by trying things out but by following through. If you want to get your pilot across the finish line, talk to us — by now we know the stumbling blocks by their first names.