Most companies approach AI engineering as a tooling problem: give engineers better models, copilots and agents, and adoption will follow.

It won't.

We learned this ourselves at Sofigate. Giving every employee AI licences built skills fast, but costs rose and business results did not follow until we changed how work was organised.

As AI becomes part of software development, the constraint shifts from model capability to organisational capability. In our experience, one developer can effectively steer only around ten agents before reviews and decisions start piling up. The bottleneck moves from writing code to deciding, reviewing and owning it.

This talk argues that your AI engineering strategy is only as strong as the organisation around it, and works through five leadership decisions: how to move from experimentation to adoption, how humans and agents divide the work, what changes in roles and practices, how to govern without killing speed, and how to measure impact rather than AI activity.

Not another tooling overview, but a short, practical playbook for engineering leaders.

Key learnings:
- Recognise when AI engineering has become an organisational problem rather than a tooling one
- Design how humans and agents divide work, review and ownership
- Replace AI activity metrics with engineering and business outcomes