AI adoption at Extenda Retail did not start with a strategy deck. It started with curious people quietly trying things, and my job was to turn that energy into momentum without killing it.
It began at the edges
In 2024 and into 2025, AI showed up the way it does in most software companies: one engineer at a time. Developers were trying GitHub Copilot and JetBrains AI. Colleagues across the business were using Gemini to draft, summarise and think out loud. Product teams were already building assistants such as Tillie, nyce.buddy and ExtendaBuddy into what we ship.
None of this was coordinated, and that was its strength. The people closest to the work found the use cases faster than any central team could have. A joint hackathon with Google Cloud showed how much latent appetite there was once people were given time and permission to build.
Guardrails before scale
Grassroots energy has a downside: shadow AI. With responsibility for information security and data protection on my desk, I could not let enthusiasm outrun governance.
So the first structural move was not a programme. It was a set of guardrails that made it safe to say yes:
- A formal AI Security Procedure
- A published list of approved tools, starting with Gemini, NotebookLM, Atlassian Rovo, Claude and GitHub Copilot
- A simple request-and-approval route for anything new
The point was to turn “can I use this?” from a negotiation into a lookup. Clear guardrails increased adoption rather than slowing it, because people stopped worrying about getting it wrong.
Why experimentation alone stalls
By late 2025 we had pockets of excellence and a long tail of people who had tried AI once and moved on. Experimentation creates champions; it does not create a habit across 300 people. The gap between the most and least fluent teams was widening.
That was the signal to shift from letting it happen to making it happen.
Turning on acceleration
In Q1 2026 we launched a structured AI Acceleration Program, working with an external partner, IAMAI. Three things mattered most.
- We listened first. An all-staff AI survey told us where people really were: what they used, what blocked them, what they feared.
- We brought leadership in early. Executive discovery workshops moved AI from an IT topic to a leadership agenda, with each executive looking at their own function.
- We invested in fluency, not just licences. An AI Learning Program, AI Fluency sessions, a shared prompt library and short weekly lessons in Slack made learning a routine rather than an event. Challenges like the “AI Swap”, where everyone automates one task of their own, made it personal.
We also made deliberate tooling choices instead of letting every team pick its own. Claude Code became our standard for agentic coding, and we committed to a central MCP gateway so that connecting AI to our systems happens with security and cost under control.
What I would tell another leader
- Don’t kill the grassroots – harvest it. Your early adopters are your future trainers and stream leads.
- Guardrails are an accelerator. Clear rules on tools and data unlock adoption.
- Fluency beats access. Licences without learning produce dashboards of inactive users.
- Know when to change gear. When champions stop multiplying on their own, it is time to add structure.
Acceleration got AI into daily work. It also made the next question unavoidable: how do we turn individual productivity into business value we can measure? That is the story of the AI Transformation program, which I cover in the next post.