Everyone’s rolling out AI tools. Most rollouts are failing. Not because the tools are bad, but because organizations treat AI adoption as a technology decision when it’s actually a design decision.
When I was tasked with integrating LLM-based copilots across three teams at Colibri — Design, QA, and Data — my first instinct was to find the best tools. A few weeks in, I realized the tools weren’t the bottleneck. The bottleneck was fit. Each team had different workflows, different skill levels, different anxieties about AI, and different definitions of what ‘useful’ meant.
The Design team was cautiously optimistic but worried about losing their voice. QA was skeptical that AI could match their domain expertise. The Data team was eager but overwhelmed by options. Same organization, three completely different adoption challenges.
The framework I developed starts with a simple classification. For every task in a team’s workflow, ask: Does this require deep human judgment? Is this repetitive but context-dependent? Or is this purely mechanical? The first category stays human. The third gets automated. The middle category — that’s where copilots shine.
For the Design team, I built content drafting assistants that understood our templates and standards. They didn’t generate courses from scratch — they generated first drafts within our framework, which designers then refined. This respected their expertise while eliminating blank-page paralysis.
For QA, the copilot did automated first-pass reviews, flagging potential issues for human verification. The key insight: I positioned it as ‘giving QA reviewers superpowers’ rather than ‘replacing QA review.’ Framing matters enormously.
For the Data team, it was query generation and reporting assistance. The simplest integration technically, but it required the most training because the team needed to learn how to prompt effectively.
Results ranged from 30% to 60% efficiency gains depending on the team. But the metric I’m proudest of is sustained adoption. Six months later, every team was still using their copilots daily. Most AI rollouts see a usage spike in week one and abandonment by month two. Ours stuck because the integrations were designed for each team’s actual workflows, not imposed from above.
The lesson: AI adoption is a design problem. You need to understand the humans first, the workflows second, and the technology third. Get that order wrong and you’ll have expensive tools that nobody uses.