Your AI rollout needs more than early adopters

Most AI rollouts begin the same way. The licences go out, a few people take to the tools straight away, and their work gets visibly sharper. Leadership sees those early wins, assumes the rest of the team is on the same path, and turns its attention elsewhere. A few months on, nothing much has moved. A handful of people are genuinely good with AI, and everyone else is about where they started.

That gap is the actual problem, and it hides well, because the early signs look exactly like progress. The teams we work with rarely struggle because the tools are weak. They struggle because individual skill with AI does not spread on its own, and turning it into something a whole team can rely on is a piece of work in itself. Here is what that work involves, and the one thing it cannot promise.

For teams here in Vietnam the starting point is lower than the headlines suggest. In Cisco’s first AI Readiness Index, back in late 2023, only 27% of the organisations it surveyed in Vietnam came out as fully prepared to deploy AI. That was a vendor survey of large companies and it is two years old now, so read it as a rough baseline rather than today’s number. The direction it points in still holds. Most teams are starting from scattered, early-stage use, not a running practice.

The wins are real. They just stay put.

It is worth being clear that the upside is genuine. In a set of three randomised trials covering 4,867 developers at Microsoft, Accenture and one large manufacturer, access to GitHub Copilot raised the number of completed tasks by around 26% on average, with the strongest gains among junior and shorter-tenure people (Cui et al., 2025). The average carries a wide margin, and that junior-versus-senior split comes from the Microsoft group alone, so treat the levelling-up effect as a direction rather than a settled law. But the headline is not in doubt. Used well, these tools do real work.

The catch is where that value lands. When Faros AI looked at telemetry from more than 10,000 developers, people using AI assistants completed about 21% more tasks and merged about 98% more pull requests on their own. None of it showed up at the level of the company. Overall delivery stayed flat, and the average pull request grew by about 154%, which is to say the extra output partly inflated rather than shipped (Faros AI, 2025). That is a vendor’s correlational study, not proof of cause, so hold the numbers loosely. The pattern it describes is the one we keep seeing. Individuals speed up, the system around them does not, because the bottlenecks were never the typing. They were review, integration, and deciding what to build.

You also cannot read the situation off how busy people look, or how fast they feel. Licence counts and weekly active users tell you people opened the tool, not that the work got better, and the felt sense of speed is no more reliable. In a small but careful trial, METR had 16 experienced open-source developers work on their own mature codebases, some tasks with AI allowed and some without. They expected AI to speed them up, and afterwards still believed it had, by about 20%. In fact they had been about 19% slower with it (METR, 2025). That is a gap of roughly 40 points between how fast they felt and how fast they were. It is a narrow setting, expert maintainers on code they know intimately, and it does not mean AI slows everyone down. The lesson that travels is simpler. Self-report is a poor instrument here, and “everyone’s using it now” tells you almost nothing about whether it is working.

More training does not close the gap

The instinct, when adoption stalls, is to run another workshop. It rarely helps for long. A one-off session lifts usage for a week or two and then fades, because what you are really asking for is a change in habit, and habits do not install in an afternoon. The often-quoted “21 days to form a habit” is a myth that traces back to a 1960 self-help book, not to any study of habits. When researchers actually measured it, following 96 people forming a daily habit, the behaviour took on average 66 days to become automatic, with individuals landing anywhere from 18 to 254 (Lally et al., 2010). That work was about simple daily behaviours rather than software, so the number is a frame and not a target. The point it makes is the one that matters. A new way of working takes weeks of repetition in the real job to stick, not a single transfer of knowledge. Enablement that ignores that is paying for forgetting.

What actually moves a team

If a workshop is not the lever, what is? In practice, the teams that turn AI into a dependable capability tend to get four things right. None of them is exotic. All of them are deliberate.

Put the tool where the work happens

The fastest way to kill adoption is to leave the tool off to the side. A separate chat window or a sandbox is fine for discovery, but people drift back to their real workflow and the side tool gets abandoned. Prakash Kota, the CIO at UKG, puts it plainly: “if it’s not in the workflow, it won’t stick”. His fix has two halves. Embed the capability at the point where the work already happens, then close the old manual path so the new one is simply how the job is done now. Engineers at Tekion describe the same instinct, moving from tools you have to remember to invoke towards ones that fire as part of the process. One caveat is worth naming. Those companies had the engineering muscle to build that integration themselves, which a smaller team may not, so the principle generalises more cleanly than the tactic.

Spread it through peers, not a single expert

When one person becomes the go-to for all things AI, you have built a bottleneck, not a capability. The skill has to move from the few who have it to the many who need it, and that happens through peers more than through any central team. Some large organisations have leaned hard into this. Citi reports building a network of more than 4,000 employees it calls AI Accelerators, sitting alongside a much smaller group of senior AI Champions, across a workforce of roughly 182,000 (reporting via Business Insider). That is on the order of one accelerator for every 45 people. The figures are Citi’s own and not independently audited, and at a team of ten to a hundred the maths is different, but the shape is instructive. You want a layer of approachable peers, not one oracle. Who you pick matters too. One adoption lead at Manulife described choosing champions through three lenses at once: who colleagues already turn to, whose performance backs it up, and whom leaders will openly endorse. It is one company’s account rather than a proven recipe, but it beats handing the role to whoever is loudest about AI.

Leaders go first, in the open

Funding a rollout is not the same as modelling it. If the people running a team do not visibly do the work themselves, the message lands that this is for everyone else. Some founders take that to an extreme that is hard to miss. The CEO of Airtable, Howie Liu, says he is personally the single heaviest AI user in the company by cost (Lenny’s Newsletter, 2025). Sendbird’s leadership treats how much its most senior people use these tools as its strongest signal of whether adoption is real, and at Personio, engineering leads were expected to have done hands-on prototyping themselves before turning up to lead the rollout. These are leaders’ own accounts and the impact is not independently measured, so take them as illustration rather than proof. The underlying move is sound regardless. Visible practice from the top does more than a sponsored budget line.

Expect resistance, and meet it straight

Some of the resistance you will hit is not about tooling at all. It is the quiet worry that the point of all this is to need fewer people. Pretending otherwise does not work. When engineers at Doctolib described their rollout at a 2025 conference talk, they talked about meeting that fear directly, with an “AI summer camp” of sessions explaining how the models actually work and a shared vision of the AI-augmented engineer written with their CTO. That is their own account of what they did, not an audited outcome, but the instinct is right. People adopt a change they understand and have had a hand in shaping. They resist one that arrives as a mandate with an unspoken motive.

Measure the work, not the activity

All of this only holds together if you are honest about what you measure. Counts of logins and prompts are easy to gather and easy to game, and as the METR result shows, even the people doing the work cannot feel the difference reliably. So look past the activity to the work itself. Did the thing that used to take three days now take one, on a task you can actually check? Did the quality of what juniors ship go up enough that review got lighter rather than heavier? Those signals are slower to read than a usage chart, and they are the only ones that tell you whether the practice is real.

That is the whole argument, really. A few people doing impressive things with AI is a promising start and a poor finish. The distance between that and a team whose work is reliably better does not get closed by sharper tools or another workshop. It gets closed by putting the tools where the work is, spreading the skill through peers, leading from the front, and measuring what actually changed. That is deliberate work, and it is most of what we do alongside the teams we work with. If that is the gap you are looking at, here is how we work with teams.

Sources

• Cui, Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. economics.mit.edu
• Faros AI. (2025). The AI Productivity Paradox: What Data from 10,000 Developers Reveals. faros.ai
• METR. (2025, July 10). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. metr.org
• Lally, P., van Jaarsveld, C. H. M., Potts, H. W. W., & Wardle, J. (2010). How are habits formed: Modelling habit formation in the real world. European Journal of Social Psychology, 40(6), 998-1009. onlinelibrary.wiley.com
• Cisco. (2023, November). AI Readiness Index, Vietnam, relayed by VietnamPlus. en.vietnamplus.vn
• Kota, P. (UKG). Increasing AI adoption with agents built to serve all employees. CIO.com. cio.com
• The quiet work behind Citi’s 4,000-person internal AI rollout. AI News, citing Business Insider. artificialintelligence-news.com
• Liu, H. (Airtable). How we restructured Airtable’s entire org for AI. Lenny’s Newsletter. lennysnewsletter.com
• Tanay, J., & Bentkowski, T. (Doctolib). (2025). Embracing AI adoption at Doctolib engineering. Devoxx. youtube.com
• That Viral MIT Study Claiming 95% of AI Pilots Fail? Don’t Believe the Hype. Marketing AI Institute. marketingaiinstitute.com