How to Adopt AI in a Software Team
Does your AI adoption strategy consist of giving out Copilot, Cursor, or Claude licenses and letting them figure out how to use them?
Over the past year, I’ve found this to be a widespread strategy in software companies—companies whose teams face high workloads and, due to lack of knowledge and time, see AI as just another technology that can be learned self-taught by each developer.
When I start working with a client who has followed this approach, the first thing I do is measure the actual use and impact the initiative has had, and what I observe is the following:
Usage is concentrated among a few power users. People who have proactively taken the time to research and learn how to unlock AI’s potential. Typically, this is less than 10% of the team, and these are people who have dedicated significant personal time to learning.
A significant number of developers barely use AI or use it very little. There are different reasons here, ranging from an initial trial that gave poor results to fear of losing control over the code, especially when they use agent mode without criteria.
The average number of use cases per developer is around 3: searching for information, generating code for new features, and generating tests. However, many are unaware of other use cases that are also very useful: designing, prototyping, documenting, debugging, refactoring, reviewing code, or building scripts and automations.
They continue working as they did before AI: same Scrum methodology, same best practices (or lack thereof), same product management, etc…
Some developers purchase personal licenses out of pocket, especially Claude Code, and this means the company has no control over data privacy configuration or visibility into metrics. I’ve seen several companies unknowingly sharing proprietary source code with Anthropic for training new models for this reason.
And all companies that approach it this way have a bittersweet feeling about AI’s impact, as they may have seen some internal success cases, but the feeling is that everything continues more or less as before.
In my experience, what does work is defining an AI adoption strategy at the technology team level where the first step is to communicate to management what they can expect from current generative AI capabilities, beyond the hype, setting expectations regarding the impact and investment they will need to make.
On one hand, the impact is not x10 or x2 as many sell, but you can aspire to an increase between 10% and 30% (although this range should not be interpreted literally, but rather as an order of magnitude of what can be expected), depending greatly on the technology, project and tasks, the person’s seniority, the tools, and their knowledge about AI.
On the other hand, the minimum recommended investment includes: tools (Cursor or Copilot in Teams or Business are the basics), training, and technical coaching. And a realistic adoption plan requires several months to reach cruising speed, where AI is already part of everyone’s daily routine.
In the accompaniments I provide to companies to help them adopt AI, I usually follow this plan:
Start by evangelizing management, setting realistic expectations and sharing experiences that help them understand the possibilities. After all, AI is not just another technology, but rather changes our way of working, and this change requires leadership to be convinced and involved to carry it out. A widely held idea about AI is that it’s an accelerator; however, looking only at the time variable can lead us to end up being faster at delivering lower quality. A more advisable strategy is to see AI as an enhancer of both speed and quality, calibrating initiatives to increase both.
Propose training for everyone, to raise the general level and reduce the learning curve that exists when one is self-taught. If we don’t provide training, it’s almost certain that there will be engineers who obtain suboptimal results, and the cost in the long run will be paid in unrealized productivity.
Complement training with pair programming sessions and technical coaching, where each developer works with a developer experienced in AI use on real code in a task or user story from their project. Because training usually provides a good theoretical framework, but putting it into practice is not so trivial. If someone has ever done pair programming with a more senior developer, they’ll know how to detect the value that can be obtained from this type of session.
Review and update development best practices, because AI will allow us to go faster, but if we don’t have solid best practices, we’ll end up with more bugs and more technical debt. Thanks to AI, we’re seeing teams adopt practices they had in the backlog, and with a lower investment than was needed before.
Measure and generate monthly reports on the use of Copilot, Cursor, Claude—it’s the only way to have an objective picture of how usage and adoption evolve.
Measure and generate engineering productivity reports to detect and quantify AI’s impact on software development. Some metrics that can be used are lead time, throughput, change failure rate.
Create an AI Guild where the team shares learnings and experiences with AI within the company. This will also allow the team to perceive that AI is a fundamental part of the work and should be present in conversations. One of the most common initiatives is to create a forum, chat channel, or similar dedicated to the AI Guild and encourage any team member to post about how they’re using AI.
Monitor news on tools and models, and share them with the team as soon as they appear. This is a field that moves so fast, and where we’ve seen updates that represent quantum leaps.
In my experience with teams that follow this strategy, AI adoption by engineers ends up being widespread, and the increase in productivity becomes visible after a few months. Additionally, there’s a clear improvement in the development process, with more and better best practices, which helps elevate the engineering team’s culture.
If you feel that AI is not meeting expectations in your team, you may need to review how you’re doing it. If you want to talk about it, you can book a call here.