Shadow AI: when 30 devs use AI on their own terms
Mike Codeur
I just got back from France. I went to visit real companies: teams shipping code to production every day, with legacy, constraints and deadlines.
I expected to talk about models and tools. We barely did.
Not a single one asked me whether Claude Sonnet 5 beats GPT-5.5. Not once. That stings a little when you are a YouTuber and that is 80% of your feed.
The latest model is a YouTuber's topic
On YouTube, we chase every release. New model, new benchmark, new comparison. It gets views, and I am part of that.
In the field, it does not exist.
The companies I visited run on models that are six months old, and they are perfectly fine with it. Their question is not "do I gain three points on a benchmark", it is "is my process reliable" and "does it cost me less".
The model has become a commodity. Like electricity: nobody asks which brand of electricity they are plugging into. What matters is what you do with it.
If you are picking your AI tool based on a benchmark, you are answering a question your company already stopped asking.
The solo dev setup works. That is exactly the trap.
Take a single dev. He has his Cursor or his Claude Code, his rules, his prompts, his skills. He tuned that setup for weeks, refined it, and knows it by heart.
It works. Genuinely. He ships faster, he is happy, and he is right to be.
The problem is that this success is strictly individual. It is not transferable, not reviewable, not maintainable by anyone else. It lives in a local folder nobody has ever opened.
Alone, it is an edge. At thirty, it is debt.
30 devs, 30 configs: shadow AI
Shadow AI is AI already used everywhere in the company, without anyone having framed it. It is neither banned nor governed. It is just there, off the books.
Look at your team today. On every other layer, the debate was settled long ago:
| Layer | Shared standard? | Who enforces it |
|---|---|---|
| Language and framework | Yes | the team, once and for all |
| Code style | Yes, one linter, one shared config | CI |
| Tests | Yes, one suite, one threshold | CI |
| Code review | Yes, a PR, a reviewer | the process |
| Deployment | Yes, a pipeline | the platform |
| AI: rules, prompts, skills | No | every dev, on their own |
Nobody would accept every dev running their own linter with their own rules, with no visibility into what their neighbour does. Yet that is exactly what happens with AI.
You know the old "works on my machine" joke. It took fifteen years, Docker, CI and infrastructure as code to settle it. Solo AI is the same thing: it works on my machine, with my rules, with my prompts. On a team, it breaks.
Concretely, at thirty devs with no framework:
- nobody knows what their neighbour is generating
- two devs solve the same problem with two different prompts and two different results
- whoever has the best setup does not share it, because there is nowhere to put it
- when they leave, the setup leaves with them
- no compounding: a few individuals improve, the team does not
- no control: neither over what goes out to the models, nor over what lands back in the code
What companies actually want
Not for you to go faster on your own. For thirty devs to share the same rules and the same skills, and for it to hold over time.
In plain terms, they want three things:
- Unify. One shared, versioned set of rules and skills that everyone uses.
- Compound. What one dev discovers on Monday should benefit the other twenty-nine on Tuesday.
- Regain control. Know what goes out, what comes back, and on what basis.
A concrete example
One company we worked with wanted exactly that: a unified system of rules and skills for the whole team. Not to go faster. To regain control.
Their ask was neither "which model" nor "which tool". It was: how do we get everyone working the same way, make one dev's discovery benefit the rest, and stop depending on individual goodwill.
This is not theory. It is the real topic of 2026.
What it changes for you
If you are a dev with a job, there is a read to make here.
The skill gaining value is not "I know the latest model". Everyone will have it in six months, and it will be a commodity.
It is knowing how to turn an individual practice into a team standard. Writing rules thirty people can read, understand and maintain. Making a group better, not just yourself.
Far less sexy than a benchmark. Also far harder to replace.
Conclusion
The topic for the coming months is not "which model should I use". It is how an entire team uses AI without everyone drifting off on their own.
AI is already inside your company. The only question is whether it is there off the books, or properly framed.
I break down the full field report in the video: Shadow AI, back from France
And if you want this kind of analysis every week, it is in The Agentic Dev.