Adoption lesson

My Team Did Not Believe AI Could Produce Good Code, yet we improved Delivery 4x.

A believable technical objection slowed adoption: if an LLM is trained on a huge amount of mostly average code, why would it reliably produce high-quality output? Progress started when we stopped arguing about whether AI was good or bad and started asking what conditions produce code we are willing to accept.

Why the skepticism made sense

An LLM is a statistical model trained on a huge amount of code. Most code is not excellent. A lot of it is average at best. Only a small portion is genuinely high quality.

The obvious conclusion was straightforward: if the model predicts likely output from average training data, then it will tend to produce average code. That is a serious argument against AI use where quality matters.

That belief slowed adoption. People were less likely to trust the tool with meaningful work. They were more likely to treat weak output as proof that the whole thing was flawed. They were less motivated to invest time in learning how to use it well.

What kept standing out

  1. AI can understand code.
  2. AI can write code.
  3. AI can produce acceptable code.
  4. The quality of the output depends on the model and the context.

That last point mattered most. If context affects output quality, then part of the problem is controllable. The model is what it is. The context is something we can shape.

The more useful question

The question stopped being: is AI good or bad at coding?

The question became: under what conditions does it produce code we are willing to accept?

That shift was more useful because it moved the discussion away from ideology and toward experimentation.

What we learned by trying

We tried giving it a lot of context. We tried giving it very little context. We tried different ways of describing the problem.

At one point we started treating the model like a drunk junior developer. That framing helped. It forced us to be more explicit, more structured, and less trusting of vague output.

That was the real shift. Not from skepticism to blind belief. From skepticism to a more precise understanding of what controls quality.

What changed in practice

We stopped treating poor output as proof that AI could not code. We started treating poor output as a signal.

  • Maybe the task was too broad.
  • Maybe the context was wrong.
  • Maybe the instruction was vague.
  • Maybe we were expecting good output from a setup that did not deserve it.

That change in thinking improved how we worked with the tool. Better usage led to better output. Better output led to more trust. More trust led to more adoption. That is how the delivery improvement happened.

The lesson

The important lesson is not that AI is magical. The lesson is that adoption was blocked by a believable technical idea, and progress started when we reframed it: AI is bad sometimes, so what can we do to help it do better?

That is a better operating model for teams. It treats output quality as something to influence through task design, context quality, review discipline, and experimentation instead of as a fixed verdict on the tool.

Get in touch

Let's make it
move

Best fit: delivery or transformation roles in distributed engineering, legacy modernization and platform migrations, AI-enabled SDLC with real quality and security guardrails.