Skip to main content
Steve Bolton

AI Maturity Assessment

Eleven questions on how your engineering organisation actually works with AI. About three minutes, and you keep the result.


What this measures

This assesses AI in your engineering workflow: how your teams design, build and ship software with AI in the loop. It does not assess AI in your product. Whether you put a model in front of your customers is a different question with different answers, and nothing below is about it.

You will get a number, and you should hold it loosely. It is a diagnostic, not a grade. Eleven questions across eleven dimensions means the score moves in ten-point steps, so a single answer can carry you across a boundary, which is why the band is what carries the meaning and the number is labelled indicative. The part worth your attention is the one you cannot see from inside: the dimensions where the honest answer was “Not sure”.

The eleven dimensions

Nothing is withheld here. The whole framework is below, one question per dimension, because the framework is the argument, and an argument you cannot read is not one you can disagree with.

Capability: more is better

Seven dimensions where a higher answer is straightforwardly better. These are the only ones that count towards the score.

  • Codebase Readiness
  • Context Architecture
  • Skills as Software
  • Testing Agent Behaviour
  • Observability & Measurement
  • Organisational Readiness
  • Meta-Agents

Calibration: trade-offs, not maxima

Three dimensions to read against the friction they create rather than as things to maximise. Lock an agent down far enough and it is safe and useless; leave it open and it is fast and exposed. Where the right point sits depends on what you are building and who for, which is why these sit outside the score.

  • Security & Guardrails
  • Cost & Economic Governance
  • Interface & Discovery

Enablement: banded on its own

More is better here too. It is not a trade-off. It moves on a slower timescale than the rest, so it is banded separately instead of being folded into the score.

  • Human Fluency

What it costs you

Eleven questions, one per dimension. About three minutes. It is free, there is no email address to hand over and no name to give. The result is a link: keep it, send it to whoever needs to see it. Your answers travel in the link itself, so there is nothing stored on my side to leak or to lose.

“Not sure” is a real answer and not a way of skipping. It is what produces the named blind spots on the result, which for most teams is the useful part. Use it where it is true.

Rather see one before you answer? Here is a worked example for Northwind, a synthetic organisation invented for this page. It is not a client and the profile is not anyone’s.

0 of 11 answered
Can a new engineer get productive quickly, or does our codebase fight them?
Have we written down and maintained guidance for our AI tools about how our systems work — or are they guessing from defaults?
Do we build reusable, reviewed AI workflows — or does everyone improvise their own prompts?
Do we verify that AI-generated work is actually correct — or trust it and hope?
Can we show with data whether AI tooling makes us faster and better — or is it a feeling?
Is AI adoption owned, resourced and led — or a few enthusiasts' side project?
Do our AI systems help maintain and improve themselves — or does every improvement need a person?
Are there real guardrails on what our AI agents can see and do — secrets, production, untrusted input?
Do we know what our AI tooling costs — and whether it is paying for itself?
Is working well with AI a competency across the org — engineers, product, design — or just a specialist few?
When a document or tool our agents rely on turns out to be wrong, do we fix the source — or do people quietly work around it?