Research · Field guide · v1.0.0
How I'd roll out AI
Six stages of AI enablement, each tied to work you can open on this site.
Buying the tool is the easy part. The work is getting people who never asked for it to use it well, inside rules they can live with, and still using it a year later. I have done that job for court and payments software across 100+ deployments. These are the stages I would run for AI, and where you can see me doing each one.
01 Where are you hiring from?
The same six stages run in a different order depending on where the job sits. Pick one and the loop reorders.
You want people across the company using AI well a year from now. A pilot that fades in a quarter does not count. Start with the work, then the rules, then the training.
02 The stages
Find the real work
Before any tool, sit with the people who will use it and write down what they actually do all day. The good uses for AI hide in the dull parts of the job, and the people who do that job already know where they are.
Inside one company: Shadow the teams with the most repetitive work first. They will show you the first wins.
Where I have done it
Every court rollout I ran started on the clerk's side of the counter. The review console below was built for a reviewer who had to finish alone, so the screen follows their checklist.
Set the guardrails
Write down which tools are approved, what data stays out of them, and who signs off before anything AI drafted leaves the building. People use AI more once they know where the edges are.
Inside one company: Legal and security sign the policy once, so no team has to negotiate its own.
Where I have done it
I published a plain-language acceptable-use policy for small offices, where every rule has a way to follow it without engineers. The security self-assessment beside it is honest about what one person cannot cover.
Teach until you are not needed
Give people one early win they did themselves, then hand them a little more than they can do today. Let them pick what to build. The goal is a team that keeps going after you leave.
Inside one company: Recruit a champion in each team and train them first. They answer the questions you never hear.
Where I have done it
Heddle turns a certifying body's blueprint into a weekly curriculum, and its students passed at a 100% rate across three cohorts. The Ukrainian course takes a learner from the alphabet to advanced reading, free. I also run a small class teaching retirees to build with AI, planned from a literature review on how older adults learn.
Build what people will adopt
Some needs a chat window will not meet. Then you build the small tool that fits the job: a form that drafts the letter, a queue that sorts the requests. Build it with the people who will use it, and keep a person in charge of the output.
Inside one company: Build the shared prompts and small tools that a whole department can reuse.
Where I have done it
The assistant on this site answers anything and knows the site when that helps. The exam-prep pipeline takes one item bank and produces every page a class prints.
Check the output
Models are wrong in confident sentences. Decide how you will know, before launch, and make the check cheap enough that people run it. A team that trusts the checks will trust the tool.
Inside one company: Pick a few tasks per team and agree on what a good answer looks like before scaling.
Where I have done it
I audited every pull request in this repository with two models and a rule set, then published where they disagreed. The benchmarks page re-checks its own scores in your browser, and the voice-fit meter flags prose that reads machine-made.
Make it stick
Adoption fades when the champion moves on. Name an owner, set a review date, and watch for the automation that quietly stops. A person still signs the hard calls.
Inside one company: Put AI use in the regular team review, and give each tool a named owner.
Where I have done it
The rollout story on the homepage carries six accounts from an old system into a new one, rules fixing what has one right answer and a person signing the two hard calls. The ops harness exists because a job once failed with nobody noticing.
03 If you are hiring for this
The Case Study Library has a lens for AI enablement that puts the rollout and teaching cases first. Or write to me and say what your team is trying to get people to use.