I’ve found myself cast as an AI adoption expert, in a field where nobody knows yet how to adopt AI properly. But ten AI systems I’ve built and rolled out in five months have taught me a few things.
So here’s my 14-step playbook. It’s where to start if you join an organization and your goal is AI systems with a measurable business result. It’s also how not to repeat the usual fate of AI projects, a demo everyone liked and nobody uses.
Study the existing processes.
Together with the people who own those processes, look for places where a small change pays off fast.
Break the chosen process down into nodes: every single action, decision and hand-off it consists of. A node is one piece of work with an owner, with what goes in and what comes out, with the data it draws on and with the number it’s accountable for. If you can’t fill those in, you haven’t broken the process down far enough.
Draw the stretch of the process where AI can be built in as one diagram: every node in it, whether a person or a model does it today. Every node passes through three states: people do it by hand; AI assists and a person drives; an agent runs it end to end. The diagram shows where each node stands now, where you’re taking it and in what order. Below is mine for a content process: 24 steps, where they stood in June, where they stand now, and where they’ll be in three months.
Then look past this project and draft the company’s automation roadmap. Mine is a map of initiatives, and there are more of them than I can build at once. Every initiative on the map has four things: the data sources it needs, an owner, the KPIs it moves, and the project that delivers it. The systems that collect data go on the map first, because everything else stands on them: without good data there is no good agent. Which data sources each initiative needs is the bottleneck, and it’s my criterion for what to build now and what to build later: if the data isn’t there, the initiative that collects it comes first. I keep this map inside an agent that works as my second brain. It holds the initiatives, the data sources behind them and what the company already knows about itself, and I ask it before I decide what to build next. And a node you build once serves the next project: the node that writes text for the website will write sales battle cards, social posts and promos.
Decide what the model may judge on its own and what stays a human decision.
Design a solution for a result the business needs. It can take one of three forms: faster work, higher quality, or something that wasn’t possible before. Two is better than one.
Decide how you’ll measure the effect. I usually prefer one adoption KPI and one business KPI: the number the process exists to move.
Work out with the team how the solution fits into their daily work. The last mile is the hardest part, and it’s where most AI projects die.
A tool is adopted one person at a time, so I read each person as a set of sliders: can they use it, do they want to, is there room for it in their week, do they expect it to help with their own work, and has it been bent to the way they work. The lowest slider decides. That’s why training alone changes so little: it moves one slider and leaves the rest.
Read the sliders on the person’s own task, with the tool in front of them. That hour shows which one is at zero, and usually moves it. The company sets the defaults for everyone, by saying out loud that this is wanted and by using the tool visibly from the top. Then count who’s still using it a month later. Nearly everyone tries.
Write the intent: one page on why this system should exist and what result it owes the business. The model drifts during a long build, and so do you. The intent is what you both come back to.
Turn the diagram from step 4 into the architecture map. For each node: the actor (a model, a script or a person), what comes in and goes out, where it runs and where the data lives. I don’t write the code and I don’t read most of it. I read the map, and when the code and the map disagree, I decide which one is wrong.
Collect the evals. Tests used to be deterministic: the same input, the same expected output, pass or fail. An AI system answers the same input differently every time, and every answer looks right, so the test has grown two more layers. Code still checks what’s mechanical: the format, every number against its source. A second model checks what takes judgment, one yes-or-no question at a time. And people check that model, because its verdict is also an answer. The cases come from real work: an input from the process and the answer a good employee would give, written before the build, and every failure in production joins them.
Build it. The model writes the code, and reading all of it isn’t your job. Your job is the three documents you wrote by hand and understand completely: the intent from step 10, the architecture map from step 11 and the evals from step 12. The code is checked against all three, and it changes until it passes.
Watch whether it meets its goals and whether it’s worth what it costs to run. If it doesn’t, switch it off and go back to step 1. If it does, scale it.
Building AI systems keeps getting new names, from prompt engineering to context engineering and now to context architecture and harness engineering. All of them are useful to cover steps 10 to 13.
But the hardest part is still finding exactly where the company is losing time or money, then designing a solution and fitting it into the process so that it delivers. With AI, this takes one more skill, the one from step 6: knowing what the model can be trusted to decide and where a person has to stay in the loop.