I have sat in a lot of rooms this past year where the mood was the same. Someone built a prototype over the weekend. Someone got ChatGPT to write a PRD. A developer got Claude Code to do in three hours what used to take a week. People are buzzing.
Then I ask: what are you doing differently as an organisation?
They are in phase 1. Everyone has a subscription, everyone has had a few good experiences, and nobody has the same setup as the person next to them. It is 200 individuals who have each found a shortcut. The organisation runs the same engine it did in 2023.
Three phases
Phase 1 is inspiration. AI is a personal tool with everyone’s own prompts and tricks, and the gain disappears the day that person changes jobs. Management measures how many people have logged in.
Phase 2 is structure. AI is a shared tool with shared instruction files, skills and context, and you have decided where in the product lifecycle it is used. Management measures whether decisions get better.
Phase 3 is agentic. Agents take whole tasks from end to end, from discovery to code. People set the direction and stand behind the result.
Most Danish product organisations are in phase 1. Many believe they are on their way to phase 3, because the developers code with agents.
But agentic coding in a team that otherwise works like it is 2023 is a faster engine in the same car.
The jump to phase 2 is what I mean by structure and governance. It is about culture, organisation, skills, processes and tools. In that order. The five layers from the playbook, each with the questions you need to be able to answer.
The five layers are described in full in the playbook on AI product leadership, which you can read for free.
Under each layer are four questions. Tick the ones you have a shared answer to, and the tally at the bottom shows where to start. The ticks stay in your browser.
Layer 1: Culture & identity
In phase 1 most people use AI in secret, because nobody has said whether it is okay. I have met PMs who write their PRDs with Claude and polish them so it does not show.
In phase 2 you have said out loud what AI is for you, in a couple of sentences everyone can repeat. And you have said what people stand behind when the machine writes the draft.
- What is AI for us, and have we said it out loud?
- Who is responsible for a result that AI did most of?
- What do we reward: building more, or choosing more sharply?
- What do we do the first time an agent sends a wrong email to a customer? Does it lead to learning or a ban?
Layer 2: Structure & organisation
In phase 1 the organisation is unchanged. AI is applied on top like a coat of paint.
In phase 2 someone owns the AI practice. Otherwise you end up with 14 ways of writing an instruction file. The playbook’s three areas of competence, generalist, architect and compliance, need to be covered, also when it comes to AI.
- Who owns our shared setup: instruction files, skills, context, access? “Everyone” means no one.
- Who has an overview of what is being built with AI across teams?
- Who says stop, and is that person involved when the setup is designed, or only once something has gone wrong?
- If a team of eight can deliver the same as twelve, what do you do with the four?
Layer 3: Skills & practice
In phase 1 the skill is “can prompt”, and it is not measured. The courses are about tools.
In phase 2 you have named the skills that count: writing, context curation, judging output, systems thinking. Context curation matters most. In phase 1 everyone writes their own context every time. In phase 2 it lives in one place, maintained by someone. That is the difference between one good output in ten and eight.
- What should everyone be able to do, and what should only some be able to do?
- Who can judge an output they did not make themselves, without the senior people drowning in reviews?
- What is a good instruction file at your company? Without an answer there are 14.
- Who maintains the shared context? An instruction file nobody has touched in six months is a lie the agent believes.
Layer 4: Processes & workflows
In phase 1 AI is used inside the existing steps: write the user story faster, write the test faster. John Cutler has put it most clearly: AI accelerates the patterns you already have. If the process was broken, it is now broken faster.
In phase 2 you have mapped the product lifecycle and decided step by step: do we use AI here, how, and who approves. This is also where phase 3 begins. An agent can take a whole step once the step is described precisely enough. An agent that finds patterns in customer interviews is just as agentic as one that builds a feature.
- Which steps do we use AI in, which do we deliberately leave alone, and why?
- Where does a person need to make the call, and where can the machine just run?
- Who has the mandate to say “we investigate first” when the CEO shows up with a weekend prototype?
- Which steps could an agent take from end to end today, if we described them well enough?
Layer 5: Tools & software
The layer most people start with, and the one that matters least.
In phase 2 you have chosen a few tools and learned them thoroughly. More important: your context, skills and evaluations live somewhere you own, so you can switch model or vendor without starting over. Choose principles, switch tools.
- Where does our shared context live, and who can change it?
- Which data may AI see? Has that been decided or assumed?
- What happens if we switch vendor tomorrow?
- How do we know what the agents are actually doing?
The order is the point
Most people start from the bottom, with licences. That is why they are stuck in phase 1.
Start at the top. Say out loud what AI is for you. Decide who owns the setup. Name the skills that count. Draw the lifecycle and decide step by step. Only then choose tools.
It takes a couple of weeks with the right people in the room. The answers do not need to be perfect. They need to be shared.
The agentic product organisation comes from doing phase 2 properly. When the context is shared, the processes described and the responsibility placed, it is a small step to let an agent take a whole step.
This is the work I find most interesting right now: helping an organisation go from 200 good experiences to one way of working. If you are in one of those organisations, write to me.
