Prefer listening over reading? 🎧 Brief (01:32 min) for the key ideas, or Deep-dive (12:36 min) for the full discussion. (AI-Generated Audio Versions)
Agents in the Middle, Judgement at Both Ends
A rule of thumb that circulates widely in conversations about AI is the 20-60-20 rule. A human does the first 20% of a task, AI does the middle 60% and a human does the last 20%. It is not a formal framework and the numbers are a mnemonic, not a measurement. Still, it keeps appearing independently, including in research from King’s College London on professional services firms, where many interviewees described using it in practice.
What interests me is what it looks like when a product team actually runs this way.
If AI does the middle of the work, the value of a product team moves to what happens before it and after it.
What it looks like in practice for product teams
Some practical examples from my teams. A product team uses agents to create and manage features. As a result, the team spends more of its time talking to stakeholders and gathering requirements, and less on producing the artefacts that follow. Once the agents have done their part, the team reviews every feature and aligns with Engineering and other stakeholders, e.g. on dependencies with other teams.
The same is true for communications and announcements. Agents draft the posts, and the majority of the team’s time goes into planning what needs to be said and managing stakeholders afterwards.
In both cases the middle did not disappear, it became cheap. What changed is where the team’s calendar goes.
The same shape across product work
| Activity | First 20% (human) | Middle 60% (AI) | Last 20% (human) |
|---|---|---|---|
| Feature definition and management | Stakeholder conversations, requirements, constraints | Create and manage the feature | Review it, align with Engineering and other stakeholders, resolve dependencies on other teams |
| Comms and announcements | Plan the message and the audience | Draft the posts | Manage stakeholders as it lands |
| Research synthesis | Choose the questions and the sources | Cluster notes, draft themes | Decide which themes matter and which are noise |
| Prioritisation | Set the criteria and the context | Score options, model scenarios | Make the call and defend it |
Notice that the human columns are all about context and accountability, and the AI column is about production. None of the human work is typing.
The first 20% is the job, not the preparation
It is tempting to read the first 20% as “write a good prompt”. I think that undersells it. The first 20% is knowing which problem is worth solving, for whom and what good looks like, and that knowledge lives in people, not in any model.
I made a related argument in Faster to Build, Harder to Get Right, where I argued that AI makes building faster but does not make deciding easier. The 20-60-20 rule is the same idea applied to a single piece of work. If the input is vague, an agent will return something fluent and wrong, which is worse than something obviously wrong, because nobody stops to question it.
The last 20% is alignment and accountability, not proofreading
Reviewing the output is part of the last 20%, but it is the smaller part. In AI Mistakes Should Not Be Allowed to Travel I argued that a person approving outputs all day turns review into approval by habit, and that the checks that can be automated should be.
What an agent cannot do is the rest of the last 20%. When a product team reviews a feature, the harder questions are whether Engineering agrees with it, what it depends on in other teams and whether the people it affects are on board. For communications, it is landing the message and managing the reaction. An agent can produce the feature or the announcement, but it cannot be held accountable for how either is received. The human share of the work is where the judgement, the relationships and the accountability sit.
What is still hard
The first 20% can wither. Writing a spec or a plan is often how a product manager works out what they think. If agents produce those artefacts, the thinking has to happen somewhere else. In the teams I lead that place is the conversations with stakeholders, but I am watching for signs that the thinking is getting thinner, because a polished output can hide an unexamined problem.
The percentages are not constant. A routine stakeholder update may fit 20-60-20. A decision about strategic direction is mostly human, and a regulated communication needs far more than a final 20%. The right split depends on the cost of being wrong, not on the rule.
It describes one task, not the job. Product work is a loop of discovery, delivery and feedback, and the hardest part of it, getting people aligned, is not on the list at all. The rule is useful for deciding where to put human attention within a piece of work. It does not tell you what work to do.
Conclusion
When the middle is cheap, a product team is no longer measured by how much it produces. It is measured by how well it frames the problem before the agents start and how well it owns the result after they finish.
The middle 60% is where AI saves time. The judgement at both ends is where the product team earns its place.