For leaders deciding what AI changes

How to reorganise your teams for the agentic era.

AI agents can now carry most of the production work of building a product: the coding, the testing, the drafts. The question for a leader is what to do with that: how to organise the work, staff the team, and keep a person answerable for every result. We rebuilt our own teams to find out, on our own internal projects, and measured what happened. This page is what we learned, and how to run the change in your own organisation.

Measured, not modelled

The speed gains were real, but the more interesting finding is what the freed-up capacity turned into.

5-7x

Faster to a working prototype

Prototyping compresses dramatically: from two to three people over a week, to one person over one to two days. Where the work was clear and bounded, AI took the execution weight and one person ran the loop end to end.

2-5

People per team

The size of team the work now needs. A handful of people can direct and check the output of dozens of AI agents, so teams reorganise around decision-making and quality control instead of headcount.

2-3x

Faster on a major refactor

A full refactor of an existing production application. This held when the team that owns the codebase drove the AI loop, with the right harnesses around the agents. The full story is further down this page.

What we believe

Three convictions our own work forced on us.

We expected the main lesson to be about speed. It turned out to be about structure, accountability, and adoption.

CONVICTION I

Speed is the easy part, the process is the transformation.

Most companies buy the tools, hand one to every person, and keep the process unchanged. That's AI enabled: the same handoffs, each step somewhat faster.
AI native means redesigning the process around the agents: fewer handoffs, teams organised around decisions, agents doing the production work in between. We've run both on our own projects; the second is the one that changed our team sizes.

CONVICTION II

Accountability never leaves a person.

However fast the agents produce, they don't own the result. The person who assigned the work does. Enterprise software carries consequences that reach beyond any one builder: the brand, the regulator, the customer. So every piece of work in our teams has a person's name on it, the checks before release stay as strict as ever, and the agents answer to a reviewer, never the other way around.

CONVICTION III

Adoption is the measure that matters.

The math tells you what's possible. It doesn't move a single person to work differently. For a company that already exists, the real transformation is getting your people to genuinely take up new ways of working, and that happens from the ground up, through the work itself. The number to watch isn't licences bought. It's how many of your people have actually changed how they work.

HOW THE ORGANISATION CHANGES

The work of a ten-person team consolidates into three people, their agents, and a set of guardrails.

Some companies hand everyone AI and keep the org chart: the same functions, the same handoffs, everything slightly faster. The map below shows the other path. Tasks that used to be divided among many individuals consolidate into a single person who has to be capable of all of them, the production work moves to agents, and the wider functions build the tools and keep the hard lines. Read it as a staffing map, because that's the decision it represents.

The five roles

The team that emerges

When AI carries the production work, narrow specialist roles collapse into broader ones. Three of the five sit inside every team. The other two serve many teams from the tools and guardrails layer.

01

"Is this worth building?"

The Product Owner

Decides what should exist, shows working demonstrations instead of writing documents about them, and stays responsible for the product after it ships.

02

"Will this hold up?"

The Tech Lead

Answers for the build. Used to spend the day writing code. Now spends it directing AI agents, reviewing what they produce, and designing the structure the whole system rests on.

03

"What do the agents work on tonight, and is their output good enough?"

The AI Engineer

The converging point for software engineers and data scientists. Masters the agentic development process: defines the work for the agents, sets up the harnesses they run in, and brings the technical judgment to know when their output can be trusted.

04

"Does everything we ship look and behave like us?"

The Design Steward

Instead of designing individual screens, builds the design standards and automated checks that let every team ship work that looks and feels like the company's products, even when AI produced it.

05

"What's failing quietly?"

The Quality Conscience

AI's mistakes rarely look like mistakes; they look plausible. This role exists to hunt for the errors that read as correct, before a customer finds them.

That's a standing change to how we staff live deliverables with real deadlines, with agents directed through the same project systems we use every day. It also changed what some of our people spend their days doing

When the work consolidates, the specialists don't disappear. They stop producing the work by hand and start building what everyone else works with. Our senior designer is the clearest example. She used to spend her days producing screen designs, one at a time, for a single team. She now builds the design standards, tools, and review checks that every product team relies on, and she advises those teams on their own design decisions. It is the same person applying the same judgment, except that it now reaches every team instead of one.

The measure we watched throughout wasn't output. It was whether people actually took up the new way of working, and kept it up once the novelty wore off. We ran the change from the ground up for that reason, with the people doing the work shaping how the agents fit into it, rather than handing down a mandate. Any company can buy the same tools we did. What decides the outcome is whether your people change how they work, and that is won one person at a time.

A simple rule for what to hand over

What your people keep, and what you hand to the agents.

A person leads

When any of these is true

The scope is vague. Nobody has decided precisely what the job is yet.
The context is messy. The requirements or the existing systems are hard to read, even for a person.
The stakes are high. A mistake would reach a customer, a regulator, or the brand before anyone caught it.

→

AI leads

When all three hold

The task has clear edges. The AI knows exactly where the job starts and stops, and has a test it can run against its own work.
The requirements and systems are readable. The AI can follow what exists and what's being asked.
Mistakes are cheap. An error stays bounded, in a sandbox or a pull request, and never reaches production on its own.

Run your own

Six weeks will teach you more than any presentation.

Start small and stay accountable

1Pick one well-defined part of your business.Use the rule above: clear edges, readable requirements, mistakes that are cheap to catch.
2Give one person real ownership of it.The result carries their name, including everything the AI produces. Keep approvals lean: one owner, at most two informed reviewers.
3Run it properly for six weeks.Normal systems, normal review standards, no special treatment. First two weeks produce speed and expose quality gaps. The next weeks build the checks that close them. The final weeks ship. When output falls short, fix the check that let it through.
4Measure adoption alongside output.The lasting question is whether the new way of working holds after the six weeks end.

Voices from the ground

What the team said after six weeks.

“

The first thing I noticed was how quiet the week got. Fewer meetings, fewer handoffs, and the work just moved. My job became deciding, and standing behind the decisions.

PL

Product Lead

“

I hand the agents a scoped task at the end of the day and review a pull request over coffee the next morning. My craft moved from typing the code to judging it.

EN

Engineer

“

We ran experiments overnight that we would never have budgeted time for before. The agents explored the options, and we decided which ones counted.

DS

Data Science and Delivery