Start Where Nobody Worked

Why large companies won’t lead AI adoption, and the way in that skips most of the eleven steps

Most large companies will need more than ten years to adopt AI. The more I see of them from the inside, the more I agree.

The models are ready for a lot of the work. The road to production is what takes years.

Eleven steps to production

When a CEO decides to roll out an agent, that decision starts a long trip.

First comes an audit, usually a quarter long. Then a roadmap that has to cover the ROI, the data, an internal champion, security and legal. Then a pilot, a decision on whether to scale it, a readiness checklist and training for the people who will work with the agent.

Count them and you get about eleven steps between “the CEO decided” and “the agent is live”.

Any of them can stop the project. These three I see most often:

• The sponsor leaves for another company, and the project loses its owner halfway through.

• The data isn’t ready for an agent. It sits in several systems, the same field means different things in different places, and nobody owns cleaning it up.

• The agent never gets access to the systems it needs, because nobody can agree on who answers for what it does there.

Each step makes sense on its own. Together they turn a two-week experiment into a two-year program.

The part nobody says out loud

There is one more obstacle, and it rarely shows up in a project plan.

Picture the manager whose department the agent is supposed to automate. Their team, their budget and their place in the organization depend on the work the agent is about to take over.

So they find ten reasons the pilot isn’t representative. The sample was too small, the cases were too simple, the real work is more complicated. Some of those reasons are fair, which is exactly what makes them hard to argue with.

I’d call it sabotage only in the mildest sense. There’s no bad intent in it, just a normal reaction from people who feel threatened, and a rollout plan that ignores it stalls right where the agent starts to matter.

Why small companies go first

Startups and solo founders have one step instead of eleven. Try a tool in the morning, and if it doesn’t fit, drop it in the evening.

Their data isn’t scattered across three ERPs, and nobody’s job depends on defending the old way of working.

I see this in my own side project, where agents do work that would otherwise take six to ten hires. Decisions there take minutes, because I’m the only person who has to agree.

So for the next few years, most of the leaders in AI adoption will be small.

The way in

If you work inside a corporation and want to move anyway, there is a way in. It goes against the standard advice to start by automating routine work.

Leave existing processes alone. Start with tasks nobody has ever done, because doing them used to take a team of developers for six months or ten analysts for a quarter.

A few examples:

• Read all 4,000 supplier contracts and find the volume discounts nobody claimed and the late-delivery penalties nobody charged.

• Review every support request instead of the usual 3% sample.

• Combine three accounting systems into the one table with margin per customer that finance has been promising for three years.

I wrote earlier this year that with Claude Code, ideas that were never worth a developer’s time can now be tested in days. Inside a corporation the effect is the same, and the stakes are higher.

Why it works

A task nobody did has no “before”. There’s nothing to compare it with, so nobody asks for a quarter-long audit with a baseline.

Nobody owned it, so nobody has territory to defend. The team sees a tool doing something that used to be impossible, and it doesn’t feel like a replacement for them.

And the owner sees money the company never had. Unclaimed discounts and uncharged penalties are a much stronger argument than “the process got 15% faster”.

How I’d pick the first task

A few filters help:

• It touches money directly: revenue, cost, cash or penalties.

• The data already exists somewhere, even if it’s messy. Contracts, tickets, invoices.

• One person can check the result in a day. If finance opens ten contracts and confirms the discounts are real, the case is made.

• Nobody’s current job is to do it.

What happens to the eleven steps

The eleven steps exist for initiatives that climb up from the middle of an organization, collecting signatures at every level.

When the owner has seen the result personally, the steps stay. Security, legal and training still happen, but they take weeks instead of years.

Nobody has to persuade the owner with a slide deck anymore. They saw it work, and they’ll push the project themselves.

The fastest way into a large company runs through the work nobody was doing.