How to Roll Out AI in a Company: One Measured Workflow, a Few Champions, Not More Licences

Matthias Heim · 2026-11-16

Measure where one piece of work waits, find use cases by handing real work to AI for two weeks, and give a few people room to lead. With a client example from the IT team of a Swiss industrial group.

Roll out AI through one measured workflow and a few employees who want to learn, not through licences. Time where one piece of work waits today, and let a small group hand real work to AI for two weeks. Support them with a small team, internal or external, that tries things out ahead of everyone else and hands what works back to the teams.

In a Swiss industrial group, the head of IT service management developed and deployed six ServiceNow stories and eight hotfixes with an AI agent in the two weeks after a two-day workshop. Most of that development had previously gone to an external partner.

This article covers the three parts, how to find use cases from real work, and what leaders and regulated companies need to put in place.

In short

Measure one workflow from request to customer. Where the work waits is your real constraint, and it moves when AI speeds up one step.

Find use cases by handing real work to AI for two weeks and writing down where it fails, not by brainstorming in a workshop.

Give a few motivated employees (we call them champions) time and permission, supported by a small team that tries things out ahead of the others.

Why the Company Doesn't Get Faster When People Do

When people get faster with AI, the company often does not. The bottleneck moves to everything after building: review, release, handover, and customers who can only absorb so much change.

A feature that gets built in days still waits for review and release, and customers still need time to start using it. Speed in one step does not change how fast work reaches the customer.

Ask your leadership team where the bottleneck sits now, and you may well get several different answers. That is why we start by measuring one workflow:

Follow one piece of work from request to customer and time every step. Where it sits waiting is your real constraint.

Raise that workflow one level at a time: first someone pastes an email into a chatbot, then drafts are already waiting in the inbox, then the clear cases go out on their own.

Do leadership's part: time to learn, an owner for the handover from day one, and data rules by data class (public, internal, sensitive) rather than by tool.

We describe this shift in more detail in why leadership is the new bottleneck .

Find Use Cases by Delegating Real Work

Use-case workshops produce lists of guesses, made before anyone has handed real work to AI. They tend to miss everyday tasks and overrate ideas that fall apart on the first real case.

Delegating works better. I tried it on my own work first: I connected one AI agent to my email, calendar, CRM and files and tried to hand over everything that could be handed over. It started with email drafts, then follow-ups from meeting transcripts, then invoices. By the end of October 2026, 64 automations were running, and none of them had been planned in advance.

Each failure pointed to the next step, and almost always for one of three reasons: missing context, a missing tool or connection, or a decision that should stay with a person. In June, for example, an agent read a meeting transcript and moved a deal I had marked as won back to "proposal". Since then, won or lost is my decision, never the agent's.

For a team, the exercise looks like this:

Pick one person and connect one agent to the systems that person actually uses.

For two weeks, hand over everything that could plausibly be handed over, the way you would with a capable new hire.

Write down every failure and its cause: context, tool, access, or a decision that should stay with a person.

At the end, fix the most frequent cause first. One fix often unlocks several tasks at once.

Each failure names what to fix next: missing context, a missing connection, or a decision that stays with a person.

Give Champions Room

A champion is an employee who wants to learn, gets time to work with AI on real tasks, and then helps colleagues do the same.

Give them a small group, time to experiment and permission to skip the standard process for a while. Then help them: sit together, coach, and make early results visible. Let them work across role boundaries, because with AI one person can do work that used to be split between several specialists.

When it works, people like this bring their colleagues in. Buying licences takes a week. Growing these people takes longer, and it is what changes how the work gets done.

The ServiceNow example from the start of this article shows how it works. The IT lead who runs the agent is not a developer. He came from support. The agent reads each story and flags what is missing, proposes the change, builds it on the development system, and he approves every deployment. After the workshop he extended the setup himself, for example to log and close maintenance notices that arrive by email. His earlier AI assistant licence now goes largely unused.

Add a Small Team That Works Ahead

Champions move faster with a small team, inside or outside the company, that works ahead of them. It tries new approaches on real problems, builds working versions quickly and hands what works back to the people who own the work.

Give that team a route to production from the start: a named person on the receiving side who takes over what works, and a release path that does not take months. Otherwise working versions pile up unused, and you are back to measuring where the work waits.

What Leaders and Regulated Companies Need to Provide

None of this happens by accident. Where it worked, leaders provided these four things:

An owner for the handover from experiment to daily work, named from day one and with a deputy, so a tool does not depend on one person. In the ServiceNow example, the client's own IT lead runs and extends the tool.

Data rules by data class (public, internal, sensitive), decided once rather than tool by tool, and checked against the Swiss Federal Act on Data Protection, FADP (nDSG), before real personal data is involved.

Human approval for anything that reaches production or customers, with a record of who approved what.

A measure: one workflow timed from end to end, not the number of licences.

If you want a team working next to yours in this way, see how we work .

Frequently Asked Questions

How do you roll out AI in a company?
Start with one measured workflow and a small group of people who want to learn. Give them time, coaching and permission to work differently, and let them bring the next teams in. Licences come after that, not before.
Why doesn't AI make the whole company faster?
Because the bottleneck moves. When one step gets faster, work piles up at review, release, handover or with customers. You only see it by following one piece of work from end to end.
How do I find AI use cases in my company?
Connect an AI agent to one person's real work for two weeks, hand over as much as possible, and record where it fails and why. The failures show you what to fix and which tasks are ready.
Who should run the first AI experiments?
A few employees who want to learn, supported by a small team, internal or external, that builds working versions quickly and hands proven approaches back to the people who own the work.

How we work · Work · Insights