How we bring 150 people along on our AI journey

And the interesting mistakes we made on the way.

Mark Vletter · Voys · AI Grunn

"Just a few things before you start."

This is what we ask of a colleague who just wants to do their job.

Install the desktop app Create an account Pick a model Add skills Connect MCP servers Learn to prompt Mind the privacy rules Keep it all up to date

Everyone gets an account. That is not a strategy. That is a subscription.

We are Voys.

  • 150 colleagues
  • 33,000 customers
  • steward-owned
  • self-managing since 2012

We live by the experiment: make interesting mistakes and learn fast.

Mark Vletter at the Voys office

So we said: go experiment. And they did.

Projects everywhere.

Great ideas, and nobody could find them.

a server at home the product'sstaging environment a tools site someone's laptop

Then came the fight.

Staging exists to test our product. Suddenly it also ran everyone's internal tools. Both sides had a point.

Productplatform Internaltools

Running for months. Owned by no one.

Projects kept running unattended, in the wrong place, long after anyone looked at them.

And every tool looked different.

Every new tool invented its own layout and its own house style.

Aa Aa Aa Aa Aa

Building got cheap. Finding, running and maintaining did not.

Where do your people already ask questions all day?

In the chat tool they already use. Only there, they ask people.

So AI becomes a colleague in the channel.

It asks what a good colleague would ask: what exactly, how often, with which systems, for whom.

ColleagueCan someone check new tenders for us every morning?
AI teammateSure. Which sources? Who needs the result? And what makes a tender interesting for us?

Nobody installs anything.

Live today

You ask the way you would ask a colleague. You don't need to know which system does the work.

A chat thread in which a colleague asks how to connect a desk phone and the knowledge bot answers with four steps

One front door. Four kinds of questions.

1I want to knowsomething 2Do thisfor me 3Do thisevery time 4I need a toolof my own one-off and loose → always and fixed

1 · Knowledge

"I want to know something."

Our knowledge lives in Notion, written for people. For AI that turned out to be a mess. You have to prepare it first: AI information architecture. We use Klai for that.

Live today: our knowledge bot answers in Slack.

2 · Skill

"Do this for me."

A skill is a written recipe that teaches the AI how we do one job. If the skill exists, the chat helps you find it. If it doesn't, you build it together.

Claire, our code reviewer, started as one skill.

3 · Workflow

"Do this every time."

Fixed steps where the process is fixed, an AI step where it gets messy, a person where a decision is needed. It reports back by itself: "I did this for you. I have one question."

An admin wires up the connections once. Everyone else just talks.

Our content publication factory runs like this in n8n.

4 · Software

"I need a tool of my own."

For anything that stays the same, software is faster, cheaper and leaner than AI. In Garage you see your tool and ask for changes in a chat box. A coding agent builds it in the background.

Herald translates our entire help centre, and everyone can make changes. Garage runs in internal test. Live demo
Garage: a finance tool with the Garage assistant open, asking for a change in plain words

AI finds the route. Then we pave it, and we maintain it.

Think of a new colleague. The first time, they think about every step. By the tenth time, it's a checklist.

What runs behind the door.

Open source wherever we can.

Front doorSlack AI teammate andskills engineHermes: asks,routes, runs the skills KnowledgeKlai Workflowsn8n SoftwareGarage, with Symphonyrunning the coding agents Control roomPaperclip: every agenthas a role, an ownerand a budget ModelsClaude, and local modelson Ollama for private data HomeYolodeck

One of them, from the inside: Claire.

Claire reviews the merge requests our developers ask her to. She started as a single skill.

Merge request The gateplain code: run or skip? The agentHermes + Claude SkillsMarkdownfiles in git Memoryself-hosted The proxyholds thecredentials posts the review

Four choices we would make again.

  1. 1Code decides whether she runs. Nothing starts unless someone asked.
  2. 2Her knowledge is plain Markdown. Odd comment? Open the file.
  3. 3She never holds a password. Assume she'll be fooled.
  4. 4Rules have to earn their place. A new rule starts on probation and retires when people stop acting on it.

What Claire changed.

Comments per merge request 5.93.81.6 AugSepOct Share of her comments acted on 52%65%68% AugSepOct

About 98 engineer hours saved a month.

An estimate. We check it with a survey in December.

One place to run it all.

Yolodeck is our hosting platform for internal tools. Not the product platform. Not a server at home. One place, one login, one way to ship.

This is the piece we are still waiting for.
Yolodeck

Where we are today.

Every component is tested and works. Now we connect them.

Live Knowledge bot in Slack Claire, our code reviewer n8n, now managed by AI In internal test Garage Next Yolodeck

Paving means maintaining.

  • Everything has an owner
  • Problems go to a shared channel, not one inbox
  • Keys belong to the tool, not to a person
  • Unused for 90 days? Switched off
  • Look before you build
  • One design system, so every tool looks like ours

Technology is the easy half.

Four levels of AI fluency.

You don't jump levels. You aim one level up, in one part of your work.

LearningAI is a light assist,the work looks the same CapableAI is woven intoyour core work Builderyou build things yourcolleagues use too Redesignerwhole categories ofwork disappear

We learn it together.

  • A frontier member in every team
  • Garage sessions every other Wednesday: five-minute demos, before and after
  • Time you are allowed to spend on this

We don't count how often you use AI. We count what changed.

Mark Vletter working on a laptop

What about jobs?

Herald took over translation work worth two full-time jobs. Those two colleagues now help everyone else do more with AI.

More with the same people. Not the same with fewer.

Three things for Monday.

  1. 1Pick one front door: the chat tool your people already use.
  2. 2Sort the questions into four: know, do, repeat, build.
  3. 3Give everything an owner and a home.

Don't bring your people to AI. Bring AI to where your people already work.

Thank you.story.markvletter.com/ai-strategyMark Vletter, Voys