When a company decides to adopt AI, the first question is usually the same: where do we start? The answer depends on where the company is right now. Where people haven’t really tried AI on their own work yet, it’s too early to set targets: there’s nothing to measure. Where employees already bring tasks of their own, you can’t do without targets and measurement.
We see three stages: try it, build the environment, measure. Each has its own job, and things go wrong when the tools of one stage get used at another.
Three stages
Stage 1. Try it. There’s one job here: getting people to start trying AI on their own tasks. Left alone, only a few enthusiasts experiment, and for everyone else it never starts, so this stage is something the company has to organize. What it takes is access: a work account, no personal cards, no workarounds, and a simple rule about what data can go in. That’s infrastructure too, just the minimal kind. No targets and no KPIs at this stage.
The result of the first stage isn’t measured in hours. You see it differently: who got hooked, which tasks people keep coming back to the tool for, where it helped and where they gave up on it.
Stage 2. The environment. Once people have tried it, you can seriously decide which internal documents and systems to connect the model to, how to set up a sandbox where an employee can build a prototype for their own task, and who will maintain all of it. At the first stage this conversation would be guesswork. Now it’s clear what to build for: people are bringing the tasks they want to hand over to AI.
Stage 3. Metrics. Once there are tasks people regularly handle with AI, those can and should be measured: how many times a month the operation happens, how long it takes, how many people are involved, before and after. These tasks are also the best place to pick a process for a pilot. Now a target has something to stand on: there’s a baseline, there’s a habit, and the employee knows that what’s being measured is work they brought in themselves. We’ve covered how to set up that measurement and work out payback in AI in Business: How to Adopt Without Burning Your Budget.
The typical mix-up: a target at stage one
A company where hardly anyone has really tried AI often starts straight at stage three: it picks a process, sets a number, launches a pilot and measures before and after. The logic is clear enough: management needs to know what it’s paying for.
But a target assumes people already understand why they need AI in their work. At the first stage, they don’t. Some have no work access at all, only a personal subscription paid for on their own card. Others have access, but AI stays a browser tab they open on special occasions. And almost nobody has yet seen where the tool saves them an hour and where it confidently produces nonsense. A target handed down from above stays someone else’s number.
Say a department is told to cut the time it takes to prepare a sales proposal by a third. Nobody can justify that figure: management doesn’t know what AI will do for this particular department, neither do the employees, and the number comes from other companies’ case studies. In the first weeks the measurement will mostly capture learning time, since people are only getting used to framing a task and checking the answer.
The outcome is predictable: the target is missed, the report looks bad, and the company concludes “we tried it, it didn’t work”, when it never really tried.
How to tell which stage you’re at
Three questions are enough, best answered in order.
Do employees have work access to AI? Work access means paid for by the company, no personal cards, no workarounds, and a clear rule about which documents can go in. If there’s no such access, you’re at stage one, even if a few enthusiasts have been using their own subscriptions for ages.
Do they actually use it for work? Ask a few people from different departments where AI helped them last week. If you hear specific tasks, and people start complaining that the model doesn’t know your templates, documents and systems, it’s time to move to stage two. If the answer is “I’ve tried it a couple of times”, you have access, but stage one isn’t done yet.
Are there tasks people already handle with AI on a regular basis? Ones that get done every week, and by more than one person. If there are, and you can say how long the operation used to take, you have something to measure. That’s stage three.
A company is rarely at one stage throughout. Sales may already be bringing tasks while accounting hasn’t tried anything yet. Then the answer to “where do we start” is different for each department.
How to try it without building infrastructure up front
The first stage has an awkward spot. For people to try AI they need work access, but it’s too early to build full infrastructure for that: nobody knows yet what will be needed.
For this stage we’re building our own environment, Worklane. It’s a workspace in the browser. An employee signs in by invitation, describes their task, attaches materials and hands the work to an AI agent, watches what the agent does, checks the result and saves it. No personal subscription needed: the models, tools and limits are set up by whoever runs the environment in the company. What you get is a sandbox you can launch for a department or the whole company before deciding what your own setup will be.
Right now Worklane is a working prototype, and it hasn’t been tested with people outside the team yet. We’ll write about how the first test goes.
Where this breaks
A stage without targets doesn’t mean a stage without an owner or a deadline. Someone has to watch who’s trying it, collect what people bring, and decide when it’s time to move to stage two. Without that, “we’re trying it” drags on forever and quietly fizzles out.
It’s hard for management to pay for access without a promised result. That’s the honest cost of doing things in this order: the first money buys not results but the chance to find them.
Not everyone will try. Lean on the people who got hooked; the rest catch up later, watching their colleagues.
When someone asks “where do we start with AI”, the more useful answer is a question back: where are you now? Targets, pilots and measurement aren’t going anywhere. Each of them just has its own stage.