More and more companies ask the same question: how do we adopt AI? Most of the time the honest answer is another question: are the foundations in place? First you need clean data, clear workflows, and measurable use cases, and only then AI. AI is just the layer on top.
I'm not saying this because AI is overrated. I use it almost every day, and I've written about it separately: how I use AI in practice. I'm saying it because I've seen too many projects where AI was bought before the foundations were in place. The result wasn't an intelligent system. It was the same mess, only faster and stated with more confidence.
AI speeds up what already exists
AI doesn't fix a broken process. It amplifies whatever is already there. If the process is clear and the data clean, AI multiplies a good thing. If the process is messy and the data scattered, AI multiplies a messy thing.
AI is an amplifier, not a repair tool.
A concrete example: a chatbot on top of an outdated knowledge base doesn't save time. It answers wrong, but does it convincingly. The customer gets a fast answer that happens to be wrong, and nobody notices until the damage is done. Before AI, a bad instruction just sat there. Now it talks.
Three things that have to be solid first
1. Clean data
AI is exactly as good as the data you give it. If the same information is scattered across emails, Excel files, and three different systems, and the versions contradict each other, AI doesn't know which one is true. It guesses. And the guess sounds certain.
A customer database where the same company appears four times under four spellings doesn't become reliable by putting AI on top of it. It becomes reliable by cleaning it first. Only then is AI any use.
2. Clear workflows
You can't automate or augment a workflow nobody has written down. If the work lives only in one person's head, AI can't learn it. It invents its own version, and that version is rarely the same as yours.
This has a useful side effect: when you write the process down before reaching for AI, half the problem often solves itself right there. You discover that two people do the same task differently, or that one step is a leftover from an old system and isn't needed anymore. AI wouldn't have found that. Writing it down did.
3. Measurable use cases
"Adopt AI" is not a goal. It's a wish. A goal is a use case with a number you can move: hours per week, error count, response time, time to first reply.
Not "AI for customer service", but "cut the first reply on 60 percent of inquiries from four hours to ten minutes". The first can't be verified. The second you either hit or you don't. If a use case can't be measured, you can't know whether it produced anything, or whether it was worth doing.
How do you know if the foundations are solid
Three honest questions are enough to start:
Do you know where your data is? One place, or ten places and two versions of the truth.
Are the important workflows written down? Or do they live only in people's memory.
Can you name one use case with a number? Or is AI still discussed in the abstract.
If the answers aren't there yet, that isn't a failure. It's a starting point, and a fairly common one. This is exactly what the IT Health Check maps: not whether you're fashionably "AI-ready", but where the data actually lives, which workflows exist, and what is worth improving first. Independently, with no product I'd be selling on top.
When the foundations are solid, AI is the easy part
Here's the good news. Once the data is clean, the process clear, and the goal measurable, AI is often the cheapest and easiest layer in the whole project. The hard part was never the model. The hard part was the data and the process.
When I automated my own sales pipeline, AI came last. First there was a clear process for who gets contacted and why, and clean data on who they are. Only on top of that does AI draft the messages. The same goes for automations more broadly: the machine prepares, the human decides, and both only work if the base is solid.
So the day will come when you need AI. Probably soon. Just not first.
Conclusion
When someone says "we should adopt AI", the right follow-up isn't "with which model". It's "for what, and on top of what". Almost always the answer reveals that the groundwork has to come first: clean the data, write the workflows down, and pick a use case with a number.
Foundations first. AI is the layer on top. That's also the order in which it pays off, and the only order in which it stays standing.
// Kalle Huttunen 1 Jun 2026