Sonntagsausgabe · Part 11
Where does your company stand on the AI ladder?
“We are doing something with AI already.” I hear that sentence in almost every first conversation, and it is almost always true. It just says nothing about where the company actually stands.
Because between “a few employees type something into ChatGPT” and “the AI handles work that used to take someone half a day” there are three stages nobody skips. Know them, and you immediately know what the next step is — and what you can save yourself.
Here is the ladder. Five stages, each with its tell-tale sign and the first step. No jargon you would have to look up.
Stage 1: AI-curious
How you recognise it: Individuals use ChatGPT or similar in the browser. Whoever does it, does it alone. There is no rule about which data may go in, and nobody knows how many accounts are in circulation. Asked what the company saves, you get an anecdote, not a number.
What is expensive about it: Not the subscription. It is the scatter. Customer data, quotes and contracts end up with vendors nobody has vetted a contract with, and the knowledge of what works stays stuck with individual people.
The first step: One page of house rules. What may go in, what may not, which tool is the official one. One page is enough, and it is written in an afternoon. After that you know where you stand — and people stop hiding what they do.
Stage 2: AI-ready
How you recognise it: The homework is done. Your programs have interfaces, so another program can operate them and not just a human with a mouse. Filing is ordered well enough that a machine finds what belongs together. There is one login for everything instead of twenty passwords. And your data comes out in full when you need it.
The uncomfortable part: At this stage the AI could work — but it hardly does yet. That feels like standstill, and this is exactly where most give up, or buy a tool instead that does not replace the groundwork.
Why it is still the most valuable jump: It gets by without AI. Everything you do here pays off even if you never deploy an agent: fewer passwords, findable documents, programs that talk to each other. In our own review, tools we had used for years failed this test — not because they were bad, but because they could not be driven.
The first step: Take your most important program and ask the vendor a single question: can another program operate this? The answer sorts your toolbox faster than any consultancy.
Stage 3: AI-assisted
How you recognise it: An agent handles first real tasks, and a human checks every single one before it counts. Not “the AI writes draft text”, but: it fetches receipts from portals, names them, files them and reports what is missing. The human looks it over and says go.
What you learn here: For the first time you see what the thing can really do — and where it fails. Usually not at the AI, but at a missing rule for the doubtful case. Anyone who discovers during the half-year closing that nobody ever defined how a receipt should be named does not have an AI problem.
The rule that helped us: Pick a first task where a mistake costs nothing. Collect, name, file, report what is missing. Booking and signing stay with the human. For us that was the half-year closing: 34 receipts from three portals, gathered and filed in 90 minutes, and the human share was checking one mail and saying go.
The first step: Write down for one week what you did the same way more than three times. Whatever sits at the top is your first agent.
Stage 4: AI-integrated
How you recognise it: The AI no longer sits beside the processes but inside them. And everything it does is recorded traceably: who initiated what, when, on what basis, who approved. Not as a screenshot, but as a log that is still readable a year later.
Why this stage is a different league: From here on, auditors, trustees and customers may read along. That is the line between “we are trying something” and “this is how we work”. For regulated industries it is the only stage that counts at all.
The first step: Take the stage-3 process that has proven itself and answer three questions in writing: what gets recorded? Where? Who can read it without asking anyone?
Stage 5: AI-native
How you recognise it: New processes are designed from the start so that a machine can run them. Nobody builds a process that only works with a mouse and a screen and postpones automation to later. The question “can an agent do this?” is asked while designing, not afterwards.
What changes: The company does not get faster because it works harder, but because the work is cut differently. Tasks that used to hang on people now hang on rules — and rules can be improved, lent out and reused.
The first step: One line in every new process: who runs this, human or machine? If the answer is “human”, it needs a reason.
The jump that pays the most
It is not the last one. It is the one from stage 1 to stage 2 — and it gets by entirely without AI.
On stage 1 you buy tools and hope. From stage 2 on you decide. Everything after that is a question of sequence and patience, not of budget. Skip the groundwork and aim straight for stage 3, and you have bought an agent that stops at a login screen.
And one more thing that rarely comes up in sales conversations: on every stage it is fine to stay as long as it carries you. Not every company has to reach stage 5. But every company should know which stage it is on.
We worked the ladder on our own company, in weeks rather than years, and every stage broke something that had looked fine before. That is exactly the point.
Which stage are you on — and what is the one step that takes you to the next?