The 5 Stages of AI Process Maturity

Everyone Is Selling the AI Operating System. Almost No One Asks About the Road There.

Every pitch paints the same picture: the company as an AI operating system, where processes run end to end through agents, systems synchronise themselves, and humans step in only at the few points where their judgement truly matters. That picture is not wrong. It is a legitimate end state. But it leaves two questions unanswered — the two that come first for any organisation: What are the sensible steps toward it, and which of them is worth pursuing today, technologically and economically?

Anyone who does not answer these questions for themselves adopts the answer from the pitch. And that answer is almost always the same: more AI, more automation — the more, the more mature. But the equation does not hold. Because “more AI” almost never means “more autonomous process” in practice. It means one more AI assistant, triggered by a human at the click of a button, its output reviewed and carried onward by hand. The amount of AI grows — the process behind it stays the same. That is precisely why the question “How much AI are we using?” leads in the wrong direction. The right question is: What is actually happening to the process?

The Wrong Yardstick

Common AI maturity models measure maturity by the degree of AI autonomy. For industrial B2B companies, that is the wrong yardstick. What matters is not how independently the AI acts, but how mature the processes and data are in which it is meant to work.

You can recognise this in an unremarkable detail. Today, the human is the manual integration layer between the systems. They hold the process transitions together in their head and in private Excel and Word lists: which enquiry is still waiting on a technical clarification, which quote needs following up, where a delivery date is not yet confirmed. This glue work sits in no system — it lives in the heads and to-do lists of individual employees.

AI maturity does not mean giving the AI more autonomy. It means moving this manual glue work back into the systems, step by step. The best indicator of an organisation’s maturity is therefore a simple question: How much private to-do list does the employee still need?

This reframes what the model actually measures. It is not a maturity model for the technology. It is a maturity model for the company — for its processes, its data, and its organisation. That is what sets it apart from the widespread, technology-centric models.

The Five Stages — a Path, Not a Ranking

The road to the end state can be described in five stages. They are not a judgement, but a map.

1. Stage 0: AI is used ad hoc — individually, without system connection. The sales rep has a reply email drafted and copies it out; value emerges by chance and depends entirely on the individual.

2. Stage 1: AI becomes a tool. Individual tasks are deliberately assisted — drafting, researching, summarising. The AI turns a customer enquiry into a structured requirements list and a first draft quote; the transfer into the system, the configuration, and the pricing remain manual work. This is the reality of most industrial companies today.

3. Stage 2: AI becomes a colleague. An agent is given a task, knowledge, tools, and rules, and handles a defined, recurring process largely on its own — with human approval at defined points. A quoting agent extracts requirements, loads history and terms, produces a valid configuration with a price indication, and prepares the quote for approval. For most processes, this is the economically optimal point today. And the jump to this stage is not a question of AI, but of process: an agent can carry a task on its own only once that task has been made clearly defined, decidable, and handover-capable. That is exactly why Stage 2 is a sign of the organisation’s maturity — not the technology’s.

4. Stage 3: several such agents interlock — one process triggers the next. This is where the manual integration layer is digitalised: the transitions between processes. The quoting agent hands off to a clarification agent when something is technically unclear, aligns dates against availability, and triggers order creation once the order comes in. This is the real stress test; few organisations reach it cleanly.

5. Stage 4: is the end state being sold: processes are rebuilt so that the agentic path is the norm, not an island solution alongside manual work. What matters is the insight that this is an emergent end state — not a target to leap to. For many processes, it is neither necessary nor economical.

Anyone who knows this map can place a pitch: what in it is Stage 2 reality — and what is Stage 4 brochure?

Human-in-the-Loop — Rightly and Wrongly Understood

Here lies a tension that is rarely stated openly. You want to remove the manual glue work — and at the same time you build systems that keep the human constantly in the loop. If the system demands confirmation at every step, the manual work has not disappeared. It has returned as monitoring work. The new to-do list is then called the approval inbox.

Human-in-the-loop is right — but only asynchronously, and confined to genuine decision and approval points. Maturity means keeping the human at the few places where their judgement counts, not as permanent supervision over a process no one trusts. A system that requires constant confirming and monitoring is not a sign of maturity. It is relocated manual work.

Assessing Processes: AI Gap or Digitalisation Gap?

Before an organisation buys AI capabilities, it should clarify a different question: Which gaps does AI actually close — and which are simply neglected digitalisation and process gaps that would have had to be closed without AI anyway?

The distinction decides success. A rough rule of thumb: if a process step is already digitally defined today but not automated, it is a digitalisation gap. An agent placed on top of it only papers over the omission — and gets deployed at a premium for something that classic automation solves more cheaply and more reliably. If, on the other hand, the step requires judgement, interpretation, or the linking of context that cannot be captured in fixed rules, it is a genuine AI gap — that is where the value arises.

It is exactly for this assessment — which capability is really missing at which stage, and where to start first — that we developed a structured format. It sorts out where an organisation can already start today and where a foundation in data and process transitions has to be built first. Not a blanket “data first, then AI” — but a process-by-process classification.

First the Position, Then the Pitch

The AI operating system is a realistic target picture — but as a direction, not the next step. The road there runs through an honest assessment of where you stand: where the organisation is today, which vision fits it, which strategy holds over time, and what each step means for the operating model. Only then can you judge which technology and which vendor fit.

The order that prevails in everyday practice is usually the reverse. Yet perhaps the most honest assessment of where you stand is close at hand: how many private to-do lists are still running in your sales team today?