AI writes the plan in four minutes.
The plan was never the problem.

A target state, a roadmap covering eighteen months, a communication plan including a stakeholder matrix: what used to be two weeks of consulting work, any language model now delivers in a lunch break. Nevertheless, transformations keep failing – and at the same points as ten years ago.

The figures on this are new and unambiguous. In Publicis Sapient’s Global Enterprise AI Report, surveyed between the end of April and mid-May 2026 among 1,550 AI decision-makers in six countries, 73% say they use AI regularly or in most business processes. Only 10% describe it as the core of their business operations – and in Germany, likewise, only 10% say AI is fully integrated across the company. 42% admit that their organisation is not built to realise the benefits.

This is not a technology gap. It is an execution gap – and it decides whether an AI transformation actually lands in the organisation or stays in the toolbox.

The German-speaking banking market looks no different. According to an analysis by zeb for Austria, Germany and Switzerland, adoption rates in large companies are 68% in Austria, 65% in Switzerland and 57% in Germany – above the EU average. Anyone who concludes from this that these institutions are further ahead than others is confusing adoption with impact.

What AI has really taken off our hands

Looked at honestly: quite a lot – namely the part that was expensive and worth little.

Analyses that used to keep three consultants busy for four weeks. Benchmarks. First process maps. Variants of a target state, set against one another. Minutes, status reports, presentations. Anyone still paying day rates for this today is paying for a craft that no longer exists.

It should be possible to say this out loud, even as a consultant. A considerable part of classic consulting work has just become worthless. Anyone who disputes this has not tried it.

Where things are still stuck – three points

Before I started in Vienna in 2005, the same institution had seen four managers in three years. Each brought a programme along, each programme had a clean concept, none was seen through to the end – the next one had already been transferred before the uncomfortable decisions came up. The workforce had made the most sensible thing of it that one can: they nodded politely and waited. Not for the programme. For the transfer notice. What came of it is described in the case study on the branch turnaround.

First: decisions. A model can set three options cleanly against each other, with costs, risks and dependencies. It cannot walk into the room where two division heads have been postponing the same question for six weeks and bring about the decision. Programmes rarely fail because of the concept. They fail at the points where nobody decides.

Second: commitment. A roadmap is a claim about the future. It only becomes true when someone demands commitments, follows up on deadlines and names deviations while they are still small. That is uncomfortable, it costs relationship capital, and it cannot be delegated – neither to a tool nor to a status meeting.

Third: accountability. If a decision was wrong, someone is liable for it – professionally, financially, in regulatory terms. A tool is not liable. Nor can it explain four weeks later why it advised one way and not another at the time.

What I take on – and what I do not

What I take on in a mandate is not a recommendation but accountability for results. I go into institutions where a decision is due and someone has to make it – with a mandate, not with a mandate proposal.

What I do not take on: roles that are unclear or not backed. Figures that are supposed to look better rather than become better. And mandates in which the cheapest day rate decides in the end – not out of vanity, but because experience shows that in such institutions nobody makes the uncomfortable decision after all. How to tell that I am the wrong person is therefore set out openly on the interim management page.

What follows from this – in practice

In an AI transformation, the question is no longer whether an institution uses AI. It is: who makes the decisions that the tool has prepared?

A division of labour that has proven itself:

1. AI handles the preparation. Analysis, variants, documentation, follow-up. Everything that is diligence and needs no authority.
2. A human makes the decision. Named, with name and date – not as a committee.
3. A named function accepts escalations. A short route upwards, no further status round.
4. Whoever makes the decision must be able to explain it. Without reference to the tool.

Anyone who does not fill points 2 to 4 gets exactly one thing from AI: piles of paper produced faster.

I work this way myself. What I use artificial intelligence for in a mandate, what I expressly do not use it for and who bears professional responsibility is set out openly on a dedicated page. Anyone who demands straight talk owes it as well.

The point

Transformation was never a knowledge problem. Execution means: deciding, committing, following through, standing by it – the four activities a language model cannot take over.

It is welcome to do everything else. I let it, too.