A shift is taking place in companies right now that hardly anyone speaks about openly. In mid-sized companies, in marketing departments, in strategy meetings, language models are being used for tasks that used to be outsourced: positioning papers, competitor analyses, market strategies, communication concepts. Sometimes as a conscious decision, but often more in passing: a CEO opens ChatGPT, a marketing manager consults Claude, a strategy officer asks Gemini to formulate three variants of something. What emerges is rarely declared as 'AI output'. It is presented as an original concept. Sometimes it is edited, and presented in committees as though it were the result of human strategic work.
... If everything is watered down, what’s left for us to grapple with?
— Steven Broschart
This practice is not inherently wrong. It is economically understandable, because classical strategy consulting is expensive, often slow, and inaccessible to many SMEs anyway. An AI delivers in hours material that once cost weeks and five-figure sums. That is genuine democratisation of expertise. Anyone who rejects it outright denies themselves real and substantial efficiency gains.
Yet the approach is characterised structurally by something that hardly comes up in most discussions - but it determines whether the efficiency gain becomes a strategic advantage or simply an expensive self-deception.
What Strategy Actually Is
Before we discuss what AI can really achieve in strategy development, we should briefly clarify what strategy is at all - because its common usage obscures a central aspect.
Strategy is not knowledge. Knowledge about your own market, about competitive position, about your own strengths is a necessary precondition, but is not itself strategy. Strategy is also not analysis. A SWOT analysis, a market study, a competition matrix describe the situation without making it a strategy.
Strategy is fundamentally something else: the holding of tensions and the subsequent positioning. Someone who wants to be market leader in the premium segment cannot simultaneously be a mass-market provider. Someone who bets on speed cannot build on perfect quality. Someone who prioritises a particular target group consciously neglects another. Real strategy therefore demands a willingness to live with conflicts, to reject options, to decide for something - not against, but for forgoing alternatives.
Yet strategy does not end with a single decision. It is not only the positioning, but also the path through the tension toward it. Because positioning is only where the actual work begins: defending the strategy over months and years against constant pressure to soften it. Every major customer asking for special terms, every employee suggesting a new business field, every crisis needing quick liquidity - these are moments when the initial tension becomes acutely felt again and the temptation is great to abandon one's own position. Strategy therefore has a concrete temporal dimension. It is not only a decision, but a discipline. Not only a choice, but an endurance of that choice over time, with all the corrections and adjustments required along the way.
It is precisely this quality - both the tension in the decision and the path through it - against which one must measure what AI is capable of in strategy work. An AI that produces a 'strategy paper' in which all options are presented equally plausibly has not produced strategy, but a material collection. An AI that formulates a positioning in one meeting but has no relationship to that position three months later has also failed to deliver strategy. The difference is often not seen in practice.
What Language Models Can Do Well
Beyond hype and scepticism, it is worth looking at what language models actually achieve.
Language models are excellent sparring partners for preparing strategic work. They can in short time condense relevant knowledge, structure arguments, make blind spots visible, and formulate counter-positions. Working this way gets you to an informed starting position faster than without them. And that represents genuine added value.
They are particularly suited to a specific form of preparation: standard establishment. If a company operates below industry standard in its marketing or positioning, AI can help quickly and cost-effectively to catch up to 'market standard'. This is a frequently underestimated use case - not all companies need to be pioneers; many first have to do their homework that competitors have long since completed. Here AI often delivers in half an hour much more than three weeks of internal discussion rounds.
They also work as sparring partners for testing arguments: if you have your own idea and want to check it, you can present it to an AI and collect its counter-arguments. That is a sensible approach to realistically assessing your own thinking - provided you are willing to take the counter-arguments seriously and, if necessary, to disagree with them. I want to pick up this line of thought again shortly.
Catching Up vs. Getting Ahead
But it is precisely at this point that an important distinction becomes visible. AI helps to reach standard - it does not help to exceed standard. This is not meant as a technological downgrade, but rather as a description of how it works.
Language models are optimised for the statistically most probable answer. They generate what is typically said and decided in similar situations, weighted by frequency in the training data. This is an advantage when catching up - if you have not yet reached the level common in your industry, AI can quickly get you there. When leading, it is a disadvantage. What exceeds the standard defines itself precisely by being not the most probable answer. A truly differentiating positioning, a truly new business idea, a genuine strategic move lies by definition outside what the models consider probable.
That has a consequence: whoever develops their strategy with AI probably ends up at a point where others using the same tools will also end up - and may already have. The result is not genuinely wrong. It is simply the same result as competitors using the same models. A 'strategy' that tends toward the statistical middle is, in the strict sense, no longer a strategy, but a de-positioning. Those who implement it have no less than others - but they also have nothing that distinguishes them.
The Architecture of the Problem
As already suggested, the reason lies in the architecture of the models themselves. Language models are trained to produce, for any input, the answer statistically most likely to be expected - driven by vast training data in which the majority of articulated positions reflect consensus, industry centre, the familiar. This is their strength when catching up and their weakness in differentiating.
This structural tendency can also be empirically demonstrated: in an experiment I conducted in May 2026, I had four different language models together play the profession-guessing game 'Was bin ich?' broadcast for years on ZDF - a simple yes-no question format that tests several reasoning abilities simultaneously. Three observations from the experiment are particularly interesting for strategy work.
First: frame fixation. Once a particular solution frame had formed in the models' minds - for instance, 'the person sought works alone and controls something' - they clung to it even when new findings later contradicted it. In not a single round did the LLM team manage of its own accord to leave a false frame once it had been established. Transferred to strategy, this means: once an AI gets into a particular strategy frame (for example, 'differentiation via price'), it will stay there even if the data gradually points to something else. Breaking the frame must come from humans. They must disagree.
Second: negation blindness. The models account for positive evidence much better than negative. If a question was answered with 'yes', they built on it efficiently. If it was answered with 'no', they surprisingly often ignored the accompanying implications. This is highly relevant for strategy work, because many strategic statements are negations: 'We are not a premium brand', 'we cannot compete on speed', 'we do not serve the mass segment'. If the AI does not integrate these negations cleanly, it produces strategy proposals that are logically incoherent in detail without that being noticed.
Third, and perhaps most importantly: The models did not think together. Even when three different language models worked together in the same game, the quality of a round usually depended on one or two individual turning-point questions - typically from a single model. More voices did not therefore generate more collective intelligence.
These three observations show clearly: language models can work efficiently within an established frame. Breaking the frame itself - and that is what strategy work is - they cannot really do.
Multi-Agent Setups: Repair, Not Resolution
It is tempting to solve the problem technically: if a single model tends toward the middle, could a setup of multiple specialised agents help - one taking the visionary, creative part, a second playing the critical counter-pole, a third putting it all together? Such multi-agent architectures are currently being tested in practice and do yield better results than single models in certain areas.
But the conceptual limitation remains: it is superficial repair, not resolution of the fundamental problem. Even a 'critic' agent is at its core the same model type and subject to the same structural distortions. What is formulated as criticism is statistically the most probable criticism - not the sharpest, not the most uncomfortable, not the kind that could truly threaten the established position. The 'Was bin ich?' experiment showed precisely that: even with multiple models, there remained fixation on a frame initially established. Multiple LLM voices are not the same as genuine difference.
Multi-agent setups can improve output. But they cannot generate the tension that a strategy decision demands. Those who overlook this have beautifully and securely formulated output, but possibly also something dangerous.
What the User Must Endure
At this point, an aspect becomes visible that practically never comes up in AI discussion - and which is central precisely in strategy work: the quality of an AI sparring partner depends not only on the AI, but also on the psychological discipline of the user. Concretely: on two abilities that at first glance seem to contradict each other.
The first is the ability to endure criticism of your own idea without being intimidated by it. If you submit a controversial, unconventional idea to a language model for review, you will very likely receive a restrained, carefully balanced, in tone friendly but in content rather sceptical assessment - simply because the statistical majority of arguments in the training data tend toward consensus.
The second is the ability to disagree with the AI when its assessment provides insufficient logical substance. Anyone who accepts everything the AI says because it formulates sovereignly replaces their judgment with its - and is manipulable. An AI language model's assessment of your idea is not the judgment of an experienced consultant, but a statistically smoothed probability distribution - and it can simply be wrong if the logic of the idea lies outside what frequently occurs in the training data.
This dual requirement - to endure criticism and be able to disagree - constitutes an essential cognitive competency, involving debate culture with oneself - mediated through a machine. Those without these abilities will either ignore the AI or fall under its spell.
It is interesting to observe that these are precisely the qualities that make a good human consultant. In interplay with the AI, these are now transferred to the person being advised.
The Manipulation Trap
Connected to this is another psychological aspect, one also often swept under the rug in AI discussion. Language models formulate sovereignly. They seem certain, balanced, professional. This linguistic sovereignty is quickly interpreted by the user as substantive sovereignty - a trap...
A hesitant, self-correcting AI would be less dangerous because it honestly communicates its uncertainty. A confidently formulating AI, meanwhile, manipulates into a position sold as secured majority opinion, though it is only statistical majority. Those building on AI recommendations do not only adopt content, they also adopt a certainty not backed by experience. This observation echoes from my essay: language models not only externalise thinking work, they also change how remaining thinking feels - namely, more certain than it should be.
This manipulation trap is not malicious. It is a side-effect of optimisation for fluent and precise language. But it has palpable consequences: strategic decisions emerging from dialogue with an AI feel more solid than they are. That is dangerous because strategy work lives precisely from the uncertainty a good consultant would keep alive.
What Remains - and What to Do
What follows from all this for practice?
First, use language models sensibly. AI is excellent as a sparring partner, argument mirror, material generator and 'standard establisher'. It is not suited as a strategy generator. This distinction should permeate every concrete use. In practice, it helps to provide the AI with a dense context corridor: a reusable file with company description, market position, competitive environment, own strategic hypotheses. A markdown file is completely sufficient, which requires some investment in creation once and can then be used in every subsequent LLM session. This approach dramatically improves output quality. An AI receiving only sparse context generates merely smoothed general statements; an AI with dense context can generate surprisingly relevant friction arguments. Still, it remains friction within market consensus, not beyond it.
Second, do not delegate the cognitive work itself. What the AI cannot achieve - frame-switching, enduring tension, genuine positioning against the middle - must come from humans. That demands the dual ability to endure AI criticism and, if necessary, to disagree with AI arguments. Those unwilling to bring this discipline should not use AI for strategy work - the danger of being manipulated into an 'inconspicuous middle position' is too great.
Third, sharpen understanding of consulting. If AI takes over the replaceable part of consulting - condensing knowledge, structuring arguments, preparing material - the machine-non-replaceable part becomes visible: experience, confrontation, enduring conflicts, forcing decisions. That is not less consulting, but more precise consulting. What was once bought as an unclear package can now be separated more clearly: what the machine can do, and what it cannot. Those who draw this distinction cleanly get better consulting and save money. Those who do not get a worse version of both.
In conclusion: AI does not eliminate the consulting profession. Yet it is also not true that it poses no threat - it does, but in a clearly defined area. The truth is that substitution works precisely where consulting was only ever information. What consulting is beyond that - enduring the tension in which every strategy stands, and accompanying it through time - remains human work. And it becomes not less important through AI availability, but more important. The easier material and middle become accessible and attainable, the more valuable becomes what material and middle are not.