AI@Strategy Conference 2026: The Programme, the Sessions, and the Business Model Question
The AI@Strategy Conference 2026 takes place in Munich on 7 and 8 October, organised by the Chair for Strategy and Organization at TUM School of Management. This post describes what the programme covers, names five sessions I find worth attending, and argues that its subject throughout is the organisation rather than the business model. Education is the main case: the need to learn stays constant, while the institutions like schools or universities that serve it are a historical answer to scarcities that are now dissolving. Mobility, healthcare and pharmaceuticals show the same pattern. when AI meets the current business models. The post closes with questions I am taking to Munich.
AI@Strategy Conference 2026: dates, venue and programme
The AI@Strategy Conference 2026 takes place on 7 and 8 October 2026 at the TUM Audimax in Munich. It is organised by the Chair for Strategy and Organization at TUM School of Management, and positioned by the organisers as a strategic prelude to Davos 2027 and the Munich Security Conference 2027. The published programme lists more than forty speakers from academia, industry and politics.
- Dates: 7 and 8 October 2026
- Venue: TUM Audimax, Arcisstrasse 21, 80333 Munich
- Organiser: Chair for Strategy and Organization, TUM School of Management
- Programme: aistrategyconference.com/#program
- Conference site: aistrategyconference.com
I have been invited to attend by the organisers, Prof. Dr. Isabell Welpe and Prof. Dr. Theresa Treffers, and I thank them for the invitation. Putting a programme of this quality together takes a great deal of work, and the reading that follows is meant as a contribution to it rather than a verdict on it.
What follows is a reading of the programme as published in early September, which the organisers describe as continuously updated.
The AI@Strategy Conference 2026 programme is about organisational adoption
The AI@Strategy programme has a clear centre of gravity: how established organisations absorb artificial intelligence.
Day one moves from opening keynotes and policy through organisational deployment to risk, ethics and governance. BMW presents on building an AI-native operating model, Allianz on running AI at scale in financial services. The final session covers workflows, governance as strategy, and regulation in financial markets.
Day two moves from organisational strategy through leadership to what the organisers call the human, physical and social dimensions of AI. Deloitte on building an AI-fuelled company, McKinsey and BCG on AI and firm performance, and an extended session on leadership under AI: Emilio Castilla on meritocracy and bias, Jacqueline Ng Lane on human-AI decision making, Hannah Riley Bowles on negotiation.
Both mornings open with pre-conference workshops on agentic AI, leadership, negotiation and the human operating model.
This is an accurate description rather than a complaint. On organisational adoption, leadership and governance, the programme is well constructed and seriously staffed. The research on the leadership side in particular is empirical work by people who have done it, not a panel of opinions.
Five AI@Strategy sessions worth attending
The programme is large and the choice is genuinely difficult. What follows is a personal selection rather than a ranking: these are the sessions I find most interesting right now, for reasons that have to do with my own work rather than with their relative quality.
I am also looking forward to seeing Christopher Tucci again. He works on digital strategy and business models, which makes him the closest academic neighbour to my own field on this programme. His session title had not been announced when I read the schedule.
From AI Assistance to AI Delegation: Evidence from ChatGPT and Codex. David Holtz, day one. Evidence rather than folklore on the shift from assistance to delegation. If delegation becomes the unit of AI-enabled work, time-based revenue models in professional services stop making sense.
How do AI start-ups manage tensions around dependence and liability? Shaz Ansari, day one. Built on a real trade-off. Dependencies are choices in the architecture of value creation: what a firm owns, what it rents, where bargaining power ends up.
Multi-Agent Orchestration Frameworks and Generative UIs. Dirk Hörig, day one. Ten minutes on the most commercially consequential question in the programme. If a customer’s agent searches, compares and negotiates, the agent becomes the customer interface.
AI for Scientific Discovery. Fabian Theis, day two. The one session in which AI changes the object of the business rather than its throughput.
Physically Grounded World Models. Matthias Nießner, day two. Language models changed knowledge work. Spatial models change what machines can do in physical space.
The subject of the programme is the organisation, not the business model
Read the programme as a whole and a pattern appears. The subject throughout is the organisation, and AI is what acts on it. How the operating model changes. How leadership changes. How work changes. How it is governed. Even the sessions I listed above, which are the most interesting in the programme, keep that structure.
The framing contains an assumption: that the organisation persists, and the question is how it adapts.
The prior question is whether it persists. New technologies can enable new business models, and those models are usually built by companies that did not previously exist. The risk for an established organisation is not that it adapts too slowly. It is that someone else serves the same need with a configuration that makes the organisation unnecessary. Adapting well and becoming unnecessary are compatible outcomes.
So the question the programme does not ask is this one: which businesses exist only because of constraints that AI is now removing, and what should be built once those constraints are gone?
Where the business model question does appear in the programme
It would be wrong to say the question is absent. It is present, but dispersed, and mostly in short slots.
Mari Sako works on firm boundaries, outsourcing and professional services. Her session on AI-driven supply chains and value creation is architecture in the strict sense: who does what, and where the boundary of the firm sits. She has ten minutes and appears virtually.
Björn Eskofier holds a chair at LMU on AI-supported therapy decisions and came from sensor-based health monitoring. Continuous measurement instead of episodic diagnosis changes what a health provider promises and how it is paid for, not merely how efficiently it operates. He appears inside a start-up lightning round.
Joachim Ziegler runs the InsurTech Hub Munich. Insurance is the clearest example of a business model built directly on a technical constraint: risk is pooled because individual risk could not be predicted. That is the constraint data logging and AI is eroding, and pooling is not a detail of the insurance model. It is the model.
Ziyue Li works on transportation analytics at TUM, which is the coordination question in mobility rather than the vehicle question.
Catherine Wang researches AI in strategic decision making and opportunity recognition, which sits upstream of any business model choice.
The five sessions listed earlier belong to the same thread. Put both lists together and a programme appears that the schedule itself does not present:
Day 1
- 12:35 David Holtz, delegation and what it does to time-based revenue
- 14:35 Shaz Ansari, dependence and liability
- 14:55 Dirk Hörig, the agentic customer interface
- 17:10 Christopher Tucci, title not yet announced
- 18:10 Lightning round, for Ziyue Li and Joachim Ziegler
Day 2
- 11:10 Fabian Theis, discovery and where value migrates
- 11:55 Mari Sako, firm boundaries and value creation, virtual
- 13:00 Matthias Nießner, world models
- 14:50 Leadership panel, for Catherine Wang
- 15:50 Start-up lightning round, for Björn Eskofier
That is a full two days, and it is a different conference from the one the programme headings describe. Two of these appearances sit inside lightning rounds, one is a virtual ten-minute slot, one is a panel on another subject. That is why the thread is easy to miss. It is not hidden. It is compressed.
The material is there. What is missing is the frame that would connect it.
Education is the clearest case: the need stays, the institution does not
The need to learn does not change. People want to be able to do things they cannot yet do. That was true before schools and will be true after them.
The school is a historical answer to specific scarcities: too few people who knew anything, books that were expensive to copy, no way to give thirty children individual attention, no way to assess competence except by having an adult watch. Year-group classes, timetables, fixed curricula, examinations and degrees as signals are configurations built on those constraints.
Every one of those scarcities is dissolving. Explanation is available on demand, adapted to the individual, at marginal cost near zero, in any language, at any hour. Assessment of what someone can actually do is becoming continuous and cheap.
This does not mean schools and universities disappear. It means their configuration stops being the obvious answer and the value they add has to be argued rather than assumed. Someone will do that work. It will not necessarily be the incumbents.
The example is not incidental to this conference. The host is a university, executive education is represented in the room, and education as a value creation system does not appear in the programme.
The same pattern runs through mobility, healthcare and pharma
Education is the clearest case but not a special one.
In mobility, you can use AI to improve the combustion engine, and that is legitimate engineering with real value. You can also use it to coordinate vehicles, in which case the scarce resource is not efficiency but road capacity and utilisation, and the unit of competition moves from the vehicle to the system. Both are available. They are different questions.
In healthcare, the same pattern-recognition capability can find additional codable diagnoses that raise reimbursement, or identify people whose disease trajectory is still reversible. One of those has a funded industry behind it. The other largely does not, and the difference is not technical.
In pharmaceuticals, if discovery becomes cheap, value migrates: to clinical data, to trials, to regulatory capability, to patient access. The laboratory’s role changes.
This is not a sector story. AI is a general-purpose technology, and it does not enter some industries and leave others alone. Every established business model contains assumptions about what machines cannot do.
We made this mistake with the internet: optimising instead of rethinking
The comfortable version of the internet story is that established companies ignored it. They did not. They invested, they hired, and many understood the technology in considerable detail.
They optimised with it instead of rethinking through it. The internet was treated as a new channel, a cheaper infrastructure, a digital version of an existing product. The technology was visible. The level at which it would change competition was not.
The mechanism is the one that matters here. Customer needs or what I prefered to say jobs-to-be-done stayed constant. People still wanted to find information, hear music, watch films, buy things. What changed was the configuration that served the need, and who ran it. The companies that won were not incumbents who converged. They were built on configurations the incumbents could see and had no reason to want.
AI itself is not a business model innovation
AI is a capability in a business model, not more, not less. A business model changes only when its parts change together: the value proposition, the architecture of value creation, the revenue model, and the team and values that carry it. These are not a list. They are a system. Change one and the others have to move.
This is the point I have worked on since my doctoral dissertation on business models in the digital economy in 2001, and it has not needed revision. New technologies can enable new business models. Established firms use them to optimise the business models they already have.
Both halves matter. The first is a possibility, not a certainty. The second is an observed regularity, and it is not caused by incompetence. It is caused by the fact that optimising an existing model has a known customer, a known cost base and a measurable return, while designing a new one has none of those at the point where the decision has to be made.
Adding AI to every function on the organisation chart does not produce a new strategy. It produces a faster version of the existing one. A company can become considerably more efficient with AI and still become irrelevant.
What has kept me on this question for twenty-five years is not the business model concept itself. It is the media model underneath it, which I developed with Prof. Dr. Beat Schmid at the mcminstitute in St. Gallen. Digital media are not electronic data processing. An active information carrier does not process. It acts, it initiates, it is an agent. Written down in 2001, that is a structural description of what is now called agentic AI. I did not predict it. The model held the seed, not the harvest.
The dissertation drew one conclusion from this that is worth rereading now: the challenge to established firms will not come from firms that use new media to do the same thing more efficiently, but from entrepreneurs who satisfy a need differently (Stähler, 2001, p. 289). Exactly one word needs replacing for 2026.
Machine agents are also not new, which the current debate tends to forget. Financial markets have had them for a long time. Algorithmic systems were already placing a large and rising share of orders in US equities two decades ago (Hendershott, Jones and Menkveld, 2011). Humans set the strategy, machines act, and the division of roles is clear to everyone involved. What is new is that the same structure is now arriving in domains organised very differently from a trading venue, where the configurations it permits have not been built yet.
Which is why analysis is only half the work. Business model innovation is a generative strategy: it does not select a position within a given space of possibilities, it produces possibilities that were not there before. The analysis is demanding, and the imagination that follows is not a substitute for it. It is its yield. Only when you understand precisely what a medium makes possible can you responsibly design what does not yet exist, and then go and build it.
The questions I am taking to the AI@Strategy Conference 2026
These are research questions rather than a checklist. The third cluster is the one I called a generative strategy above, and it is worth being precise about the provenance: I did not have that term in 2001. The substance was in the dissertation. The name came much later, but this is precisely what we need for finding new business models.
What are the economics of AI, and what are they becoming?
Two curves run in opposite directions. Which of them governs competitive advantage, and over what horizon? The cost of being at the frontier grows by roughly 2.4 times per year (Epoch AI). The cost of reaching a given capability level falls by one to two orders of magnitude per year, and open-weight models trail the leading closed models by about four months (Epoch AI, Epoch AI). Whether access to a model is a durable resource or a commodity depends on which curve you are standing on.
What happens to a business model built on metered access when the model runs on the customer’s own hardware? If frontier-grade capability is four months from open weights and an order of magnitude cheaper a year later, local execution stops being a preference and becomes an ordinary economic choice. Marginal cost per query, the vendor relationship, the data flow and the switching cost all move at once. Those are four different parts of the same business model.
If the same capability can serve a need or serve the billing, the revenue model is a selection mechanism acting on technology. What governs that selection? The same pattern recognition can identify a still reversible disease trajectory or an additional billable diagnosis. One of the two has a funded industry behind it. That outcome is not technical, and it is usually reached without anyone recording that a decision was taken.
Which patterns become legible once the search space exceeds human intuition?
Which regularities lie in plain sight that no human has been able to read? Protein folding stayed open for fifty years, not because biochemists were insufficiently diligent, but because the space exceeds human intuition. The same is plausibly true of the genome, the microbiome and every domain we currently describe rather than understand. The research question is which of them are now legible, and how one would know before trying.
Which value propositions become possible once such a space becomes navigable, and which architecture and revenue model do they require? This is the step the debate skips. New epistemic reach is not a business model. It becomes one only when it yields something a customer needs, an architecture that can deliver it, and a way of being paid that survives contact with the incumbent system.
How do we strategise when the goal cannot be known?
How do we distinguish a limit set by physics from a limit set by unsolved coordination? Both present as walls, and organisations optimise against both with equal conviction. Transaction cost economics explains what coordination costs do. It does not tell us which of them are about to fall. The difference decides whether you improve the engine or reorganise the road.
What would a method look like that treats the goal itself as the object of search? Our instruments resolve uncertainty about the route and presuppose certainty about the destination. A generative strategy faces the opposite case, and a hard evaluation problem with it, because its output cannot be specified in advance. An organisation that can only pursue goals it can state beforehand will only ever reach places it has already imagined.
The programme shows what a field puts on a stage. It does not show what is discussed in the breaks. I will report on the difference.
Understand. Imagine bigger. Act.
Sources
Epoch AI. Training compute cost of large-scale AI models. Link. Puts the growth in training cost for the largest models at about 2.4 times per year, roughly a doubling every eight months.
Epoch AI. LLM inference price trends. Link. Measures how fast the price of reaching a fixed capability level falls: between nine and nine hundred times per year depending on the benchmark, and about forty times per year for GPT-4-level performance on PhD-level science questions.
Epoch AI. The gap between open and closed models. Link. Since January 2026 the most capable open-weight models have trailed frontier closed models by an average of four months, or eight points on the Epoch Capabilities Index.
Hendershott, T., Jones, C. M. and Menkveld, A. J. (2011). “Does Algorithmic Trading Improve Liquidity?” The Journal of Finance, 66(1), 1–33. Link. Documents the growth of algorithmic trading in US equities and its effect on liquidity; used here only for the observation that machines have been transacting autonomously in financial markets for a long time.
Stähler, P. (2001). Geschäftsmodelle in der digitalen Ökonomie: Merkmale, Strategien und Auswirkungen. Dissertation, University of St. Gallen. Link. Introduces the business model as a unit of analysis and the media model of the infosphere on which the argument above rests; the sentence quoted here is on p. 289.
Pre-conference analysis, 6 September 2026. This article will be updated under the same URL during and after the conference.
