"Where do we start with AI?" and "how do we cascade it across the group?"

by Yannis Larios

The two questions every C-level executive asks me. Always.

Over the past year I have sat with the leadership teams of over four dozen companies across Europe, reviewing where their AI strategies stand and where they should go next. Listed industrial groups. Pan-European financial institutions. Multi-brand retail groups. Family-controlled mid-caps. Professional firms.

Different scale, sectors, countries, starting points. Almost identical conversations. Those same two questions arrive nearly word for word, every time: "Where do we start with AI?" and "How do we cascade it across the group?"

The first has a simple answer: start at the top, with the whole C-suite, for a few days. Simple to say, hard to do, because most organisations instinctively do the opposite. Cascading is where companies come apart.

AI is not a conventional software rollout. It does not arrive, get installed and close as a project. It must be business-owned, technology-enabled and risk-governed, and it improves through use, iteration and licence to experiment. Leaders raised on disciplined implementations find that uncomfortable, and they are right to.

The 10x change sits at the C-level table

A tenfold rethinking of what the business sells, how it serves and how it decides can only be authorised by the people who run the company. That authority belongs to the C-suite, and it sits nowhere else.

Without a structured way in, the thinking stays small. More than once, a chief executive has sent me his list of what AI would change in his company before we had done any work together. Most items were a faster version of what the company already did. Bill quicker. Report sooner. Better documents. Capable people who have never been given the space to reimagine rather than improve.

The distinction is simple enough to test in a morning. The efficiency question asks how to do today's work faster. The tenfold question asks what the company could sell, serve or attempt that was impossible last year. Most leadership teams have never been asked the second one.

Which is why the sprint is with the full C-suite. Every member. Chief executive, finance, operations, commercial, technology, people, legal. A few days, not months. Around it sits a crash course on the AI tools and their strategic implications: what they do, where they fail, and why an answer that held twelve months ago may not hold today.

Then the part that matters most. One or two days inside the company's own context, where the team brings its live decisions and works them through with AI at the table. A pricing decision deferred twice. A market entry nobody has had the capacity to size. A customer proposition rejected three years ago on economics that might hold today. This is hands-on, small-scale training for the people who run the company, and they do the work themselves. Watching someone else demonstrate it changes nothing.

What comes out is a hands-on understanding of where AI can change the shape of the business tenfold, held by the people with the authority to act on it.

A caution, because I have watched this fail. The energy in that room decays fast. Two weeks later the quarter reasserts itself and the insight becomes an anecdote at the next board dinner. The antidote is to leave the room with two or three initiatives already owned, funded and scheduled. A sprint that ends in enthusiasm rather than commitments teaches your leadership team that AI is interesting, which is worse than not running it at all.

Where AI compounds

The C-level sprint sets the direction and the strategic appetite for AI. The compounding, though, happens in the departments, every day, across every function you run.

A framing I learned on Stanford's Generative AI in Business programme, and have followed ever since, is Augmented Intelligence. Read "AI" that way. Artificial Intelligence sits apart from your people. Augmented Intelligence works through them, amplifying judgement, analysis and throughput.

Look at how your people spend their days. In almost every company I have seen, capable people act as expensive "routers". Moving a file between systems. Copying a report into a spreadsheet. Chasing an approval across three inboxes. Reconciling one list against another. They sit on the seams between systems that were never built to speak to each other, and they hold those seams together by hand.

That is where AI agents belong. They take the routing, and they release good people to do work that matters. Marketing, sales, finance, procurement, HR, customer service, legal. Every function carries the same seams and the same people stuck on them.

Individually, these AI use cases may look unglamorous. Yet the productivity they release is remarkable. Run across every function and repeated across the years, the results compound tremendously, and the business arrives somewhere optimisation could never have taken it.

So an AI strategy with nothing running underneath it is a memo. Equally, AI use cases spreading with no strategic direction produce a list of tweaks. Both are required.

Cascading: the question before the structure

When one team solves something with AI, how long before the rest of the business has it?

Every AI use case teaches you something. What worked. What the data really looked like. Which vendor delivered and which one presented well. Where it broke, and what it cost. In most organisations those lessons stay where they were learned, and six months later another department, or another company in the group, spends the same money to learn them again.

Carrying that learning across the business is what an AI operating model is for, and the choice matters most in groups that own many companies, where one lesson can be paid for five or six times before anyone notices. I see five recurring operating patterns, each under its own conditions, and each moves what has been learned in a different way.

Centralised puts one team in charge of priorities and delivery. Everything is learned in one place, which makes transfer trivial and local fit difficult. It suits a business where the same process runs the same way everywhere.

A centre of excellence concentrates scarce expertise and lends it out, so lessons travel with the people who carry them. Right when skill is the constraint, as it usually is in the first year. Its failure mode is predictable: demand outruns the centre, the queue lengthens, and the businesses quietly go around it.

Hub and spoke keeps standards, platforms, governance and accumulated lessons at the centre while local teams find and run the opportunities. It is the most deliberate of the five about transfer, treating it as a standing responsibility rather than something that happens when people remember. It is where most of the groups I work with settle, and if you asked me to choose blind, I would start here.

Federated shares principles and decision rights explicitly, and transfer happens through the agreements themselves. It suits a centre that leads by agreement, which describes most holding companies with strong operating chief executives. It demands more written-down discipline than groups expect.

Networked moves people and initiatives to wherever the need is, so knowledge travels in their heads rather than through a process. The most adaptive of the five and the most demanding. I would not start here in an organisation's first coordinated year of AI.

One word of realism: running a single AI programme across one shared function in every company sounds efficient, and it stalls when those companies sit on different systems, different data and different processes. It works where the function is already shared: group finance, treasury, a shared service centre, a central marketing team.

Whichever structure you choose, govern AI deployment as a portfolio of AI use cases at department level rather than a queue of projects. The discipline I build with clients rests on three rules. Every AI initiative carries a DRI, a directly responsible individual, and a measurable KPI. The criteria for stopping or scaling are agreed before the work starts. And the portfolio respects a ceiling I rarely see respected: beyond three or four live initiatives per department, governance burden and leadership attention start consuming the value.

None of this exists to slow the experimentation down. It exists to make sure the experiments that work get noticed, funded and repeated, and the ones that don't get stopped before they quietly consume a year.

What surprises me

I sit in strategy conversations with the C-suites of serious, well-run companies, and yet that lingering anxiety sits in the room. They know there is no AI strategy on the table and no AI governance behind it, and it weighs on them. It should. Under the EU AI Act’s recently amended Article 4, providers and deployers must take measures to support AI literacy among those who use AI systems on their behalf. The underlying obligation is already in force, it covers all AI rather than only high-risk systems, and it is a part of the Act that many boards I meet have not yet discussed.

 The AI tools are already inside those companies. Some arrived on a licence, most because people brought their own. Strategy moves slower, understandably. Reviewing where an AI strategy stands takes an afternoon. Designing one takes senior attention that nobody has spare, and few leadership teams have yet had the chance to weigh one operating model against another.

The starting point is those decisive days with the people who can rethink and reimagine the business, followed by the method that lets everyone else act on it. It certainly isn't the AI pilot, run by a department to postpone a decision that belongs upstairs.

None of the companies I have watched get this right worked it out alone. They borrowed the pattern from someone who had already seen it fail elsewhere and then built properly, and it spared them a full budget cycle of expensive learning and misdirected AI spending. That experience transfers. And it is the shortest route I know from the first question to the second.

Yannis Larios advises companies on AI strategy, governance and operating-model design, and facilitates the delivery of C-level AI programmes.