Two ways of thinking about AI are competing for your budget: the broad bet and the narrow one. European companies can only afford one of them first, and it is not the one getting the headlines.

Share of the world’s advanced AI compute — United States: 75%+ · China: ~15% · Europe: under 4%. Figures cited by sovereign AI strategist Nina Schick, 2026.

Let’s start with the part nobody in a European boardroom enjoys saying out loud: we are not going to build the winning model. The compute is not here, the capital is not here at that scale, and the energy price makes the arithmetic worse every year. That race has two runners and we are not one of them.

That would be a depressing opening if building the model were the point. It isn’t. The value of AI does not accrue to whoever trains it. It accrues to whoever applies it to a problem that was expensive to solve yesterday. And on that measure, the scoreboard is wide open.

Where European companies actually stand

Eurostat’s numbers are the ones to hold on to. In 2025, 20% of EU enterprises with ten or more employees used AI technologies at all, up from 13.5% a year earlier. Denmark leads the continent at around 42%. That means in most European markets, four out of five companies you compete with are doing nothing with this technology beyond an individual employee quietly pasting text into a chat window.

The worker-level picture is the same shape. Research from the St. Louis Fed and CEPR put generative AI use among US workers at 43% in early 2026, against a European range of roughly 26% in Italy to 36% in the UK. The European Central Bank’s survey of firms found companies expecting to put around 9% of their total investment into AI in 2026, so the intent is real and the money is moving.

Read those two facts together and you get something more useful than a lament. Europe is behind on building and behind on using. The first gap is structural and you cannot close it. The second gap is operational, it sits inside your own company, and it is currently the single cheapest competitive advantage on the table, because almost nobody around you has closed it either.

The bar to be ahead of your competitors on AI is embarrassingly low right now. That window closes on a schedule you don’t control.

The two ways to think about AI

Almost every conversation I have with a CEO or a CMO about AI is really a collision between two different mental models. Both are legitimate. The trouble starts when a leadership team runs on one without noticing the other exists.

The broad bet: AI as the destination

This view treats AI as a platform shift. The question is what the company becomes when the technology matures, how the operating model changes, which roles disappear, what the category looks like in three years. It is directional, ambitious, and it is the view that gets capital moving.

Failure mode: a strategy deck, a steering committee, and no change to any actual process. Or the opposite — a large platform purchase made before anyone named the problem it solves.

The narrow bet: AI as a tool against a named bottleneck

This view ignores what AI might become and asks where, in this business, this quarter, a specific constraint exists that this technology can move. It is unglamorous, it is measurable, and it produces evidence rather than opinions.

Failure mode: forty disconnected experiments that each work slightly and add up to nothing, because no one ever asked what they were building toward.

The clearest live demonstration of the split is the one everyone is watching. The American centre of gravity is frontier capability and scale: build the most capable system, get it everywhere, win the platform. The Chinese centre of gravity is diffusion: release capable models openly, push them into factories, vehicles and workflows, and compound the advantage at the application layer. In manufacturing the difference is visible — analysis published by AI Frontiers put AI deployment in production at 67% of Chinese industrial firms against 34% of comparable US firms. On Hugging Face, open Chinese models now lead on total downloads, and the models built on top of them have overtaken those built on American foundations.

One honest caveat, because you will be challenged on it if you repeat this: these are centres of gravity, not separate worlds. Plenty of American manufacturers do narrow, boring, high-return AI work. Plenty of Chinese labs are chasing the frontier as hard as anyone. And a deployment rate tells you what was installed, not what paid off. Use the comparison as a lens, not as a verdict on two nations.

Why doing nothing feels so reasonable right now

Most leaders I speak to are not sceptical about AI. They are stuck. The reasoning goes: the technology changes every six weeks, the vendor landscape is unreadable, our data is a mess, and if we commit now we will have committed to the wrong thing. So the safest move is to wait for clarity.

The market data says you are in large company. Gartner’s 2026 CMO Spend Survey found 98% of CMOs using or piloting AI, while only around 30% say their organisation is ready to scale it. Seventy percent call becoming an AI leader a critical goal for the year — and seventy percent also admit their internal processes are not mature enough to implement and scale it. BCG puts the share of companies struggling to achieve and scale value from AI at 74%. McKinsey’s global survey found around 88% of organisations using AI somewhere, roughly a third scaled beyond pilots, and only 39% able to attribute any enterprise-level profit impact at all, most of it below five percent.

That is not a picture of a technology that doesn’t work. It is a picture of a technology being adopted without a decision behind it. Piloting is what organisations do instead of choosing. It feels like motion, it costs real money, and it produces no compounding advantage whatsoever.

Why this lands hardest in marketing

Marketing has the largest AI surface area of any commercial function and, in most companies, the weakest measurement discipline. That combination is dangerous. Gartner has CMOs allocating 15.3% of marketing budgets to AI initiatives. The Duke/AMA CMO Survey has generative AI touching roughly 15% of marketing activities, with 91% of respondents saying implementation takes too long.

Here is the uncomfortable mechanic. AI reduces the cost of producing marketing output. If your marketing does not currently count — if you cannot trace what a campaign contributed, if the attribution is guesswork, if content volume is the proxy for effort — then AI will not fix that. It will let you produce three times as much of what was already not working, faster and cheaper, and the noise will bury the signal completely.

AI does not make bad marketing good. It makes bad marketing cheap, which is worse, because now you can afford much more of it.

The companies that will get real leverage out of this are the ones whose marketing was already a system before AI arrived: a defined funnel, known conversion rates, a measurement setup you trust, a clear idea of which constraint is actually limiting growth. If you have that, AI is an accelerant pointed at a known bottleneck. If you don’t, AI is an expensive way to generate more unattributable activity. Fix the system first, then point the technology at it

The rule: think broad, act narrow

The sequencing is the whole argument, and it is not a compromise between the two models. It is a specific order.

• Think broad so that your narrow moves point somewhere. The three-year view is what stops a pilot programme from fragmenting into unrelated experiments. It tells you which capability you are accumulating.
• Act narrow because that is the only mode a European mid-market company can actually execute, and because narrow moves produce evidence. Evidence is what earns budget for the next step. Vision does not survive one bad quarter; a documented 40% reduction in campaign production time does.

Three questions to run any AI proposal through before it gets funded:

• What is the constraint? Name the bottleneck in one sentence without using the word AI. If you can’t, you have a technology looking for a problem.
• What does the three-year version of this look like? If this works and you do it twenty more times, what capability have you built? If the answer is “we’d be faster at some things”, it’s a tool purchase, not a strategy.
• How will we know within eight weeks? Not “how will we know eventually”. A narrow move that cannot be judged inside a quarter is a broad bet in disguise.

What broad looks like: the questions for the next three years

These do not need answering this month. They need to be live in the leadership team so that the narrow moves accumulate in one direction.

• If producing content costs close to nothing, what becomes scarce in your category? Usually: distribution, trust, proprietary data, or genuine customer relationships. That scarce thing is where your strategy has to move.
• What proprietary data do you hold that a competitor cannot buy? Sales call recordings, service histories, quote-to-close patterns, application data. Are you even collecting it in a usable form? This is the one asset AI makes dramatically more valuable and most companies are throwing it away daily.
• When your customers research through an AI assistant instead of a search engine, who owns the source material it reads? That is a content and PR question with a three-year lead time.
• Which parts of the customer relationship stay human on purpose? Decide it deliberately now, rather than discovering the line after you’ve crossed it.
• What would your team do with three times the throughput at the same headcount? If you have no answer, the productivity gain will quietly evaporate into more output nobody asked for.

What narrow looks like: what you can start this week

Every one of these is small enough to run without a committee, and each produces a number you can put in front of a board.

• Mine your own customer language. Put twelve months of sales calls, support tickets and reviews through a model and extract the actual words your customers use about their problem. Compare it to your website copy. This is the cheapest, highest-leverage move on this list and almost nobody does it.
• Check what AI assistants say about your company today. Ask five of them what your firm does, who it serves and what it costs. The answers are being given to prospects right now, whether or not you have looked. Then fix the source material they are reading.
• Compress one briefing loop. Take a single recurring deliverable — a campaign brief, a monthly report, a product description set — and rebuild the path from input to first draft. Measure cycle time, not volume produced.
• Vary creative against one live campaign. Same offer, ten angles, let the spend decide. The real work here is the measurement setup, and discovering you don’t have one is a useful result.
• Enrich inbound leads before a human touches them. Company data, fit scoring, routing. It cleans your CRM as a side effect and it shortens response time, which is one of the few marketing variables with an unambiguous link to revenue.
• Automate competitor monitoring into a ten-minute weekly read. What changed in their messaging, their ads, their positioning. Judgement stays human, collection does not.

Where not to point it

Being specific about the limits is what makes the rest credible.

• Anything reaching a customer unreviewed. An editorial gate is not bureaucracy, it is the thing standing between you and a brand incident.
• Personal data through tools you have not assessed. GDPR did not go away and the AI Act arrived on top of it. Check before, not after.
• Cutting junior roles before capturing what they know. Juniors are how tacit process knowledge gets documented. Remove them first and you automate a process nobody understands any more.
• Platform purchases made before the bottleneck is named. If the vendor identified your problem for you, be suspicious of the diagnosis.

How you will know it worked

Measure this

• Cycle time on one named process, before and after
• Hours saved, in the abstract
• Cost per qualified lead, cost per result
• Volume of content produced

Not this

• Share of AI output accepted without rework
• Number of tools or seats deployed
• Time from enquiry to first human response
• Percentage of staff “using AI”

If a pilot cannot move one of the metrics in the left column within a quarter, stop it. That discipline is what separates a programme from a habit.

What small step is actually worth taking in your company?

That question has a different answer in every business, and it depends far more on your existing marketing system than on the technology. It is also not a question worth answering alone, or from a vendor’s demo.

I spend my time in two places: as a marketing strategist working with companies on what makes their marketing count, and inside Yuntos Lab, where we test this technology before it goes anywhere near a client, and then implement it across our own agencies. That gives me a fairly direct view of what works, what is still theatre, where it is genuinely dangerous, and where it is an obvious yes.

If you are a CEO, CMO or marketing manager reading this while quietly doing nothing because the choice feels impossible: get in touch. An hour of challenging each other on where your leverage actually sits is worth more than another quarter of waiting for clarity.