Benedict Evans published his new presentation, AI Questions, on 18 November. Across 86 slides, two findings stand out for anyone who runs a European organisation. Giving everyone AI is not the same as changing how the organisation works, and the data so far shows mostly the first. Meanwhile, AI agents are about to test every business model that depends on customers who never compare, switch or complain. This is what the deck shows and what to do with it.

A boom among the builders, not yet among the users

Evans is an independent technology analyst and a former partner at Andreessen Horowitz. Twice a year he publishes a large deck on where the tech industry is heading, and this one replaces AI eats the world from May. His summary of 2026 fits in one line: it is still very early.

Inside the tech industry, early looks like a gold rush. Anthropic, which filed for its IPO in October, reported third-quarter revenue of $16.96 billion, more than twelve times the figure a year earlier, mostly driven by agents that write software. Meta, Amazon, Alphabet and Microsoft plan to spend more than $700 billion on data centres and chips this year, more than twice what the global telecoms industry invests. Their capital spending now grows faster than their cash flow, so the money increasingly comes from debt, leases and new share classes.

At the same time the models themselves are commoditising. Anthropic and OpenAI still lead, but Google, Meta and SpaceX have caught up and Chinese models sit close behind. Stripe agreed to buy the model router OpenRouter for more than $7 billion, Nvidia is buying Hugging Face for nearly $13 billion, and large companies now buy models from more than one supplier. Evans’ provisional thesis is that models become infrastructure and power moves up the stack, to the products and workflows built on top. That is also why the lasting asset for your organisation is not the model you rent but the context you own.

None of that boom tells you much about adoption in your own organisation. That is where the second half of the deck starts.

Adoption is broad but shallow

In the United States, about 45% of workers use generative AI at work, according to the survey data in the deck, but fewer than 15% use it every day. Even in the tech sector, occasional users outnumber daily ones. Spending data from the corporate card company Ramp is more extreme. The top 1% of firms spend more than $10,000 a month per employee on AI, the median firm about $16.

A Deloitte survey of the UK workforce, with a weighted base of 25,000 people, shows what this looks like from the inside. Among those using AI at work, about two-thirds have permission from their employer, but only a third have tools paid for by that employer. About half have had any training, and a third believe management understands AI. Evans calls it the spray-and-pray phase of deployment.

European statistics look better on paper. Eurostat reports that 20% of EU enterprises with ten or more staff used at least one AI technology in 2025, up from 13.5% a year earlier, and Belgium reached 34.5%. These figures count whether a company uses AI, not how deeply it has changed anything.

Diffusion is not transformation

Evans’ most useful slide for leadership teams compares three decades. In the 1980s companies put spreadsheets on every desk, but their finance operations ended up running on databases and systems such as SAP and Oracle. In the 1990s every PC got a browser, but that did not rebuild any business around the internet. In the 2020s organisations roll out Copilot or ChatGPT to everyone.

Timeline of three decades where a tool on every desk is followed by the systems that really changed how companies worked
Every platform shift started with a tool on every desk. The transformation came later, from systems that rebuilt how the work was organised. After the framework in Evans’ AI Questions.

The pattern creates a trap. A company-wide licence scales but does not transform, while a pilot may transform one process but does not scale to hundreds of workflows. In an essay from September, Evans explains why the gap persists. Most people are not tool builders, even when AI writes the code. Writing code was never the hard part. Knowing what to build is, and so is changing a workflow that crosses five departments, three systems and several regulatory regimes. The same gap shows up in software teams, where AI makes developers faster but not their teams.

It also takes time. After two decades, the public cloud still runs only about 30% of enterprise workloads, according to the Goldman Sachs CIO survey Evans cites. He describes changing how a company works as an industry rather than a project, with consultants and integrators doing much of the work. He also asks whose questions these really are, and splits them into three. How do we buy, build and deploy this, which looks like IT procurement but is also about leadership and culture? How does it change our operations? And does it change our product and our market?

From faster and cheaper to new questions

Evans sees three stages. First you absorb the technology and do the old things faster and cheaper, then you innovate, and only then does the market itself change. He looks for the new through two lenses. Automation makes existing work cheaper. Abstraction changes the level of the question, from automating Salesforce data entry to asking why churn is changing or which customer to call today.

Cheaper tasks often lead to more work rather than less. When barcodes automated the supermarket checkout from 1974, stores could manage inventory far better and ended up stocking roughly five times as many products. But infinite cheap interns work for attackers too. During what the security world calls Black October, high and critical vulnerabilities reported by more than twenty large technology vendors and projects rose to over 8,000 a month, a multiple of a year earlier. For security teams in Europe this lands on top of reporting clocks that are already running.

Inertia is no longer a moat

The most original chapter is about friction and inertia, which Evans treats as a whole field of economics. Many businesses earn money because customers cannot be bothered or find switching too complicated. An agent does not get bored.

For UK mobile operators, Goldman Sachs estimated that 5% to 20% of free cash flow is at risk in the medium term if out-of-contract customers let agents optimise their plans. The bank expects consumer adoption to take three to five years, names Orange and KPN as defensive buys and sees a higher risk-reward profile at BT, Deutsche Telekom, Vodafone and Telefónica. In the US, Apollo’s chief economist Torsten Slok asked whether an agentic bank run is coming, with assistants moving household cash out of checking accounts paying 0.1% into accounts paying 3.3% to 5%. A study of health insurance markets in the deck shows people paying thousands of dollars to keep their plan, mostly out of inattention and the hassle of switching.

Four cards for telecoms, bank deposits, insurance and public services, each showing what customer inertia is worth
Where hassle and inattention pay today, and what an agent that compares, switches and complains puts at risk. Sources in the article.

Public services already feel it, because complaining has become almost free. Researchers from the Hertie School, the Cooperative AI Foundation and GovAI documented 84 cases across 11 jurisdictions where AI-written requests and claims flooded government services, from German social courts to French administrative courts and UK ombudsman services. They warn that the fastest responses, such as new fees and barriers, tend to hit the least digitally literate first. They recommend auditing the most exposed services and integrating digital identity into them.

Consumer agents are still tiny in commerce. Meta’s Muse passed 10 million users, but during Halloween week Uber counted around 500 orders and four rides from third-party agents out of roughly 320 million trips. So this is a question of years, not quarters. It does raise a strategic question Evans puts bluntly for retailers and platforms: do you want to become someone else’s API call, or keep the customer relationship?

What European organisations can do now

  • Measure depth, not seats. Track daily use per workflow and what changed in the process, not how many licences are active.
  • Redesign one cross-functional workflow end to end. Budget for process, data and change management, not only for the tool.
  • Keep models swappable. Route work to the cheapest model that is good enough and measure the return on tokens rather than token volume.
  • Map where your revenue depends on inertia. Out-of-contract tariffs, low-interest deposits, automatic renewals and complaint handling are the first places an agent will look.
  • Prepare your channels for agents. Decide how agents identify themselves, what they may do and how your offers are read by machines, as protocols such as Google’s Universal Commerce Protocol start to spread.

Two answers to every AI question

Evans closes with a line he has used for years. Every question about AI has one of two answers: nobody knows, or look at what happened the last time everything changed. His best example comes from music. Recorded music revenue more than halved after the internet arrived and recovered only when the question changed from how to avoid buying a CD to how to get all the music there is. If the last platform shift is a guide, the advantage did not go to whoever installed the browser first, but to whoever asked what being online made possible. The board-level question for 2027 is therefore not how many employees use AI, but which question you can now ask about your business that no software could answer before.

Frequently asked questions about AI Questions

Where can I download Benedict Evans’ AI Questions?

The PDF of 86 slides is free on the Presentations page of ben-evans.com. There is no video of the November version yet. The video of the May version, recorded at SuperAI in Singapore, is on the same page.

What is new compared with AI eats the world?

The thesis that models commoditise and value moves up the stack is unchanged. New are Anthropic’s financials from its IPO filing, the first hard data on agents in commerce and a full chapter on friction, inertia and what agents do to them.

What does Evans mean by diffusion is not transformation?

That giving every employee a new tool spreads the technology without changing how the organisation works. In earlier shifts, the real change came later, from systems and processes that were rebuilt around the new technology.

Does Evans say AI is a bubble?

He does not answer that question in the deck. He shows that capital spending is outgrowing the cash flow of the largest tech companies, which makes the build-out dependent on capital markets, and that the next frontier model tends to eat the profits of the last one.