From the Arrow Forum 2026 · July 16, 2026 · Munich, Germany · This is my 40-minute keynote, reconstructed from the slide deck. (It was not recorded.)
You can’t predict what a worker looks like in 2031, but you can identify what’s true across every plausible path there and build for that. Two things are near-certain: AI capabilities keep climbing, and AI diffusion — how fast organizations absorb what AI can already do — stays slow. The gap between them is the whole opportunity, and closing it is about to become the biggest expansion of IT’s job in a generation.
What do we know about the next five years?
The talk opens on “What is a worker in 2031?” and reframes it as “what happens in the next five years of work?” The spine is a running list Brian returns to and revises across the talk: AI capabilities will continue to increase; AI diffusion is slow; closing the diffusion gap will greatly expand IT; knowing how to close it will be the key to the future of IT and work.
Capabilities vs. diffusion
Two clocks: capabilities (what AI can do, still climbing) and diffusion (how fast AI gets absorbed into organizations, slow). The space between the curves is the diffusion gap. Nearly eight in ten companies report using gen AI, yet just as many report no significant bottom-line impact — the problem isn’t capability, it’s absorption. Knowledge work is roughly 20% visible (emails, documents, meeting transcripts, chats) and 80% invisible (thinking, reasoning, judgment). Traditional IT only ever operated in the visible 20%. AI changes the mandate: IT now has to reach the invisible 80%, aiming to make all of knowledge work visible and supportable.
How do most companies plan to get there today? “Develop an AI strategy”: the CEO declares “we are an AI-first company,” buy Copilot licenses for every worker, then — nothing. The missing step three is the point. Buying licenses is not a plan for closing the diffusion gap.
Stress-testing “AI capabilities will continue to increase”
A futurist doesn’t predict the future; a futurist works with probabilities and builds for what’s true across all of them. Two assumptions hide inside “capabilities will keep increasing”: that technical progress continues (probably, not guaranteed), and that the most advanced models stay available — which now depends on the US government, on whether a bubble pops, and on the Chinese government. A run of headlines from a single ten-week stretch in 2026 shows how fast frontier access is granted and revoked: Anthropic ships Mythos Preview to ~50 companies, then Fable 5 and Mythos 5; the US government forces Anthropic to cut those models off; days later the Chinese lab Zhipu answers with GLM-5.2, a full open-weight model in the Opus 4.8 / GPT-5.5 class; DeepSeek closes a $6.5B external round; the US government asks OpenAI to limit GPT-5.6 to trusted partners, then reverses and allows both Anthropic and OpenAI to ship. Frontier access is now a policy variable, not a given — and “even if the bubble pops, the technology still exists” is itself an assumption that may not hold for AI the way it held for railroads and dark fiber.
The one thing that survives every asterisk: open-weight models. GLM-5.2 is already released, MIT-licensed, and frontier-class. So the first item on the list gets rewritten — from “AI capabilities will continue to increase” to “Sonnet-class AI is real, and guaranteed to exist.” That’s the reliable floor. No government and no bubble can take it away.
Stress-testing “AI diffusion is slow” — enter the FDE
Diffusion has a mechanism for change: forward-deployed engineers. The money pouring into them, from the same stretch of 2026: Palantir’s Q1 shows US commercial revenue up 133% year over year, with the CEO crediting the FDE model; OpenAI announces a Deployment Company with $4B+ invested and 150 FDEs on day one, partnered with BCG, McKinsey, Capgemini, and Accenture; Anthropic and DXC align to train tens of thousands of FDEs; AWS commits $1B to build its own FDE organization; Microsoft announces “Microsoft Frontier Company,” $2.5B and 6,000 FDEs; Anthropic and Blackstone set up a $1.5B joint venture. An FDE is a forward-deployed engineer. To get one, you either hire one of those firms, or figure out what an FDE actually does and do it yourself: go into the organization and turn the invisible 80% into visible, encoded, usable knowledge. So the second item gets rewritten too: “AI diffusion is important; FDEs can speed it up.”
How to close the gap: the seven-stage roadmap
The mechanism is the seven-stage roadmap, with a context vault at the center from stage three on: faster search (a better Google, one-and-done prompts); thinking partner (back-and-forth, uploading documents, dictating); cognitive extension — the second brain (give the AI access to everything instead of bringing documents to it); multi-tool agent (the AI reaches into the world — computer-using agents, browsers, APIs, MCPs); fleet (multiple AIs coordinating with each other and with other systems’ AIs); pod (a human plus their AIs, running beyond business hours); and the published self, an optional fork where you publish your own context vault so others’ AIs can subscribe to it.
At the pod stage, three worker types emerge: cognitive owners (context plus judgment, the source of expertise), cognitive operators (who run the agent fleets), and cognitive curators (who maintain the context vaults and skill libraries). Daily token consumption per worker climbs by stage — roughly 100K at faster search, 1M at thinking partner, 10M at cognitive extension, 100M at multi-tool agent, 1B at fleet, 10B at the self-running pod — which is why token management becomes a first-class IT problem.
The EUC audit
The current EUC model assumes one person, one screen, one set of apps, one set of hours. AI breaks every one of those assumptions, but most of EUC transitions over rather than disappearing. VDI stays, used by humans and AI workers both — shallow UX-wrapper apps struggle, middle horizontal SaaS gets squeezed, and deep regulated systems of record don’t move, so AI still needs somewhere to run against them. Image management becomes skill management; app virtualization becomes skill virtualization; profile management becomes context management. Group policy becomes agent policy. Session recording becomes agent observability, and beyond that, cognitive observability. Performance management becomes token management. Endpoint management becomes cognitive endpoint management. The receiver becomes the cognitive workspace. And the control plane stays — and gets bigger.
The close: build your own brain
Workers still need to work. Work still needs to happen somewhere. Somebody has to make that somewhere work — safely, observably, cost-effectively. That somebody is EUC and IT, and the job is bigger than it has ever been. The next step is stage three: build your own brain. You have to feel it before you can govern it. You have to have one before you can manage thousands.
Key formulations
“A futurist does NOT predict the future. A futurist works with probabilities.”
“Nearly eight in ten companies report using gen AI — yet just as many report no significant bottom-line impact.”
“The CEO declares we’re an AI-first company. Buy Copilot licenses for everyone. Then… ???”
“Sonnet-class AI is real, and guaranteed to exist.”
“You have to feel it before you can govern it. You have to have one before you can manage thousands.”