EUCtech Denmark 2025 · Billund, Denmark · May 22, 2025 · closing keynote with slides.
AI agents will behave like knowledge workers. They’ll use the same workspaces as humans, those workspaces won’t be tied to devices, and this usage pattern will last far longer than anyone expects — the first full public airing of the workspace-as-control-plane thesis.
Which AI, exactly
“AI” means several different things at once: AI capabilities baked into products people buy, AI in the apps a company builds internally, and AI shaping end-user workplaces directly — workplace AI, the specific kind this talk is about. The question isn’t whether AI will affect the workplace; it’s already happening, today, not as some future tsunami but the way a tsunami actually arrives — a little water at your shoes, then your socks, rising bit by bit until suddenly it’s everywhere.
Small percentages, big aggregate
Every job is a series of tasks, and today AI can already do some of them — often only 5% or 6% of any one person’s job. Six percent for one person isn’t interesting. Six percent across an entire company is very interesting, and it’s part of why layoffs are getting blamed on AI even though AI isn’t doing most of anyone’s job yet — it’s doing a little of a lot of people’s jobs, and that adds up at the organizational level.
Workplace AI is worker-led
This transformation looks like the earlier consumerization of IT wave: bottom-up, not dictated from above. AI tools live inside the same workspace workers already use — as apps, web pages, sidebars — and the balance between AI and human effort isn’t a switch, it’s an analog dial that shifts back and forth. Personally, the balance hasn’t reduced Brian’s workload; it’s increased his output. Running ChatGPT, the Microsoft stack, and the Google stack simultaneously, he estimates doing the work of what would have been a team of three people four years ago — and expects that number to keep climbing as the tools keep improving.
From chatbot to agent
AI moved from a separate application you copy-paste into, to something watching your screen, to something making suggestions, to something taking small permitted actions while you watch, to something you trust enough to leave running while you get coffee. Alongside that, the models themselves evolved from general text chatbots to domain specialists, then gained image, audio, and video capability, then genuine visual reasoning — GPT-4o-class models that can solve a maze or identify a country from a street photo by reading the scene, not just pattern-matching to a fixed answer. Reasoning and multi-step planning followed, letting a model feed the output of one task into the next. Sam Altman’s September 2024 essay “The Intelligence Age” put it plainly: “AI models will soon serve as autonomous personal assistants who carry out specific tasks on our behalf.”
Why build a custom agent when the computer already exists
Asking an AI how to build a task-specific agent surfaces a familiar answer: stand up an API platform, wire in tokens, host it in the cloud. But the agent only needs to do what you already do on your laptop every day. So skip the custom backend and let the AI use the computer directly — the Computer Using Agent, or CUA: an AI that observes the screen, moves the mouse, and types, the same way a human operates a machine. Every major lab has one — Claude computer use, OpenAI Operator, Microsoft’s and Google’s equivalents — and you can watch them think in real time: pointed at LinkedIn and asked to count likes on the latest post, Operator visibly reasons through “LinkedIn is a website, I need to go there” before clicking around.
The benchmarks, month over month
OSWorld, an open desktop-task benchmark launched in March 2024, put humans at about 72% completion. The best AI scored 1% at launch; by this talk, in May 2025, it was at 42%. Microsoft’s Windows Agent Arena tracks similarly — humans at 74%, the leading agent starting at 19% and reaching 52% within about a year (with a caveat that the top model wasn’t independently verifiable). The specific model behind the OSWorld leader, UI-TARS, runs locally on a beefy laptop — not a cloud service, something you can watch click around on your own machine.
Why micro-apps failed, and CUAs won’t repeat the mistake
Micro-apps — the Citrix, VMware, and industry-wide bet on purpose-built connectors for individual workflows — never scaled because every workflow needed custom building, not every system had a connector, and someone had to decide, build, and maintain each one. Organizations run thousands of workflows; covering a handful with bespoke micro-apps while ignoring the other nine thousand isn’t a strategy. Computer-using agents solve this the way humanoid robots solve general-purpose labor: instead of a bespoke robot per task — a drilling attachment, a vacuum attachment — you build one robot shaped like a human, because the world is already built for humans, and it can pick up the drill, the iron, and the vacuum you already own. The same logic applies to AI agents in knowledge work: instead of engineering a custom interface per task, let the agent use the existing human-shaped desktop, browser, and applications already there.
The 1990s parallel, again
Wrapping an old, human-shaped interface in a temporary layer to modernize it is exactly what Citrix did in the 1990s to Windows apps stuck on a single desk — delivering Management, Access, Performance, and Security (MAPS) without rewriting anything. People called that dead by 2000, certain everything would be rewritten as web apps. Decades of “X is killing Windows” followed — Java, Flash, AJAX, Silverlight, HTML5, SaaS, Linux, Chrome, mobile apps, PWAs, cloud, UWP, Electron — and Windows outlived every one of them. “AI is going to kill Windows” gets the same response: heard it before, thirty-one years running.
The workspace, defined
A workspace is apps, data, worker identity, security, and context, wrapped together — inert until a worker connects and activates it. For thirty or forty years that worker has been human. From an IT administrator’s standpoint, there’s genuinely no meaningful difference between a human worker and an AI worker connecting to that workspace: same guardrails, same performance and security boundaries, same auditing. It’s just a remote keyboard and mouse either way. When a human closes their laptop lid at the end of the day, that device disappears — but if the workspace also lives in the cloud (VDI, DaaS, or even a small managed Linux container running Chrome Enterprise for a BYO laptop), the AI worker can just keep running there after the human logs off.
The core thesis
AI agents will behave like knowledge workers. They will use the same workspaces as humans. Those workspaces will not be tied to devices. And this usage pattern will last longer than anyone expects — not because it’s the theoretically best architecture, but because it doesn’t require anyone to rebuild anything, and it inherits every security, auditing, and management tool already in place. This is happening whether anyone believes it’s good or not.
Personas, not an AI persona
The familiar EUC persona exercise — modern worker, task worker, knowledge worker, power user, developer, executive — doesn’t get a new “AI persona” bolted on. AI is a layer inside every existing persona, and every AI itself varies: big model or small, on-prem or cloud, from Microsoft, OpenAI, something custom-built, something a GSI supplied. All personas carry both human and AI elements now; there’s no separate AI category.
Close
AI agents are real and evolving quickly. No full environment rebuild is required, no wholesale rethink of EUC strategy — just adaptability. Every persona is affected, not equally, and that’s fine. The workspace remains the point of orchestration whether it’s physical, virtual, or entirely invisible because the AI is generating the interface from scratch on whatever device happens to be in front of a worker. This is happening regardless of preference. The good news: nobody is behind. Everyone in the room is positioned to lead the next stretch of this industry.
Key formulations
“AI is not coming. The AI is already in the building.”
“I am a super-worker… I think would be a team of like three people.”
“Why do humanoid robots use the same form factor as humans? The world is built for humans.”
“AI agents will behave like knowledge workers. They will use the same workspaces as humans.”
“This is happening whether you like it or not.”