EUCTech 2026 · Aboard the Oslo-Kiel cruise, Norway · June 3, 2026
Day 1 keynote, ~45 minutes, for a room of end-user computing veterans. Later shared publicly as a special solo edition of the Citrix AI Hotsheet podcast.
Watch
This keynote was released as Citrix AI Hotsheet Episode 2 (audio and video):
The narrative around AI flipped. Last year it was "AI isn't worth it." Now it's "AI is too expensive" — companies imposing usage and spend caps. That flip is itself evidence that AI works, maybe too well. Two clocks are running: capabilities (what AI can do, still climbing) and diffusion (how fast organizations absorb it, hitting a wall). Almost every "AI ROI is failing" headline lives on the diffusion clock, not the capability clock.
The mechanism is AI-caused congestion. A worker who can suddenly produce more doesn't help if the rest of the business can't absorb the extra output — the bottleneck just moves. It's the same thing developers already see: AI writes code faster than it can be reviewed, tested, and secured. "It's not that AI can't do it. It's that companies can't absorb it."
The invisible 80%
Emails, documents, and transcripts are the visible outputs of knowledge work — maybe 20% of it. The real bulk is invisible: the thinking, the reasoning, the judgment that comes from years in a role. "These aren't the knowledge work itself. They're the outputs of knowledge work."
IT lives in the visible 20%. The invisible 80% is where the business-transformation consultancies live. AI has to transform all of work, which means EUC has to step outside its 20% bubble. That's why AI hasn't produced instant ROI: most efforts only ever touched the visible 20%.
How AI enters work: the seven steps (2026 edition)
An update to the seven-stage roadmap Brian first published in June 2025. Every worker is somewhere on this path:
Faster search — a better Google, one-and-done prompts. Where most workers (and most AI skeptics) still are.
Thinking partner — back-and-forth, uploading documents, dictating. Enabled by smarter models, longer context, and document upload.
Cognitive extension — the second brain. Instead of bringing papers to the AI, give the AI access to everything in a context vault it reads and writes. Brian started here in January 2026.
Multi-tool agent — connect the AI to the world: MCP, connectors, browser control, and computer-using agents. OSWorld: the median human scores ~72-74, AIs now score in the 80s. Not to automate — to extend the knowledge worker's reach across all their tools and data.
Fleet of AIs — multiple AIs working and talking to each other and to other systems' AIs.
The pod — the new unit of work: a worker plus their AIs. Work is no longer bounded to business hours. Three worker types emerge: cognitive owners (context and judgment), cognitive operators (run the fleet), and cognitive curators (maintain context and skills).
The published self — an optional fork: publish your context vault so others can subscribe (brianmadden.ai). Underpins a future of blended expert contexts inside companies.
The current EUC model, audited
Today's EUC model assumes one person, one screen, one set of apps, one set of hours. AI breaks those assumptions. But most of EUC transitions over rather than disappearing. Brian translates each primitive into its AI-era successor:
VDI stays — used by humans and AI workers. Grounded in "AI is not eating all software": shallow UX-wrappers struggle, middle horizontal SaaS gets squeezed, deep regulated systems of record (Epic, SAP, Oracle, mainframes) don't move. AI uses those apps; it doesn't replace the deep ones.
Image management → skill management; app virtualization/layering → skill virtualization/layering; profile management → context management. Same inheritance and layering mechanics, a new atomic unit (a knowledge nugget or a skill).
Group policy → agent policy.
Session recording → agent observability, and beyond it, cognitive observability — tracing what context shaped the AI's judgment, not just governing the agent. Layered context vaults can go rogue or be poisoned.
Performance management → token management. Token routing: the same Excel task can cost 200K tokens (a computer-using agent driving Excel), 100K (browser automation), 10K (reading the .xlsx directly), 5K (a Python script), 2K (reasoning in context), or zero (handing it to a human). Plus where the model runs and device-trust and PII checks.
Endpoint management stays — but endpoints become wearables and ambient devices, with the AI generating the UI on demand wherever you are.
The workspace client → the cognitive workspace — icons for tasks and knowledge sources, not apps.
The control plane stays and gets bigger. Something has to manage the humans, agents, cognition, models, routing, and security.
The throughline: the words change (agent, cognitive, skills, context) but how and why we do the work doesn't. The core IT skills transfer. The job gets bigger.
Book one and book two
"The Last Chapter of EUC" is the last chapter of book one — the past 30 years, all spent on the visible 20%. Book two is the next 30 years, on the invisible 80%: thinking, reasoning, judgment, taste. Page one of book two is step three — build your own second brain. "You have to manage one before you can manage thousands." Book two doesn't have a title yet, and the people who run workspaces are the ones who get to write it.
Starter prompt
Page one of book two is building your own second brain. Here's the starter prompt: