These are the notes from a talk I gave in Copenhagen to the DanofficeIT team, customers, and large Danish enterprise accounts. A largely similar version of this material is on video (the EUCTech 2026 keynote); this smaller, conversational format surfaced sharper formulations and Q&A exchanges that didn’t come up on that stage, especially around second-brain trust, data integrity, and the consulting business model.
The AI narrative has flipped from “AI doesn’t work” to “AI costs too much,” and both are diffusion stories, not capability stories. To get real AI ROI, organizations have to reach the invisible 80% of knowledge work — thinking, judgment, reasoning — not just the visible 20% that IT has always managed. The seven-phase roadmap is how you get there. For EUC professionals, every primitive already known carries forward; the words change but the job gets bigger.
Why Citrix’s 37-year pattern applies to AI
Citrix has spent its entire history wrapping existing technology to give it modern capabilities without requiring organizations to rewrite everything. In the 1990s, that meant giving Windows apps the reach of web apps. Now it means giving AI the same access to applications that human workers already have, without rewriting those applications. AI is going to navigate computers regardless; the question is whether it does so through a governed, policy-compliant Citrix session or through an ad-hoc consumer tool with no audit trail.
The AI narrative flip and the diffusion problem
Six weeks before this talk, the dominant story was: eight in ten companies using AI, a 95% failure rate, no ROI. By the time of delivery, the story had flipped: AI costs too much, companies rationing tokens, corporate reeling from AI bills. That flip is evidence AI works. Two clocks run at different speeds — capabilities, still climbing, and diffusion, hitting a wall. Most “AI ROI” disappointment is a diffusion problem. AI-caused congestion: a knowledge worker who produces ten reports instead of four doesn’t help the company if the business can only absorb four. The bottleneck just moves. Eventually the CFO asks: does this increase revenue or lower expenses? If neither, why are we paying for it?
The invisible 80%
Emails, documents, and meeting transcripts are maybe 20% of knowledge work — the visible outputs, the digital exhaust. The other 80% is invisible: thinking, reasoning, judgment, skill, experience. A lot of knowledge work is staring out the window looking at birds. EUC and IT have lived entirely inside the visible 20%. AI changes that, because AI can now do the invisible parts too — which makes them digital, and therefore manageable. EUC’s universe just got much larger.
The seven phases
Every worker is somewhere on this path, and the key framing is that you can only see one step ahead. Someone on phase one (faster search) can just about see phase two (thinking partner), but phase three (cognitive extension) looks like a different planet. People who dismiss any of this are standing at the phase where the next one still seems visible and everything beyond is incomprehensible from where they stand.
Faster search is one question, one answer — most of the world is still here. Thinking partner is longer conversations, loading documents, real back-and-forth. Cognitive extension, the second brain, is the inversion: don’t take your documents to the AI, bring your AI to your documents — not a database, actual folders of markdown files in GitHub that the AI reads and writes. Within a day or two of starting, it was surfacing connections across things mentioned weeks apart. Multi-tool agent connects that cognitive extension into actual tools: MCP, browser control, desktop computer-using agents — on the OSWorld benchmark, humans score around 72, models now score 80–85, extending reach rather than automating it away. Fleet of AIs is multiple AI systems talking to each other — your AI talking to your organization’s AI, or a partner like DanofficeIT building a bot that talks to Citrix APIs and carries all of DanofficeIT’s accumulated best practices for its customer base. The pod is the new atomic unit of knowledge work: one worker plus their AI fleet, context vault, and skills, operating continuously — three worker types emerge, cognitive owners (context and judgment, the source of expertise), cognitive operators (who run the fleet), and cognitive curators (who maintain the skill and context libraries), and the bottom two look a lot like advanced IT work. The published self, an optional fork, means taking your context vault and making it subscribable — mcp.brianmadden.ai does exactly this, letting any AI tool connect and access the full knowledge base.
Why automations aren’t the path
Task automation only touches the visible 20%, the digital exhaust. If a task were easily automatable, it would have been automated ten years ago with RPA. The cognitive extension approach means the AI has access to everything and context about everything, so it can help with any of the work, not a scripted subset. The AI isn’t told to “do the expense report” — it’s asked to help prepare for an event, and it pulls the schedule, emails, calendar, and CRM on its own, because it has full context and knows how to reach those systems.
Doing this in Citrix today
All of this runs on Citrix today, without new features. Install Claude on a Citrix VDA, publish the application in Workspace. Create a second user account — “Brian Madden Robot” — with read-only access; the agent logs on as that user. Session recording stays on for everything the robot does, 100%, always, because the robot doesn’t have privacy rights. Workers’ session recording has caused real scandals (Microsoft Recall, Facebook); the agent doesn’t care. App Protection on. DLP on. Chrome Enterprise Premium with its own managed profile. The robot can read everything it needs; it can’t send emails or delete calendar events. All of it is in production on existing Citrix infrastructure today.
EUC primitives translated
Every EUC primitive carries forward into its AI-era successor — it’s a find-and-replace of users, profiles, apps, policy, and sessions into cognitive owners, context, skills, agent policy, and agent sessions. VDI stays, used by humans and AI workers alike. Image management becomes skill management. App virtualization and layering becomes skill virtualization and layering, with the same resultant-set logic running company-wide down to departmental down to individual. Profile management becomes context management. Group policy becomes agent policy. Session recording becomes cognitive observability. Performance management becomes token management. Endpoint management stays, with the AI generating the UI on demand on whatever device is nearest. And the control plane stays — and gets bigger.
Token management
Token consumption scales dramatically by phase: roughly 100K tokens a day at faster search, 1M at thinking partner, 10M at cognitive extension, 100M at multi-tool agent with screenshot processing, 1B at fleet, 10B at always-on pods. Brian used 291 million tokens in his first month of cognitive extension; someone he follows reports 20–30 billion tokens daily in his agentic system. Tokens are supply-constrained, so the job of token management is to maximize economic value per token, not minimize spend — routing a given task to a computer-using agent driving Excel (200K tokens), browser automation (100K), reading the .xlsx XML directly (10K), a Python script (5K), reasoning in context (2K), or just handing it to a human (zero). Model choice, where it runs, device posture, PII exposure — all factor in. This routing is an IT governance layer that didn’t exist two years ago.
Q&A: on trusting your second brain’s data
A concrete illustration of the integrity problem: Brian’s AI built a profile on a colleague based on meeting transcripts. Because he only records disagreements for the AI to process — nobody dictates the meetings where everyone agrees — the AI had flagged an adversarial relationship with a colleague he’s 99% aligned with, and was quietly filtering comments through that incorrect profile. He caught it only because something felt off, went directly to the file, and deleted the entry. The mechanism: selection bias in what gets captured creates systematic distortion in the AI’s model of the world. Talk to your AI only about problems and conflicts, and it builds a problem-and-conflict-dominated worldview. The fix is file-based storage you can read, inspect, and edit directly — a vector database abstracts this away; a folder of markdown files doesn’t. These kinds of questions are exactly the ones that have to be solved. The knowledge-integrity problem was always a challenge for companies, but it was never something end-user computing thought about. Now it’s EUC’s problem too.
Q&A: on who owns your second brain
“If a folder full of text files and a twenty-euro ChatGPT subscription can do my whole job, I want to know about that first. For real.” After six months of working this way, the understanding of what the AI can and can’t do, and what value a person actually provides, gets much clearer. The context vault also becomes more valuable as AI improves — the same vault produces better outputs across Claude 4.5, 4.6, 4.7, 4.8. That’s unusual: most technology assets decay as AI advances, while the context vault compounds.
Q&A: on the consulting business model
“The days of a consultant coming in and leaving the PDF after a project — those days are dead.” What replaces it: the consultant develops a living context vault as part of the engagement, and clients’ AIs plug directly into it. DanofficeIT’s best practices, industry knowledge, and configuration guides become a subscribable second brain that’s always current. The consulting product shifts from a deliverable at project close to an ongoing knowledge relationship.
Q&A: on cognitive observability
Session recording can serve not just to observe agent behavior but as a cognitive-provenance system — tracking which source document or meeting transcript a piece of AI reasoning actually came from. Provenance matters in regulated industries, and research on this is active across the field. Session-recording infrastructure is already the right scaffolding for it.
Key formulations
“If a folder full of text files and a twenty-euro ChatGPT subscription can do my whole job — I want to know about that first.”
“The days of a consultant coming in and leaving the PDF after a project — those days are dead.”
“You can only see one step ahead. Phase 3 from Phase 1 looks like a different planet.”
“My agent doesn’t care if it’s recorded.”
“Selection bias in what gets captured creates systematic distortion in the AI’s model of the world.”
“I don’t take my documents to the AI. I bring my AI to my documents.”
“The first book of EUC is 1990–2025. We are writing the first page of book two.”