EUC World: Independence · Silver Spring, Maryland · October 22, 2024 · opening keynote with slides, delivered in the same building where BriForum started twenty years earlier.
Twenty years after the first BriForum keynote on this same stage, a tour of EUC’s past, present, and future — and an argument that the Windows desktop, of all things, is becoming the API for AI agents.
Twenty years, same stage
BriForum ran roughly twenty times over twelve to fifteen years, and of the people who spoke at the very first one in this building, a handful are still in the room. The photos from 2005 are almost comic in retrospect — a talk called “The Future of Citrix” that predicted Citrix Zeon 12, 1024×768 resolution, and finally solving session printing. The common thread across two decades isn’t any specific technology; it’s the community in the room.
The career arc, before BriForum even existed
The real starting point predates BriForum by a decade: 1994, fixing local businesses’ PCs by hand — installing RAM, sound cards, IRQ and DMA jumpers, one bad network cable capable of taking down the whole office. From there the career moved through Novell NetWare, desktop engineering, SMS, Citrix, ASPs, blades, VMware, VDI, HCI, MDM, MAM, EMM, consumerization of IT, digital workspace, future of work — a book, a certification, a conference for each one. AI is just the next line on that same list.
Imaging, and the faster horse
The original daily grind of desktop administration was gold-master imaging: build it, deploy it, patch it, watch it get janky, reimage, repeat. Systems Management Server, from 1994, didn’t change that circular model — it just automated it, which meant patch pushes could now break everything at once instead of one machine at a time. It was a faster horse: automation without a fundamentally new approach.
Citrix, and why MAPS mattered
Citrix’s actual 1998 marketing message was MAPS: Management, Access, Performance, Security. Install an application once instead of on every machine. Access it from anywhere, on any device. Good performance for client-server apps running centrally in the data center. Security, because only pixels crossed the wire — no data at rest on the endpoint. Nobody used words like “modernization” or “remote work enablement” back then. Citrix won because it was a fundamentally better way to do the job, full stop.
VMware, VDI, and the long complicated love affair
VMware’s pitch was Citrix without the complexity, since it was “just” regular VMs — which promptly produced VM sprawl everywhere. VDI followed the same up-and-down arc for two decades: too slow in the early 2000s, suddenly viable once ENV storage and copy-on-write arrived by 2012, then genuinely great once GPUs got added. COVID in 2020 forced the issue completely — VDI in the cloud, paid for every month forever, became simply how things are done, something nobody would have predicted sitting in this same room in 2005.
Mobility’s slow arrival
Real mobile work took roughly five years to reach the average office worker after 1994, arriving via the 1999 RIM BlackBerry 850 — push email, no login required, plain-text Outlook settings so nobody could tell you weren’t at your desk. It ran on a 386 chip, and Pat Gelsinger, later Intel’s CEO, snuck his initials into the design in three places as a staff architect on that chip. The iPhone launched eight years after the BlackBerry with no video camera, no apps, no Exchange integration, and the business world dismissed it as a toy. Real business ran on BlackBerries — until Satya Nadella’s first public act as Microsoft CEO in 2014 was putting Office on the iPad, eight months before Office reached iPhones at all.
Consumerization of IT, in one image
Once employees could get native Office apps and better tools on their own devices, IT’s old answer — “how do I turn this iPhone back into a BlackBerry” — stopped working, because employees started just buying and using whatever they preferred regardless of IT’s blessing. The honest origin story: CIOs got mad they couldn’t get Netflix on their BlackBerries, and iPhone adoption in the enterprise took off from exactly that irritation. The lesson holds: workers gravitate to the tools they want, and IT’s job is to orient around that rather than fight it.
Where that left device management
The eventual accommodation: iOS or Android for mobile, Mac or Windows for desktop, VDI or Citrix for delivering Windows apps regardless of the endpoint. Unified Endpoint Management promised one platform to rule all of it, but the vision broke down against real-world complexity — too many use cases, too many vendors, too many partial fits, managed by different teams in different countries under different regulations. What actually matters to a CIO isn’t the platform choice; it’s security and compliance (don’t get ransomed, don’t end up on the news), operational visibility (are employees able to work), and digital employee experience — the capabilities layer above the technology stack, not the stack itself.
Why Windows isn’t dying, again
Every few years since 2005 has produced a new “Windows is dead” story — ASP.NET, Java, AJAX, HTML5, VDI, Chromebooks — and the market-share data, zoomed out far enough, shows a decline so gradual it runs past the year 2060. The old joke about cockroaches, Twinkies, and devices still running Windows after any apocalypse still lands, because Windows keeps outliving its obituaries. What matters isn’t Windows the brand; it’s the desktop as a model — the framework that lets applications talk to each other, copy and paste, authenticate, audit, and print. That framework isn’t going anywhere even as more computing happens elsewhere.
The AI turn
ChatGPT launched in 2022, but serious engagement with it didn’t start until early 2023 — six months of using it as a limerick generator before the “wait, you can ask it work questions” moment. Before that, the entire cultural frame of reference for AI was science fiction: 2001, Terminator, Her. This is, by Brian’s own account, the single biggest shift of a career that started right as the internet itself was becoming mainstream. Today’s ChatGPT is the DOS prompt of this era — text in, text out, primitive relative to where the interface is inevitably headed.
Recall, Copilot Vision, and the drip
Microsoft Recall’s initial version just took desktop screenshots every five seconds for later search. The trajectory described months before this talk: understanding intent rather than just archiving pixels, integrating with Microsoft’s 150-plus Copilot products, then eventually taking actions — moving the mouse, doing the boring parts of a job first. Three weeks before this keynote, Microsoft shipped Copilot Vision, a separate Edge-browser feature limited to pre-approved sites, letting the AI co-browse and help fill out forms — Microsoft’s first real step from watching the screen to touching it.
Small, fast, specialized models
Simon Willison’s viral example: recording a 35-second screen capture of scrolling through invoices, dropping the video into an AI studio, and extracting structured JSON data for under a tenth of a cent, no engineering required. Purpose-built small models that only interpret UI elements can process a screenshot faster than a video frame rate — cheap and fast in a way a general-purpose model like ChatGPT isn’t. A parallel line of research from the same month reads user intent directly from mouse movement and clicks. The likely shape of the field: not one enormous general model, but many small specialized ones wired together, echoing how distinct regions of the human brain handle different kinds of cognition.
Windows Agent Arena
Microsoft’s own benchmark defines roughly 150 tasks an AI agent should be able to do on a Windows desktop — change a setting, send an email, install something — scored against shared VM templates. At the time of this talk, agents were succeeding about 17% of the time. A day before the keynote, Microsoft announced Copilot AI agents would begin tackling real employee tasks within weeks.
The desktop as the API
Every application already lives on the desktop, wired together with thirty years of institutional customization — copy-paste, plugins, custom images, internal tools that no automation vendor like Zapier will ever have connectors for. Rather than reinventing that integration from scratch for AI agents, point the agent at the desktop itself: it inherits every application, the existing security context, audit trails, and management tooling for free. That, more than any specific product, is the argument for why the Windows desktop becomes the entry point for enterprise AI rather than a purpose-built cloud-native agent stack.
What moving the desktop to the cloud unlocks
Once that desktop is remoted, its pixel stream can be split off for more than just display: screen and text archiving for legal or forensic purposes, DEX analytics for performance, workforce-intent analysis to understand what people are actually doing, compliance recording with GDPR-aware redaction — all without the AI agent needing to run inside the VM itself, and without rebuilding VDI sizing or architecture from scratch.
The close: 2025 is the year of VDI
If AI agents need an always-on, remotely accessible desktop to operate from, VDI stops being a legacy technology to defend and becomes the actual on-ramp for enterprise AI. The same MAPS argument that sold Citrix in 1997 — centralized management, access from anywhere, consistent performance, no data at rest — applies just as cleanly to AI agents in 2024. This isn’t happening because AI agents are good yet — most are still slow and error-prone — but the direction is clear, arriving in small increments rather than one dramatic leap, and the entry point for all of it is the desktop this room has spent thirty years building.
Key formulations
“I’m not worried about Windows in the box. I’m worried about me in the box.”
“The desktop is the API for all of your AI agents.”
“2025 is going to be the year of VDI because of AI.”
“We already have this. It’s already done.”