AppManagEvent 2025 · Utrecht, Netherlands · October 10, 2025 · closing keynote with slides.
Applications are dissolving. AI platforms are becoming the real gateway to work, connecting to modern apps, browsing web apps, and increasingly skipping apps to edit files directly. App management has to evolve into capabilities orchestration.
The ROI headlines miss the point
Nearly eight in ten companies report using generative AI, and just as many report no significant bottom-line impact. A widely-cited (and widely-discredited) MIT report claims 95% of generative AI pilots fail. The discredit doesn’t stop it from being quoted everywhere, because it’s useful for exactly the narrative it drives: that corporate-mandated, IT-led AI — the $500,000 AI service desk that still can’t reset passwords, the security tool generating 10,000 unreviewed false alerts a day — isn’t delivering. Every CEO wants an AI strategy so as not to be Blockbuster in the Netflix documentary, and it rolls downhill to the CIO and CISO. But this entire framing misses where the actual action is.
Where the real activity lives
Thirty-one years in end-user computing teaches you the users are what matters. Regular workers are using ChatGPT, Claude, Gemini, Copilot, Mistral every day — chatting, brainstorming, analyzing documents, drafting strategy, asking for career and personal advice, generating images and video. This summer brought an AI browser war: Chrome got Gemini, Edge got Copilot, Dia arrived as an AI-native browser. Every mainstream browser now has an AI engine watching what you do with the web.
Not chatbots anymore
Claude’s connectors panel reaches into SaaS tools, desktop applications, and anything with an MCP endpoint. ChatGPT just added its own apps ecosystem. Calling these tools “chatbots” undersells them — AI productivity platforms, agent platforms, whatever the name, they’re becoming the actual gateway to work, reachable by voice or text, watching screens, generating video and audio, present as web, Windows, Mac, iOS, and Android apps simultaneously.
Agent bridging and computer-using agents
For apps with no API, ChatGPT’s agent bridging lets the AI operate a browser directly — watching becomes driving. For apps with no web interface at all, computer-using agents from every major AI vendor process the screen and operate the mouse and keyboard the way a human would. The OSWorld benchmark, 369 English-language desktop tasks, put humans at 72% completion when it launched in April 2024. AI scored 12% then, 61% by August 2025, 70% by October. These agents are still slow — screenshot, analyze, act, screenshot again — but specialized models built just to read UI screenshots can now process a full HD frame in under 30 milliseconds, real-time speed on the horizon the same way self-driving cars went from crawl-stop-crawl to smooth.
The personal/work line
Workers trust their personal ChatGPT or Claude with almost anything — the same way a personal phone feels safer to overshare on than a work-owned one. The future probably isn’t one universal work chatbot; it’s closer to a work persona and a personal persona for AI, the same separation that already exists for devices.
The workspace doesn’t fundamentally change
A workspace is apps and data, identity, security, and context. For twenty years that’s meant a human worker attached to that workspace. Now the worker attached to it can be human, AI, or both together — and IT’s job barely changes in kind. The same guardrails built for human mistakes (bias, deletion, hallucination-equivalent errors) apply to AI workers too. One genuine new wrinkle: because AI can now log on and keep working after a human closes their laptop, VDI suddenly has a real justification it didn’t always have — a persistent cloud desktop an AI can keep using after the human logs off.
Two workers, one workspace
Increasingly it’s not “human or AI” attaching to the workspace — it’s the AI platform sitting in the middle, with humans talking to their AI and the AI talking to the backend systems. The human interface can be keyboard and mouse, voice through earbuds, or a video avatar; the AI-to-backend interface can be pixel-scraping an old HDX session, reading the DOM directly, or calling a modern API. Over time, everything leaving the workspace stops being built for humans and starts being built for AI to consume, with the AI generating whatever human-facing interface is needed on demand.
The 1990s parallel, again
The web was supposed to kill Windows apps in the 1990s: instant access, any device, one place to update, better security. The catch was rewriting every application from scratch. Citrix exists because corporations don’t chase the theoretically cleanest solution — they keep what works and wrap it. The current AI transition is the same shape: a temporary, slightly awkward period where AI has to use human-style interfaces (browsers, desktops, mice) as a workaround, the same way humanoid robots are built human-shaped just because they have to operate in a human-shaped world.
Applications built by asking
Claude and Copilot can now edit Excel, Word, and PowerPoint files directly, because those formats are just XML underneath — no application needed, just sufficient understanding of the file format. Once that’s true, the old ten-users-per-app ratio inverts toward ten apps per user: every worker becomes an accidental citizen developer, asking their AI to pull receipts and file an expense report every Monday without ever thinking of it as “building an app.” These creations are ephemeral, undocumented, constantly changing implementation (Python one week, Rust the next), highly interconnected, and often invisible to IT entirely — not shadow IT so much as alternate universe IT, with no real relationship to how app management works today.
Near-term shifts
When creating an app costs less than the meeting to discuss creating it, the whole economics of applications changes — the value moves from the code (now nearly free) to the specification of what the code needs to do. Security has to shift from application-level to environment-level, since vetting every ephemeral AI-generated app individually is impossible. Radical observability — session recording, tight guardrails, short leashes — becomes the substitute for pre-approval. Citizen development needs a supported path built for ordinary workers asking AI to do things, not just power users vibe-coding. And policies need to survive an application lifecycle that can compress to seconds.
Longer-term shifts
Apps were built for human consumption; AI doesn’t need that interface, only the data and the capability to act on it. Traditional UI erodes in favor of AI-native interfaces, with legacy apps surviving as long-tail exceptions reached through remote protocols. Context — understanding what a worker is actually trying to do, across devices and data types — becomes the decisive capability. And the discipline itself shifts from application management to capabilities orchestration: not locking down individual apps, but understanding and constraining what tools workers’ AI can actually do.
Close
Managing applications has an expiration date, even if it’s a distant one. What survives and appreciates is expertise in managing digital workspaces — the environment everything else has to run inside and be secured within. This is happening regardless of preference, and it’s accelerating: work that used to take 18 months to catch up to now takes two.
Key formulations
“When creating an app costs less than the meeting to discuss creating it, everything changes.”
“This is not shadow IT. I call this more like alternate universe IT.”
“It’s like the protocol droid with three million languages — it now talks Excel directly.”
“You cannot vet 10,000 applications. But you can secure the workspace where they operate.”
“Context is the killer feature moving forward. Whoever masters this wins.”
“Your job managing applications has an expiration date. Your skills in managing digital workspaces are more valuable than ever.”