SHI Fall Summit 2025 — October 15
Brian gave this keynote at SHI's Fall Summit 2025 in Somerset, New Jersey, to an audience of IT leaders and practitioners. He deliberately kept the focus close: not where AI goes in ten years, but what's already real and what IT needs to do about it in the next 12 to 18 months.
The real action is in user land, not IT projects
Most enterprise AI projects fall into two failure-prone shapes: horizontal-but-shallow (chatbots and copilots rolled out to everyone with no real guidance) or vertical-but-slow (a single line of business rebuilt from scratch, over budget and over schedule). Meanwhile workers are quietly finding their own tools and using them for real work. "For a lot of these folks, ChatGPT or Claude or Gemini is their primary interface into the work they do."
From shadow IT to shadow strategy
Calling this "shadow AI" undersells it, Brian argued — an idea he's written about separately. Past waves of shadow IT (Gmail, Dropbox, personal iPhones) were about better tools. This wave is different — workers are handing AI the strategic thinking they used to do with human colleagues. "This isn't about tools anymore... it's full-on shadow strategy." The shift is subtle from the outside: workers are still opening the same chat window they were in 2023, but what they're asking it to do, and how much they're leaning on it, has quietly escalated from "help me write this email" to full colleague.
Every browser just became an AI browser
The AI-browser wars — Comet, Dia, Gemini bolted onto Chrome, Copilot bolted onto Edge — all happened in the past three to six months, not in some future roadmap. And the platforms behind them stopped being chatbots: Claude's connector list now reaches into Google Drive, OneDrive, calendars, Salesforce, and any custom MCP server; ChatGPT added its own app ecosystem the same week as this talk. "These things are so much more than just chatbots... it's not just another application. It's like: I use Claude, and I only drop down into the applications if Claude can't do what I want to do."
The OSWorld benchmark, and what it means that AI can now just use a computer
For applications that don't have modern connectors, computer-using agents (from Anthropic, Microsoft, OpenAI) operate the mouse and screen the way a human would. OSWorld — 360 plain-English desktop tasks — is the benchmark for how good they are at that. Humans clear about 72%. AI agents scored 12% in April 2024; 61% by August 2025; 70% by this talk in October. "Humans' ability to operate a computer is not that much better or different than an AI's ability to operate a computer."
The post-application realization
The turn in the talk: Claude editing Excel files directly, without ever opening Excel, because spreadsheet files are just structured data and AI is good at structured data now. "It's actually editing the Excel file directly — like CRISPR gene editing." If an AI can draw a chart from data on request, the case for keeping the application around — rather than just the data and a capability to act on it — gets thinner every month. Brian frames this as a four-stage realization he went through himself in real time, ending with: the application might not be the thing you actually need.
Applications you can no longer manage the old way
If any worker can create a fully functional "application" just by asking an AI to automate something — one that might run for fifteen seconds and never exist again — the old application-management playbook (procure, test, approve, audit each app) stops being viable. The unit of management has to shift from the individual application to the environment those ephemeral creations run inside.
AI as colleague, not tool — and why the rollout has to be bottom-up
Brian's closing frame: think of AI not as an application you deploy, but as an extension of each worker — brought in individually, the way workers already bring their own phones, and secured the same way you'd secure a human worker (guardrails, audits, data-loss prevention, session recording), not the way you'd standardize a corporate laptop image. Mandating a single company-wide AI platform from the top, he argued, will fail the same way mandating a single laptop model would.
His close: "This is happening whether you like it or not. By the way, you don't get a vote. You don't have to like it, but this is happening. You're not falling behind. Keep on doing what you're doing."
This is a summary. The full talk transcript lives in Brian's second brain: read it on GitHub.