A deliberately near-term view: not where AI goes in ten years, but what IT actually needs to know and do in the next 12 to 18 months.
Two failing project categories
Most enterprise AI projects fall into one of two shapes: horizontal and shallow (chatbots and copilots rolled out to everyone with a “good luck figuring out how it works”), or vertical and deep (a single line of business rebuilt around AI, expensive and slow). Both categories run over budget and under-deliver. The real activity is happening below both of them, in user land — what workers are quietly bringing and doing on their own.
From shadow AI to shadow strategy
Workers are finding ChatGPT, Claude, Gemini, and Grok on their own, and for many, one of those tools is now the primary interface into their job, ahead of whatever portal or application the company built. Calling this “shadow AI,” the way earlier waves got called “shadow IT,” undersells it. Shadow IT was Gmail working around attachment limits, iPhones replacing BlackBerries, Dropbox replacing VPN file shares — workers reaching for better tools. This is different: workers are brainstorming strategy with ChatGPT the way they used to with human colleagues. It’s full-on shadow strategy, not shadow tooling.
Helper to colleague
AI has evolved from a subtle helper (”help me write this email”) into something closer to a full colleague, and the shift is invisible from the outside — workers are still opening the same chat window they were in 2023, but what they ask it, how much they lean on it, and where they go first by default have all changed underneath.
The summer of AI browser wars
Capabilities are breaking out of the chat window entirely. New AI-native browsers (Comet, DIA) arrived alongside old browsers getting AI bolted on (Chrome with Gemini, Edge with Copilot) — every mainstream browser now ships with AI built in, not as something coming, but as something that happened in the past three to six months.
Beyond chatbots: connectors and computer use
Modern AI platforms aren’t chatbots anymore. Claude connects directly to Google Docs, OneDrive, calendars, Salesforce, and other SaaS tools — ask it about a presentation or a calendar and it just answers, no copy-paste required. It reaches local files, notes, terminals, and PowerShell directly. Anything without a built-in connector can get one via MCP, a standard interface any service can expose to any AI. ChatGPT is building the same thing with its own apps ecosystem. Nobody has settled on what to call these platforms — AI productivity platforms, agent platforms, even “AIOS” — but whatever the name, they’re becoming the gateway to work itself: workers go to Claude or ChatGPT first, and only drop into a traditional application when the AI can’t do what they need directly.
Agent bridging and computer-using agents
For applications with no modern connector and only a web interface, this year’s development is agent bridging: the AI can now drive a browser directly, clicking through an ugly old expense system or travel portal the way a human would, all day if needed. For applications that live only on the desktop, computer-using agents from every major AI vendor operate the mouse and keyboard directly. On the OSWorld benchmark — 360 desktop tasks scored yes or no — humans complete about 72%. AI was at 12% in April 2024, 61% by August 2025, and 70% by October: functionally on par with a junior employee who occasionally makes mistakes, same as a person would.
The post-application realization
A live example: Claude can now edit Excel files directly, because .xlsx is just an open XML format and the model doesn’t need to open Excel’s UI to manipulate it — it edits the file the way CRISPR edits genes, not the way a human clicks through menus. The same applies to Word, PowerPoint, PDFs, and code. If an application’s real value was capabilities plus a human-readable interface, and the AI has its own capabilities and doesn’t need the interface, the application itself starts to become optional. A four-stage realization: first, “we need Excel, so AI will operate Excel for us”; then, “why does AI need Excel if it can edit the file directly?”; then, “if I need to see something visually, AI can just draw it”; and finally, the application layer collapses into the data layer — what actually matters is permissions, audit, and DLP wrapped around the data, not the application sitting on top of it.
Applications you create by asking
The same evolution that took ChatGPT from party tricks to “I don’t Google anymore” is coming for every application. Eventually a worker asks their AI to check a spreadsheet every morning and email a summary to a distribution list — and that request is itself a small application, created without procurement, licensing, testing, or a security review. IT can’t manage that world the old way, app by app; it has to manage the environment those creations run inside instead.
Full-on agents and co-workers
An agent, in the simplest useful definition, is an AI environment that runs on its own, makes decisions on its own, works overnight, and routes around roadblocks without a human watching every step. At an internal Citrix QBR, when everyone else introduced their teams, Brian introduced his as ChatGPT and Claude — genuinely how he works with them. The framing that matters: don’t think of AI as an application. Think of it as an extension of a worker.
Bring your own AI
Every worker will bring their own AI and use it their own way, the same way BYOD eventually normalized workers choosing their own laptop within a security baseline. Which AI platform to standardize on can’t be mandated top-down any more than a single corporate laptop model could be — some companies will have real reasons to restrict choice, but most will land on BYO-AI within guardrails, the same guardrails already used to secure human workers: audits, DLP, session recording, applied evenly to AI identities.
What IT does, today and tomorrow
Interaction itself is diversifying beyond typing into voice, video avatars, and cross-device continuity — a conversation started on a phone continues on a desktop, on earbuds during a walk, and on a tablet at home, with the interface built on demand for whatever device is in front of the worker. Underneath, the AI reaches modern apps via APIs and MCP, controls web apps via browser automation, drives legacy apps as a computer-using agent, or skips the application layer and edits files directly. IT’s job doesn’t change in kind: manage workspaces that provide a secure space for work, provide guardrails that protect corporate data, and observe, protect, and guide — today for human workers, tomorrow for AI ones too.
Takeaways
AI platforms and agents are real today, not a future scenario. The path into the enterprise runs bottom-up through knowledge workers, not top-down through mandated tools. No full UI rebuild is required — just a workspace system designed for adaptability. Everything already being done for human workers will apply to AI workers too. The workspace persists as the point of orchestration regardless of form factor. This is happening whether anyone votes for it or not.
Key formulations
“It’s full-on shadow strategy.”
“I don’t even want to call them chatbots.”
“These tools are becoming the gateway to work for human knowledge workers.”
“It’s actually editing the Excel file directly — like CRISPR gene editing.”
“You don’t get a vote. You don’t have to like it, but this is happening.”