Impact of AI: Explored podcast, with James O’Regan and Gerjon Kunst · recorded in person at EUC Forum, London · December 17, 2024 · ~45 minutes.
Brian’s second appearance on the show, this time recorded in person, and still an independent consultant at the time. A year-end review and forward look, heavy on AI agents entering the enterprise through Windows desktops, the Claude computer use demo, non-human licensing questions, and the democratization of AI through open-source models.
Why Windows desktops, not new cloud agent infrastructure
Every human job breaks down into roughly 200 small tasks, and the easiest ones to hand to AI first are things like expense reports — find receipts, match them to a calendar, submit. But doing that properly isn’t a matter of wiring an AI into an Expensify API. The AI needs the actual desktop: the photo library with the receipt, the calendar confirming you were there, the credit card statement to reconcile against. That means giving the agent control of the mouse and keyboard on an existing Windows desktop, not standing up a new agent in a cloud container with a bespoke connector into every system it needs. The security, compliance, and connectivity infrastructure a desktop already has is the infrastructure an agent needs — nobody has to reinvent it. People reflexively recoil at “you want AI to use a Windows desktop?” without noticing that Windows has already survived thirty years of things that were supposedly about to kill it: Java, HTML, Chrome, web apps, mobile. Millions of enterprise Windows desktops are still here heading into 2025.
Why the “easy” path through Copilot Studio isn’t easy
In practice, customers try AI agents and hit the same wall: clicking the “AI agent” button in Copilot for Microsoft 365 is easy, opening Copilot Studio is easy, and then you’re building a chatbot, wiring Power Automate connections, standing up APIs and security, and it gets complicated fast. RPA vendors rebranding as AI agent vendors have a reasonable path forward, but the underlying complexity doesn’t disappear.
Claude computer use, honestly assessed
The public demo makes mistakes and looks unpolished, reminiscent of early ChatGPT — and the usual test people run is asking it some obscure PhD-level question it gets wrong, missing that most humans would get the same question wrong too. The right bar isn’t perfection; it’s crawl, walk, run against existing human-desktop infrastructure that already handles every task a worker does. An agent might handle the two easiest of someone’s 200 daily tasks today, three next month, five the month after — and that compounding curve, not any single benchmark, is what actually matters. The same logic that applies to self-driving cars applies here: an agent doesn’t need to beat the best human worker, just the worst one, to create value.
What actually changed in 2024
ChatGPT had been publicly available less than two years at the time of this conversation. The foundational models themselves didn’t take another giant leap during the year, but the surrounding capabilities did: voice on desktop, reasoning, Canvas for interactive editing. It’s the same pattern chip design went through — megahertz, then wider address spaces, then coprocessors, then multiple cores — expanding outward once the core stops being the only lever. There’s no model collapse and no end of scaling happening; growth is just spreading across a lot of different dimensions simultaneously instead of one obvious axis.
Predictions for 2025
The democratization of AI through open-source models is the headline: a roughly $5,000 box, or even a MacBook M2 or M3, can now run models close to frontier capability, fully customizable, no guardrails, entirely under one person’s control — a portable, ungatekeepable version of what used to require a hyperscaler’s infrastructure. Physical AI is arriving faster than expected too — humanoid robots like Tesla Optimus, priced under $50,000, already lifting 50 pounds and folding laundry, with home humanoid assistants plausible within this generation’s lifetime. And workplace shadow AI use among ordinary, non-technical friends and colleagues is already pervasive — further along than most people building enterprise AI strategy seem to realize.
Key formulations
“The path to AI agents is to let AI use our computer.”
“It’s not model collapse, it’s not end of scaling — it’s just going in a million directions and a million dimensions all at the same time.”
“It just has to be better than my worst employee, not my best employee, to provide value.”
“I’ve never been so unsure of the future while being so excited about the future.”