What We're Thinking

A one-person cybersecurity firm just hacked McKinsey's internal AI platform in two hours. CodeWall's AI agent accessed 46.5 million chat messages, 57,000 user accounts, and the system prompts that govern how the platform behaves. McKinsey patched it fast. Says no client data was compromised. But the breach isn't the story. The decision that created the exposure is. McKinsey built Lilli in-house. Strategy planning, data analysis, client presentations. 25,000 AI agents for 40,000 employees. AI consulting is 40% of their revenue. They chose to build instead of buying enterprise AI from well-funded companies that invest a lot of capital on security, red-teaming, and infrastructure. Companies that patch vulnerabilities across thousands of customers at once. When you build your own, you own all of that. The security posture. The patching cadence. The feature roadmap. And you're doing it with a team that has a hundred other priorities. I see this pattern a lot. Engineering teams build internal AI tools because they can. Knowledge management systems, internal search engines, custom chatbots. "Can" isn't "should." The question isn't whether your team is smart enough. They probably are. The question is whether maintaining, securing, and updating it is the best use of their time when tested alternatives exist and ship improvements weekly. There are real cases for building custom AI. Proprietary workflows that don't exist in any product. Domain-specific models trained on data no vendor has. Differentiation that depends on owning the stack. But internal search? Knowledge management? Company chatbots? Those are solved categories. You're competing with companies that have orders of magnitude more capital dedicated to exactly those problems. Start with what's available. Prove the limitation. Then build. The complexity has to be earned. McKinsey's Lilli suggests it wasn't.

AI won't replace your financial adviser. But your financial adviser who uses AI will replace the one who doesn't. Same goes for your accountant. Your lawyer. Your ops team. Your consultant. The disruption isn't human vs. machine. It's the people who use the tools vs. the people who don't.

The promise of a 24/7 AI agent is that it makes you more productive. After a week of living with one, my experience is different: you end up getting sucked into a lot more things than before. His name is Spock (a friend asked why not Data, but Spock is flesh and blood). He runs OpenClaw on a dedicated Mac Mini, connected to my knowledge base, calendar, and task system. Always on. I message him from Telegram, and he responds with context-aware answers drawn from years of my notes. Here's what I've learned: The security architecture is the real work. Two macOS accounts, firewall layering, API key compartmentalization. Spock is like a new employee with limited admin rights. You don't hand over the keys on day one. "Always-on" has real costs. Disk encryption means manual intervention after restarts. Remote access means another attack surface. Token costs: month one was $100, month two doubled to $200. And I'm already sweating weekly session limits three days into the week. It sent a calendar invite to someone without asking. I shared an email about a meeting and asked how to handle scheduling. Instead of suggesting, it created the invite and sent it. From its own calendar. Instructions matter. "Suggest" and "do" should be very different words to an autonomous agent. But it works. Within 12 hours I went from "how would I use this" to "what else can I give it." It manages my tasks, processes meeting notes, curates my reading list. Sometimes I question if the agent is working for me or I'm working for the agent. The technology is ready. But "works" and "production-ready" are different things. Every lesson, security, costs, boundaries, is a governance question that enterprises will face at scale. Better to learn them on a Mac Mini first.

Sarah Friar (OpenAI's CFO) and Vinod Khosla recently discussed the state of AI adoption. They talked about multi-agent systems, the 10-year adoption curve, and the future of work. But when asked for examples of AI actually delivering value? Contract review. Accounting automation. SDR workflows. Sarah described her finance team's workflow: contracts pulled overnight, AI flags non-standard terms, suggests rev-rec treatment. Khosla shared a company doing $150M ARR with ONE accountant. Another replaced 10 SDRs with 1 SDR supervising AI. These aren't moonshots. They're systematic, boring wins. Meanwhile, Khosla dropped this stat: single-digit percentage of users leverage even 30% of AI's capabilities. Sarah's framing: "We've handed them the keys to a Ferrari. They're still learning to back out of the driveway." This isn't a technology gap. It's a capability gap. And Khosla estimates a 10-year journey to close it. That's paradoxical given how fast technology is changing. They say it's a decade-long journey. They also say 2026 is when multi-agent systems mature. I'm not sure both can be true, especially when their best examples are simple automation. But what I do know: change is hard. Getting people to learn new technology while doing their normal work is hard. That's why simple automation is the entry point, to get people hooked, to "refind their jobs" as Sarah puts it. It's where you build the muscle. Getting to agents requires something most organizations haven't developed: the sophistication to handle decisions, context management, and governance. Only 14% of enterprises are using agentic features today. Not because the technology isn't ready. Because they haven't earned the right to use it. They skipped the fundamentals, bought the platform, and now wonder why their agents hallucinate or make decisions no one authorized. Earn your complexity before you buy it. The organizations closing the capability gap aren't the ones with the most sophisticated tools. They're the ones building capability systematically, proving they can govern simple before attempting complex.

After watching The AI Daily Brief's take on the battle for personal context, one pattern stood out: Apple just handed Siri's brain to Google. OpenAI is building hardware with Jony Ive. Hardware companies are ceding AI to specialists. AI companies are pushing into hardware to own the interface. Apple controls your pocket. Google powers the brain. OpenAI wants both. The AI Daily Brief frames this as the battle for context. I think the interface is the bigger prize. The model is the commodity. The interface is the moat.