AI as an Enterprise Operating System
Tim O'Reilly interviews Dan Guido, CEO of Trail of Bits, about what it actually takes to make a company AI-native. The framing: most enterprises are deploying AI wrong — handing out ChatGPT licenses and waiting for magic. The Solow paradox of AI (90% of executives report no measurable productivity gain) isn't evidence that AI doesn't work; it's evidence that almost nobody has done the necessary organizational work.
Guido describes three levels of AI adoption: AI-assisted (tool), AI-augmented (redesigned workflows), and AI-native (structural — AI as teammate, not tool). The third is what he calls an "enterprise operating system": expertise compounding as code, where every engagement makes the next one faster and better.
The hardest part is employee resistance. When Guido announced Trail of Bits was going all-in on AI, only 5% of the company was with him — 20% actively resisting, 75% passively. He studied the literature and identified four biases to address: self-enhancing bias, identity threat, opacity, and intolerance for imperfection. His countermeasures: an AI maturity matrix (visible levels make it impossible to claim you're already good enough), skills repositories (credit for encoding expertise), curated marketplaces with hardened defaults, and an AI handbook.
Trail of Bits runs a hackathon every two months with defined learning objectives, pairing, and follow-through harvesting of reusable artifacts. The hackathons are training programs that happen to produce useful output, not production sprints. Everything feeds into three skills repositories (internal, public, curated) and a shared configuration repo ("turning scar tissue into infrastructure").
On data: agents force a permissions audit. Every shortcut an organization took with data over twenty years is being called at once. Guido advocates a centralized foundation with innovation on top, measured by what fraction of your team's data is reusable by others.
The recipe: standardize on one agent workflow, write an AI handbook, create a capability ladder, run short adoption sprints, capture everything as reusable artifacts, and make autonomous agents safe with sandboxing and hardened defaults. Guido published the full playbook openly — the skills repository, curated marketplace, configuration repo, and sandboxing tools are all public.