There's a Harvard Business School case study from the early '90s that asks a simple question: computers had been in offices for years, so why hadn't productivity numbers moved? The answer, the case argued, was that companies had stuck a computer in the corner and left every other process untouched. The filing cabinets were still there. The org chart was still there. The workflows were still there. The computer was a guest, not the center.
Boris, the creator of Claude Code at Anthropic, brought up that case study in a recent CNBC interview, and it's the most useful frame I've heard for what's happening with AI right now.
The companies that are getting hundreds-of-percent productivity gains from AI are not the ones that gave their teams a chat window. They're the ones that put Claude — or whatever model — at the center of how the business actually runs, and rebuilt everything else around it. Anthropic does this internally. So do their most sophisticated customers. Boris was very specific: not "a few percentage points," which is what he used to see in his old developer-productivity job at Meta. Hundreds of percentage points. And accelerating.
Most companies are not there yet. Most companies have done the corporate equivalent of keeping the filing cabinets and putting a computer in the corner. The filing cabinets in this case are the meetings, the manual handoffs, the ticket queues, the multi-tab knowledge work, the Tuesday-afternoon status calls. Until those go away, the gains stay theoretical.
What does "AI at the center" actually look like? Boris described his own day: he no longer hand-writes code. Hasn't for about six months. He prompts Claude. Claude writes, tests, shows him the result, he says yes or asks for a change. At any moment he has "a few agents or sometimes thousands of agents" doing work for him in parallel. He uses the new co-work product to book his own travel — he didn't sit on Expedia, he asked the agent. He's not editing a Word document. He's directing a fleet.
This is the bit business leaders keep missing. The unit of work has changed. The right question isn't "how do I get my team to use ChatGPT more?" The right question is "what does my company look like when a single person can run a thousand agents?" That's a totally different org chart. A totally different cost structure. A totally different definition of who's senior and what seniority means.
The proof points are not subtle. NASA used Claude Code to plot the Mars rover's course. Shopify and many of the largest enterprises are restructuring delivery around Claude. On the long tail, there was someone on Twitter using Claude Code to grow tomato plants — monitoring the webcam feed, tracking nutrients, adjusting the routine. Boris noticed that pattern — people using a coding tool for non-coding work — and it's exactly what triggered the co-work product. Real builders watch how their tool is being misused, and they ship for that. That's a separate lesson worth bookmarking.
A few of his other claims are worth holding onto:
Switching costs as a moat are evaporating. Claude can rewrite the integrations. The lock-in that was holding mediocre SaaS tools in place is thinner now than it was 18 months ago. Some moats survive — distribution, network effects, brand — but "we're hard to leave because the migration is painful" is no longer one of them.
Security shifted phase. Three months ago, models couldn't reliably find vulnerabilities. Now they're exceptional at it. That cuts both ways — defenders find their own bugs first, attackers find them faster too. Anthropic's stated bet is that the good guys get the best model first and patch ahead of the bad guys. That bet only works if defenders are actually using these tools. Most security teams haven't internalized the shift.
Coding is becoming a basic literacy. This is the line that will get the most pushback and is most likely to be right. In the 1400s most people couldn't read. Now most can, and an entire civilization runs on that base. Boris thinks talking to an agent and getting working software back is the next layer of universal literacy. Professional writers still exist; professional engineers still will too. But the ground floor is moving up.
Now is genuinely the best time to start a company. Not the worst — the best. He'd "not be surprised" if there are 10x or 100x more startups in ten years. The cost of building has collapsed. The cost of distribution is the next thing to fall. If you've been waiting, you are running out of reasons.
The takeaway for any operator reading this: stop running pilots. Pilots are how you keep the filing cabinets. Pick one workflow that's central to how your business makes money, put an agent at the center of it, and rebuild every adjacent process around what the agent can now do. Then do the next workflow. Don't ask "where can AI help?" Ask "what does this team look like if the agent is the default operator and the human is the editor?"
The companies that ask the second question are the ones already pulling ahead. Everyone else is still waiting for the productivity miracle to show up while leaving the filing cabinets in place.
Source: Head of Claude Code on the future of work and productivity
No comments:
Post a Comment