Quanta Bits - March 21, 2026

Hi everyone,

Happy Spring, happy Nowruz to our Iranian friends, and happy St. Patrick's Day to our Boston neighbors.

This week I want to sit with one sentence that got to me. Sebastian Duesterhoeft at Lightspeed Venture Partners told the Financial Times: "AI is not enterprise software in the traditional sense of going after IT budgets: it captures labour spend." That changes who owns the conversation, what gets measured, and how fast this moves. More on that below.

AI vendors are becoming consultants. Consultants are becoming AI companies. OpenAI is hiring "technical ambassadors" embedded in customer orgs and negotiating PE joint ventures to deploy AI in portfolio companies. McKinsey is running 25,000 AI agents alongside 60,000 humans. The line between vendor and advisor is dissolving from both sides.

Data architecture is the real gating factor for AI. Celonis, which sells process mining, surveyed 1,600 leaders and found 76% of enterprises can't support agentic AI operationally. The AI bottleneck isn't the model. It's everything underneath it.

Agent management is becoming an org design question. IBM has thousands of agents deployed across 300 client projects. The emerging role is someone who manages scope of problems, with agents doing the execution.

AI displacement is hitting the entry point, not the middle. Anthropic's labor market research shows a 14% drop in job-finding rates for young workers entering high-exposure occupations. Companies aren't firing people. They're just not hiring new ones into those roles.

MIT's Ethan Mollick presented data at the launch of MIT's Shaping the Future of Work Center this month: roughly half of workers now use AI daily, and on the tasks where they use it, they report a three-times productivity gain. Yet no company reports a three-times productivity gain. Your people are getting faster. Your numbers don't show it. That gap is not a technology problem. It's an organizational one. People are not sharing the gains with their employers.

Last issue, I shared Anthropic's labor market research and it got reactions. People saw themselves in those graphics. A colleague pushed back with the Jevons Paradox: when something becomes cheaper to produce, demand doesn't shrink. It explodes. Andrej Karpathy, former head of AI at Tesla, made the same point on No Priors: software has always been scarce because it's expensive to build. AI is making it dramatically cheaper. That means more demand, not fewer jobs.

I grew up on Star Trek, so I'm a default optimist on technology expanding what's possible. But what I keep seeing is organizations treating AI adoption as a technology checkbox: get people to use some AI tool, show progress, call it done. That's efficiency theater. Real transformation means rethinking how workflows operate when humans are no longer the constraint. It means redesigning roles, not just automating tasks. And the shape of your organization is going to change.

Block, Jack Dorsey's payments company (Square, Cash App), cut 4,000 jobs in February, roughly 40% of its workforce, explicitly because AI enables "a new way of working with smaller, flatter teams." The same quarter, OpenAI announced plans to nearly double its headcount to 8,000 employees, hiring for traditional roles: sales, engineering, product development. They signed a new lease on over a million square feet of office space in San Francisco. One company sheds thousands of roles. The company building that AI is hiring thousands of humans as fast as it can. Both are making workforce decisions, not technology decisions. And that's the real shift.

Every technology wave eventually affects headcount. What's different now is that AI skips the IT budget entirely. It goes straight to the workforce line. The buyer isn't the CIO anymore. It's the CEO, the CFO, and the CHRO, and they're moving faster than any enterprise software cycle before them.

Sebastian Duesterhoeft, a partner at Lightspeed Venture Partners whose firm wrote a $1 billion check to Anthropic, told the Financial Times: "We took a view that AI is not 'enterprise' software in the traditional sense of going after IT budgets: it captures labour spend."

That distinction matters. Bureau of Labor Statistics data puts US white-collar labor spend at roughly $4 trillion annually. Enterprise software budgets are a fraction of that. When your ERP made accounting faster, the CIO owned the decision. When AI replaces the accounting workflow, the CFO and CHRO own a workforce planning decision. BCG's 2026 AI Radar confirms it: 72% of CEOs now claim primary AI decision-making authority, double last year. This stopped being a CIO conversation.

The evidence isn't theoretical. Alation CEO Satyen Sangani told customers that data stewards are becoming agent managers, analysts are becoming agent builders, operations leaders are becoming workflow architects. McKinsey now runs 25,000 AI agents alongside 60,000 human consultants and is hiring "5Xers," generalists who manage agents across multiple workstreams. NVIDIA CEO Jensen Huang argued at GTC 2026 that engineers are shifting from writing code to writing specifications.

These aren't predictions. These are organizational decisions already being made at companies with tens of thousands of employees. Digital transformation taught us this. Cloud transformation taught us this. Moving the mess to new infrastructure doesn't fix the mess. If you replace headcount with tokens without rethinking the org, you get the same dysfunction running on compute instead of people.

The Jevons Paradox gives me optimism, but only if organizations do the actual work of reshaping. Some already have.

Ramp is a $32 billion fintech with 25 product managers who shipped over 500 features last year and crossed a billion in revenue. Half of Ramp's production code is AI-generated, headed to 80%. But the real story is who writes the code. Product managers, designers, account managers, even salespeople submit production pull requests through an internal tool called Inspect. CPO Geoff Charles put the cultural mantra simply: "My job is to automate my job." Ramp tracks AI token usage per employee and requires AI proficiency in hiring.

That is the kind of organization that is already competing in your market, or will be soon. They didn't get there by running a pilot and reporting savings. They got there by rethinking what every role does. Your company probably can't move that fast. But the question is whether you're moving at all.

If you're a CFO, this is no longer an IT line item you review quarterly. AI spend is competing directly with your workforce budget. You own the cost structure those decisions reshape.

If you're a CHRO, reskilling stopped being a nice-to-have. The Financial Times reported Goldman Sachs is building AI agents with Anthropic that automate accounting, compliance, and client onboarding. Not augment. Automate. If your enablement plan is still "give people access to Copilot and see what happens," you're behind.

If you run operations, start by mapping which workflows are actually AI-ready and which ones have data problems that make AI useless. That inventory is your first real metric.

All of these roles need to drive cultural change. Mollick's data shows 3x gains that never reach a dashboard because the culture punishes visibility. People need to feel safe surfacing productivity without fearing they're automating themselves out of a job.

And if you're an individual contributor: you have an obligation to upskill. The roles are changing whether you prepare or not. Karpathy put it simply: "it's a skill issue." That's not a threat, it's an invitation. The constraint is your willingness to learn, not the technology.

via SuperHuman Newsletter

Okara Launches "AI CMO" for $99/Month. Singapore-based Okara released an autonomous marketing agent that deploys AI agents from a single website URL for SEO, content, and social. Interesting concept, but raises a question buyers should be asking: what permissions and access rights do these agents actually have? (Okara, March 2026)

AI is smoothing out the friction where trust actually forms. Harvard Business Review's Amy Gallo found that half of people surveyed view colleagues who send AI-generated work as less creative and reliable. Harvard's Linda Hill calls productive disagreement "creative abrasion," the mechanism that generates new ideas. When AI polishes that out, it removes the moments where trust and better ideas form. (HBR, March 2026)

MCP is making AI integration easy and AI governance impossible. Model Context Protocol simplifies connecting agents to enterprise tools. The problem: MCP servers are "extremely permissive." Zendesk's SVP of product: "It's the wild, wild West. We don't even have a defined agent-to-agent protocol." The audit trail breaks down the moment you have human → AI → human → AI chains. (VentureBeat, Feb 2026)

A one-person firm hacked McKinsey's AI platform in two hours. CodeWall, a single-employee cybersecurity firm, accessed 46.5 million chat messages, 57,000 user accounts, and 384,000 AI assistants on McKinsey's internal platform Lilli. McKinsey built it in-house for strategy and client work rather than deploying tested enterprise AI. The complexity wasn't earned. (FT, March 2026)

Shadow AI is already inside your sanctioned software. Canva's growth engine is its AI suite. Notion's AI features account for half its revenue. Your IT approved these products. They didn't approve the AI that shipped with the last update. Meanwhile, 22 of the top 50 mobile AI apps were built in China and exported globally. (a16z, March 2026)

▶️ Pro-Worker AI: MIT (Autor, Mollick, Acemoglu). The MIT panel referenced in this issue's essay. Mollick and Acemoglu debate whether AI productivity gains are genuinely pro-worker or just subtask automation.

▶️ Inside Ramp (Geoff Charles, Peter Yang). The interview behind this issue's case study. What happens when every role becomes a builder role.

📖 The Shape of the Thing (Ethan Mollick). The transition from co-intelligence to the agentic era. The window to shape how organizations adopt AI is open but finite.

▶️ Andrej Karpathy on Code Agents (No Priors). Code agents heading from single-shot generation to persistent loops that research, build, and self-correct.

🗞️ Jensen Huang on Accelerated Computing (Ben Thompson, Stratechery). Post-GTC 2026 interview. Engineers shifting from writing code to writing specifications.

Andy Weir, 2021 | 476 pages | ★★★★★

I haven't watched the movie yet, but I thought I'd dig this up from my past notes. Really enjoyed this book, especially the audio version. The narration was entertaining and felt like watching a movie. The story is fairly good. While The Martian was expansive in scope with a lot of landscape to cover, this feels a bit claustrophobic at times, being in a spaceship for the most part. I also liked the imagination behind the alien form represented in the book.

The technicalities sometimes got in the way of the story. And it had a chance to explore broader civilization questions the way the Three-Body Problem series does, but it stays closer to the survival mechanics. Still, the ending was different and overall it was a good entertaining sci-fi read.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖