Demoed an AI Agent to a Room Full of CFOs — Six Chased Me Out to Complain

Source: a16z | Published: 2026-04-08T14:30:00Z

In a 50-person team, maybe one person can actually document their workflow. Most people simply can't tell an agent what to do.

Most People Can't Draw a Flowchart of Their Own Work

A deceptively simple test: walk into any organization, pick someone at random, and ask them to draw a flowchart of something they do every day. They almost certainly can't. On a 50-person marketing team, maybe one person can fully document the workflow.

This means that when you put an AI co-working tool in front of these people and ask them to tell the agent what to do, their ability to articulate is extremely limited. "Algorithmic thinking" — the ability to decompose work into executable steps — remains incredibly difficult for people in the vast majority of roles.

That Anthropic growth marketer's story went viral on Twitter: one person used Claude Code to essentially automate the work of five to ten people. But the reason he could pull it off is that he's inherently a systems thinker with strong enough technical chops. This isn't a path most people can replicate — at least not yet.

The Spreadsheet Story Will Repeat Itself

An older story might illustrate where this is heading. A top business school graduate joined her firm right at the inflection point of computer adoption — she'd never used a spreadsheet in grad school. The company's solution: let her hire as many interns as she wanted. So her first year on the job was spent managing an entire room of "human agents" — college kids who showed up and built spreadsheets.

Two years later, she and all her colleagues had become "spreadsheet people." The era of needing a room full of helpers to crunch numbers just evaporated. The old workflow was punching out an M&A model on an HP calculator, running two iterations, and calling it done; with spreadsheets, one person could run thirty iterations alone.

Agents are in that transitional phase right now: you think you need 50 agents orchestrated by one exceptionally smart person. But soon, these things will collapse into each other and become a "marketing capability pack." You ask it marketing questions, you tell it to execute. The "rocket science" of coordinating 42 agents will evaporate quickly, and the core competitive advantage will return to domain expertise itself.

Software Must Be Built for Agents, but the Interface Isn't the Point

If agents outnumber humans by a hundred or even a thousand to one, your software has to be built for agents. This is already happening — many companies spend as much time on their agent interface as on their human interface.

But a widely circulated take may be wrong: many people say building for agents means having great APIs, great IDLs, great documentation. Real-world experience suggests the opposite — what agents actually care about is the semantic layer: cost parameters, data durability, platform reliability — not how polished the interface is. Agents are naturally good at finding their way. They'll use collective intelligence to pick genuinely good backend systems rather than being swayed by marketing docs.

The implication for software companies: if your tool is closed off to agents, agents will eventually find a better alternative for the companies they serve. Enterprises used to pick software based on Gartner reports; in the future, they might rely on agent recommendations — and agent recommendations are based on actual performance, not brand. Of course, Silicon Valley will quickly find ways to corrupt this meritocracy, like stuffing ads into agents. As someone in the discussion put it: "Replicate the steak dinner experience for agents."

You Can't Hand an Agent an M&A Data Room

Box just launched an official CLI tool that lets Claude Code operate the entire Box system through natural language. The results are stunning — you say "upload this entire folder from my desktop to Box," and it does it. But the excitement is followed by a string of cold sweats.

Imagine a 5,000-person company where everyone is running agents, all connected to the same shared document repository. One agent moves a file from folder A to folder B, another agent is mid-write on that file, and a third is trying to delete it. And that's just the coordination problem.

Security is thornier. When your agent collaborates with a colleague's agent, it might gain access to resources you were never supposed to reach. You can't simply "treat agents like people" — human colleagues have their own accounts and their own responsibility boundaries; you wouldn't log into their email. But agents are different: they have no right to privacy, you need full supervisory access, you need to be able to step in at any time and say "you messed up, I'm rolling everything back." But if you can log into its account, how does it protect confidential information when collaborating with other people's agents?

An agent is fundamentally an extension of you. There's almost no way around this.

Prompt Injection Is a Real Threat

There's an unsolved technical problem right now: you cannot reliably prevent an agent from leaking information in its context window. If you tell an agent "don't leak information X," enforcing that remains technically very hard.

The corollary: any information that enters the context window can theoretically be extracted via prompt injection attacks. If someone knows your agent's email address, they can send it a social engineering email — ten times easier than social-engineering a human. So for now, letting agents independently access sensitive resources like M&A data rooms simply can't be done securely.

Many people's stopgap solution: give the agent its own phone number, its own Gmail account, even its own credit card. Use Gmail's built-in permission system for governance. This works for personal productivity, but it's nowhere near sufficient for enterprise use cases.

Enterprises Will Lock Down, Startups Will Sprint

This gap in security and governance will create an enormous speed divide. Startups start from zero with nothing to blow up, so they'll embrace agents at full speed. Large enterprises will freeze out of fear — until some kind of consensus emerges.

This mirrors the early diffusion of open source. The debate was identical: you can't let some engineer just copy open-source code into a commercial product — there are licensing issues, quality issues, security issues. Eventually an entire set of norms developed. The difference is that the open-source debate happened in conference rooms where nobody was watching; the AI debate is playing out on everyone's Twitter timeline, with everyone trying to skip straight to the endgame.

But you can't skip to the endgame. Someone has to build first, step on mines first, and produce a real case study first.

SAP Isn't Going Anywhere

A seductive narrative keeps resurfacing: agents can write code and use APIs, so all existing enterprise software will be replaced. But "vibe-coding your way to SAP" is absurd.

The domain knowledge in SAP doesn't just live in some elegantly designed data layer. It's distributed across the UI, the middleware, the usage patterns. For twenty-five years, humans have been the bottleneck for unlocking these software capabilities — how many people can't even make a dual-axis chart in Excel? Agents are great at navigating complex software interfaces, helping people find features and generate reports. But the underlying systems of record aren't going away.

The idea of rebuilding ERP from first principles is always tempting. SAP was founded in the 1970s with one set of assumptions; today you'd design with completely different ones — but ten years from now you'd think those decisions were wrong too. Layers don't disappear; they get stacked. Organizational boundaries, compliance requirements, security policies — these are what dictate the existence of layers, not technical efficiency.

Agents Will Make Enterprise IT Even More Fragmented

An underestimated risk: agents will build de facto new systems of record on their own. Just as every company's shared drive used to accumulate the best documents and the best financial models — people who couldn't find what they wanted just created something new — agents will amplify this problem. They'll frantically spin up new integrations and data flows in the gray zone that IT considers "middleware."

CIOs have seen earlier episodes of this movie: let the marketing department buy their own website builder and set up campaign pages, and next thing you know the email list leaks and the whole company gets sued. So their fear of "letting agents freely build integrations" is very concrete — it's not that agents themselves are unreliable, it's that once you allow people and agents to freely create connections between systems, your systems of record can break.

Someone in the discussion suggested that for a considerable period, agents in the enterprise will likely be "read-only": consuming data, generating reports, but not allowed to write to or modify core systems.

Engineering Token Budgets Will Be the Hottest Topic of the Next Two Years

A very practical question is haunting every CTO and CFO: what should the AI compute budget for the engineering team be? Answers on Twitter range from 1% to 100%.

Public tech companies spend between 14% and 30% of revenue on R&D. Whether AI compute costs are double the engineering team's headcount cost or just 3% more — that gap directly impacts earnings per share. The moment Claude Code Max launched, people were hitting rate limits after three prompts. This isn't a hypothetical discussion.

Startups will handle it by burning through available capital and pretending it's not a problem. Big companies will freeze out of fear. Somewhere in between, companies willing to make the bet will use their financial muscle to go all-in on a specific niche and become the leaders.

Wall Street's Zero-Sum Thinking Is Off by an Order of Magnitude

Everyone is trying to figure out the economics of AI, but they're underestimating the size of the opportunity by at least an order of magnitude.

This is identical to the PC era. People viewed the PC market as finite because they treated MIPS consumption as a fixed number. Nobody imagined what would happen when you put MIPS on every desk. Nobody imagined software and hardware would be sold separately — until one person did.

Cloud computing was the same. People looked at the server business and said: "We're just moving our 60,000 servers a year into someone else's data center." Nobody foresaw that people would consume a thousand times more compute. Salesforce faced the same dynamic: the CRM market was $2 billion a year — $2 billion worth of servers, Oracle licenses, and multi-year deployment consulting. If you let salespeople sign up with zero friction, every single one of them would.

One investor said he'd been investing for ten years, with roughly 240 infrastructure companies in his portfolio, and over the past six months every single one's growth curve had gone asymptotic — because more software is being written now than at any point in history. And on-device AI consumption on phones hasn't even started yet.

The Vacuum Tube Lesson

Every technological leap in history has produced the same panic. During World War II, people calculated how many vacuum tubes would be needed to meet computing demand — the conclusion was that all of South Dakota would need to be covered in vacuum tube warehouses, with people on roller skates racing down the aisles to swap tubes. Then someone invented the transistor.

When IBM sold mainframes priced per MIPS, they shipped more MIPS every year while charging less, without even realizing it — until someone pointed out they were on a declining curve, manufacturing MIPS faster than they could monetize them.

"The exact same thing will happen with tokens. That's guaranteed."

The current token anxiety is real, but it's a phase problem, not a permanent constraint. The solution might be more supply, algorithmic breakthroughs, or hardware revolutions. What's certain is that it won't stay stuck at "everyone clutching a token card, queuing up in the cafeteria to swipe in" forever.

New Business Models for the Agent Era

Vast amounts of information and software resources are undervalued by a factor of a hundred, simply because no one is willing to pay five cents to access a data point or one dollar to use a tool. But if you give an agent a budget and a protocol, it can spend three dollars to pull medical literature during a deep research task and complete the transaction automatically. This opens up an entirely new world of micropayments.

Of course, the promise of micropayments has appeared in every technology generation, and enterprises always gravitate toward bulk licensing — predictable, no thinking required. But tokens as a significant portion of COGS are pushing the entire industry toward usage-based billing. The shift from perpetual licenses to subscriptions took a generation; the shift from subscriptions to usage-based billing is happening right now.