AI companies are going vertical

Three things happened in the same week. Anthropic shipped Claude for Legal — a marketplace with dozens of agents trained on legal roles, from NDA review to policy drafting. They shipped Claude for Small Business, the back office plugin we just mentioned. And OpenAI wired ChatGPT directly into your bank account, with a Finances tab that reads transactions from Schwab, Fidelity, Chase, Robinhood, and 12,000 other institutions through Plaid, with Intuit integration on the roadmap for tax filing and credit-approval modeling.

Our read: AI companies are done building horizontal platforms and waiting for customers to figure out the workflows. They're showing up with the workflows pre-built, packaged by industry, and giving them away. The signal isn't that one company did this. It's that two of the three biggest AI companies did it in the same week, in different verticals, independently. A three-person law firm can skip their legal SaaS vendor and get NDA review from the same tool they use for email. A small business owner can skip their accounting software and say "I'm stressed about making payroll" and get a cash forecast. A ChatGPT Pro subscriber can skip Monarch or Copilot Money and get spending analysis from the same app they use for everything else. The buyer's evaluation just collapsed from "which vertical software" to "does my AI platform have a plugin for this."

Here's who should be nervous. Every vertical software company that thought AI would arrive as a feature they'd bolt on. The companies selling payroll tools, legal workflow software, personal finance apps, SMB accounting platforms. Their pitch has always been "we understand your industry." That pitch just got packaged into a plugin that installs in two clicks.

Meanwhile, one person spent $1.3 million on API tokens in a single month (tracked via CodexBar, an API cost dashboard). That's the ceiling on what power users will pay when the workflows are good enough. The floor is a free plugin that handles your payroll. The SaaS company's three-year roadmap is now another company's plugin description page.

Wade Foster, CEO of Zapier, published something this week that every exec running AI workflows should bookmark. His team built AutomationBench, an open-source benchmark that measures how cheap AI models perform on real production automation — the two-step, explicit-instruction workflows that run at scale inside companies every day.

The gap: The expensive models (think top-tier reasoning engines) cost $1.80 to $6.31 per task. The cheap ones cost $0.003 to $0.02. That's a 100x to 1800x spread depending on the vendor. And on the workflows Zapier tested, the cheap models handled production work just fine.

Why this matters: Most companies are running every AI workflow on whatever model their engineering team picked for the hardest problem. That's like paying senior engineer rates for data entry. The vast majority of production automation doesn't need deep reasoning. It needs reliable execution at a price that doesn't make finance send you a Slack message.

The three-question filter. Before your team routes any workflow to a model, run it through these:

1. Does it need multi-step reasoning? If the task is "extract these three fields from this PDF and put them in a spreadsheet," that's pattern matching, not thinking. Cheap model.

2. What's the cost of a wrong answer? Customer-facing copy, financial reconciliation, compliance tagging — anything where a mistake costs trust or triggers a review. Use the expensive model. Internal routing, data transforms, status updates? Cheap model.

3. Is it high-volume? Anything you run more than 100 times a day is where the gap bites. Ten thousand tasks a day at $1.80 is $18,000. At $0.02 it's $200. That's the difference between "AI is saving us money" and "AI is a new line item that finance is asking about."

Try it: AutomationBench is free and open-source. Run your own workflows through it before your next model contract renewal.

Cerebras went public and popped 70% — The AI chip company raised $5.6 billion at $185/share on Nasdaq, then closed up roughly 70% on day one. The OpenAI partnership and inference-chip story drove demand. If you're tracking where the money is going in AI infrastructure, this is the week's clearest signal. Cerebras

Grok Build CLI is here (and it's $300/month) — xAI's new agentic coding CLI runs on Grok 4.3 with a 2 million token context window and up to 8 concurrent agents. It's gated behind the new SuperGrok Heavy tier at $300/mo, which puts it in a different market than Claude Code or Codex. Early and rough, but the specs are aggressive. xAI

This guy recovered $400K in Bitcoin using Claude — 5 BTC locked for 11 years because the owner got high in college and set the password to something unprintable. Paid recovery services, tried trillions of guesses, nothing worked. Claude found the old wallet file, spotted a bug in the recovery tool, fixed it, and extracted the keys. Worth reading for the story alone. Tweet thread

Notion shipped a CLI built for agents — ntn brings the entire Notion API to your terminal, plus everything you need to build and deploy Workers. Built for humans and coding agents alike, meaning your agent can now read and write to your Notion workspace from the command line. Notion Developers

Jury tossed Musk's lawsuit against OpenAI — A California jury unanimously threw out Elon Musk's case against Sam Altman and OpenAI, ruling he'd filed after the statute of limitations expired. The jury deliberated for roughly two hours after a three-week trial. Musk claimed Altman stole a charity; Altman testified Musk wanted control and suggested it should pass to his children when he died. OpenAI's lawyer called it "nothing but an effort by Mr. Musk to slow down a competitor." BBC

The Pope wrote an encyclical on AI — Pope Leo XIV signed his first encyclical on AI, ethics, and labor on May 15, exactly 135 years after Rerum Novarum (the landmark encyclical on labor and capitalism). The Vatican announced a dedicated AI study group the next day. Whatever your read on the institution, the timing is deliberate and the signal is clear: the biggest questions about AI and work are no longer just for tech conferences. Vatican