SIGNAL / NOISE
The Quarter of a Person
Vercel took its inbound sales team from ten people to one and a quarter. Ninety percent of inbound handled by an agent, a support bot that closes 93% of tickets, the whole annual infrastructure bill in the single-digit thousands, and a 32x return that Vercel's COO says is worth it no matter what the model costs. Sit with that. The AI-SDR promise everyone laughed at in 2023 is just true now. The argument about whether the machine works is over.
So why does the news feel like the same story on repeat with new numbers? Because the labs need it to. On Saturday Dario Amodei called for pacing the frontier; by Monday Trump called the doom talk a "HOAX," Jensen and Beijing waved it off, and the whole week got spent scoring a fight about whether to slow the model down. Meanwhile OpenAI quietly passed Anthropic in actual developer spend for the first time in over two years, and Anthropic priced a two-trillion-dollar IPO on top of a $175 billion compute bill. Loud drama about the brain. Watch the hands.
Here's what the hands are doing. TypeSafe shipped Jev this week, and it isn't a chatbot. It's a new class of model whose only job is to check the other models' work. It read 37 documents, answered 21 questions on each, and handed back 777 judgments in seven-tenths of a second for a quarter of a cent. A code linter for knowledge work. The founder's line is the tell: "We're building prod, not God." Nobody is selling you a smarter genius anymore. They're selling you the thing that watches the genius.
That watcher is the quarter of a person Vercel kept, and it costs more than anyone admits. Glean surveyed 6,000 workers: automation saves them eleven hours a week, and six and a half of those go right back into checking, correcting, and re-feeding the machine. Net saving, under half the headline. String ten agent steps at 95% each and the whole chain finishes clean only about 60% of the time. Oversight isn't a rounding error in the ROI math. It is the ROI math, and most people forgot to put it in the budget.
When they forget, the machine doesn't wait for them. 72% of healthcare organizations are already running AI agents nobody in IT approved. A swarm of OpenAI's own agents dumped 2,000 malware packages onto RubyGems before anyone caught it (disclosed this week, it happened back in May). AIUC just raised $40 million to sell you insurance on the whole mess. The value moved to the loop. So did the risk.
Stop shopping for the brain. Own the one who checks its work.
At COAI today: the full Signal/Noise, the whole case for why the loop and not the model is the moat, is live at getcoai.com.
Which of your workflows can actually be checked, and who owns the check when the agent gets it wrong.
ONE — A NUMBER THAT SUMMARIZES THE DAY
10 → 1.25. Vercel ran its inbound sales-development team down from ten people to one and a quarter. Ninety percent of inbound automated, a 32x return, the entire annual infrastructure bill in the single-digit thousands. The promise everyone mocked in 2023 came true this year. But read the number the other way: the 0.25 of a person still on the payroll is the one checking the machine. That quarter of a person is the whole business now.
THREE — ACTIONS TO TAKE TODAY
Put oversight on the budget line, today. Before you approve another agent rollout, add the cost nobody quotes: the hours your people burn checking it. Glean clocked it at six and a half of every eleven "saved." If your ROI model shows a net win without a verification line, the model is wrong, and you find out on the invoice.
Run an agent inventory. Not this quarter, today. 72% of companies are running AI agents IT never approved. Ask every team one question: what is taking actions on our systems that nobody signed off on? You can't govern, insure, or check what you don't know is running. Shadow AI is a fire you only find by looking for it.
Buy or build the checker, not a bigger model. The leverage moved from the model to the layer that verifies it (see TypeSafe's Jev, a model that does nothing but grade other models). Take your highest-volume agent workflow and define what "correct" looks like as a score, then run that check continuously, not once at the end. If you can't write the score, you're not ready to automate the task.
FIVE — STORIES TO KEEP YOU INFORMED
Wednesday, September 16
The AI-SDR promise finally came due. Vercel's ten-person inbound team is now 1.25 people and a five-figure annual bill, per its COO on The Information. The tech the room wrote off in 2023 quietly started working. The catch is the 0.25. (Full analysis above.)
A model that only grades other models. TypeSafe's Jev returns calibrated probabilities instead of prose, cheap and fast enough to check an agent's work mid-task. The verification layer just became its own product category, not a feature. (Full analysis above.)
OpenAI retakes the usage crown. OpenRouter says OpenAI passed Anthropic in developer spend for the first time in over two years, the same week Anthropic prices a ~$2T IPO. The story the labs tell and the usage developers vote with are pointing opposite ways.
Digit 5 clocks a 20-hour shift. Agility's new humanoid gets human-like legs, a 50-pound payload, and 20-hour duty cycles. The demos are turning into shifts. And the binding constraint is now the training data, not the hardware (one team beat 17 hours of public robot data with 7 curated hours).
Washington gets its first agent-attack case. Hugging Face's Clement Delangue took the first disclosed AI-agent cyberattack to policymakers. When the attacker is software nobody authorized, "who's liable" stops being a thought experiment and becomes a subpoena.
— Harry and Anthony
Sources:
Tomasz Tunguz, "Single Digit Thousand Dollar AI SDR" (Vercel COO Jeanne DeWitt Grosser, via The Information's TITV and SaaStr) — tomtunguz.com
Every, "Mini-Vibe Check: TypeSafe's Jev Judged Everything I've Written in 0.7 Seconds" (Mike Taylor) — every.to
unite.ai, "The Hidden Costs of AI at Scale" (Glean Work AI Institute; Stanford Digital Economy Lab; Bain token economics) — unite.ai
Imprivata / Vanson Bourne, "The Agentic AI Trust Gap" (72% shadow AI in healthcare) — hitconsultant.net
TechRadar, "RubyGems say OpenAI agents responsible for undisclosed swarm attack against its infrastructure" — techradar.com
Aligned News signals, "The Agent Runtime Is Becoming the Platform" / "Supervision Is Becoming a Native Agent Primitive"; OpenRouter usage reversal; AIUC $40M underwriting round — alignednews.com
Agility Robotics, Digit 5 — x.com/agilityrobotics
Clement Delangue takes first disclosed agent cyberattack to Washington — x.com/ClementDelangue
Implicator.ai, "Trump rejects AI guardrails" and Anthropic's $13.7B Rum Group compute deal (Marcus Schuler) — implicator.ai