Rent the Brain. Own the Judgment.
Rent the Brain. Own the Judgment.
Sequoia just published the playbook for owning your AI. It was written for companies
with research teams. Here's the version for companies with 30 employees—and the
part that actually applies to you is the part they mention first.
01. THE SENTENCE THAT SHOULD BOTHER YOU
On August 19th, Sequoia Capital published a piece by Sonya Huang called Own Your Intelligence: A How-
To Guide. It came out of an event where Harvey, Mercor, LangChain, Trajectory, and Fireworks walked
through how companies are building their own AI capability instead of renting it from OpenAI or
Anthropic.
One line from it is going to get quoted for the next two years: “Not your weights, not your product.”
The argument underneath it is the one that matters. Huang cites Satya Nadella';s framing: buying
intelligence from a frontier lab means paying twice. Once with money. And once with the proprietary
knowledge you have to hand over to make that intelligence useful to you.
Read that second part again, because it isn't really about AI.
You have been paying twice for a long time. Every time you explain your business to a consultant, a new
hire, a contractor, an agency — you pay in cash, and you pay again by handing over knowledge that
walks out the door when they do. AI just made the transaction fast enough to notice.
Now, before you get excited: you are not going to train a model. Neither am I. That part of the article is
written for companies with machine-learning teams and nine-figure valuations.
But the framework underneath it translates almost perfectly. And the piece is honest enough to tell you
which layer matters most — it's the first one on the list, it's the cheapest, and almost nobody does it.
02. WHAT SEQUOIA ACTUALLY SAID
Four layers, in order. I'm using their words before I translate them.
Evals. A repeatable test that measures whether the system does the work well. Huang quotes Harvey's Gabe Pereyra: "If you don't have a good benchmark, you can't train models." She notes that most evals start as a founder squinting at outputs and vibe-checking whether they feel right and that the goal is turning that judgment into something repeatable. Harvey's Legal Agent Benchmark has over 1,200 tasks across 24 practice areas, graded against more than 75,000 expert-written rubric criteria.
Harness and context. An agent has three parts: the model, the context, and the harness. The harness is everything around the model — routing, retrieval, tools, memory, fallbacks, traces. Harrison Chase's framing: "the main job of the harness is to bring context to the model at the right point in time."
Post-training. Actually changing the model. Fireworks' Lin Qiao's decision tree: missing facts, use context or retrieval. Wrong format, fine-tune. Wrong taste, preference-tune. Bad at a specialized task, use RL. Too slow or expensive, distill it.
Online learning. Trajectory's Arjun Karanam made the sharpest point in the whole piece: models keep getting smarter, but every session feels like their first day on the job. You could put Terence Tao in an accounting firm and on day one he wouldn't be the best accountant there. What's missing isn't intelligence. It's experience.
Hold onto that one. It's the whole article.
03. THE TRANSLATIONHere's the same four layers, sized for a company doing $1M–$25M with 5 to 50 people.
Layer 1 — Evals become standards. You already vibe-check. You look at a proposal, a sales call recording, a service ticket, and you know within ten seconds whether it's right. That judgment lives entirely in your head, which means it doesn't scale and it can't be delegated.
An eval is that judgment written down as a scored test. Ten real examples of the work. A rubric that says what a 5 looks like versus a 2. Anyone — a person or a machine — can be measured against it.
This is the highest-leverage thing on this page and it costs you an afternoon. Sequoia's own sequencing agrees: have the eval before you decide anything else. Because until you can measure the work, every tool decision is a guess dressed up as strategy.
Layer 2 — Harness becomes process. Chase's line about bringing context to the model at the right time is a perfect description of what a good operating system does for a human being. Your new hire doesn't fail because they're stupid. They fail because nobody gave them the right context at the moment they needed it. Same failure mode. Same fix.
Layer 3 — Post-training becomes hiring and training. This is the layer you skip. You will not fine-tune anything, and anyone selling you a "custom AI model" at your revenue level is selling you a wrapper. Use the frontier tools. They're extraordinary and they're cheap.
Layer 4 — Online learning becomes captured experience. This is where the money is, and it's the layer almost every small company throws away.
Sequoia calls it a trajectory: the full path through a task. The context the model saw, the tools it used, the answer it produced, and critically what the user edited, undid, or retried. A failed task becomes a new eval. Missing information gets added to context. A bad tool response becomes a process fix.
Now map that onto your business. Your best salesperson runs a hundred trajectories a month. Your best technician solves problems nobody wrote down. Every correction, every workaround, every "actually, here's how we handle that one" is experience being generated and then immediately deleted. You are running the loop. You're just not capturing it.
04. WHERE I'D PUSH BACK ON SEQUOIA
You should know who's talking before you take the advice.
Sonya Huang's own investment portfolio includes LangChain, Harvey, and Fireworks — three of the five companies that presented at the event. That doesn't make the argument wrong. Their proximity to this work is exactly why it's worth reading. But when a venture firm convenes its own portfolio companies to explain why you should buy what they sell, name it and read accordingly. Huang, to her credit, is upfront that owning your intelligence "is not a mandate nor a recommendation to get off the labs" and that for many workloads the frontier APIs are still the right answer.
Three more things worth saying plainly:
The numbers in the ecosystem are noisier than they look. Mercor's widely reported $2 billion revenue run rate is gross marketplace volume — contractors take home 60–70% of it, per Bloomberg. Real, impressive, and roughly a third the size of the headline. When a stat gets quoted at you, ask what's in the denominator.
“Own your intelligence” is a rich company's problem. The four reasons Sequoia gives for owning — cost at scale, latency, proprietary data, controlling your destiny — three of them require volume you don't have. If your AI bill is $400 a month, inference economics are not your constraint. Frontier model prices have fallen fast and keep falling. Betting against that is a bad bet at your size.
And the failure data is brutal. MIT's Project NANDA study found 95% of enterprise GenAI pilots produced no measurable P&L impact. Read that precisely, because most coverage doesn't: it found that 95% produced nothing that showed up in the P&L within roughly six months. That's a measurement finding as much as a failure finding, and the study is qualitative — 150 interviews, 350 employee surveys, 300 public deployments. But one detail in it should stop you cold: mid-market firms scaled AI in about 90 days. Large enterprises took nine months.
You are not behind the big companies. You're faster than they are. You're just aiming at the wrong layer.
05. WHAT TO ACTUALLY DO THIS QUARTER
Three moves. Do them in order. None of them require an engineer.
Move 1 — Write one eval. Pick the single most repeated piece of judgment work in your company. The proposal. The estimate. The intake call. Collect ten real examples — your three best, four average, three that went wrong. Write a five-point rubric describing what separates them. That document is now the most valuable operating asset you own. It's how you evaluate AI tools, new hires, and your own work, with the same ruler.
Move 2 — Capture trajectories for 30 days. Take the one workflow from Move 1 and record what actually happens: what your best person was given, what they did, what they changed, and where they got stuck. Not a process doc written from memory — an observation log written from reality. Memory writes the process you think you run. This captures the one you actually run, including the corrections, which is where the value is.
Move 3 — Then, and only then, pick tools. With an eval and thirty days of trajectories, tool selection stops being a guess. You test candidates against your rubric and you have an answer. Most owners run this backwards: buy the tool, then try to figure out whether it helped. That's how you end up in the 95%.
Total cost: your attention. Total time: one quarter.
06. THE PART NOBODY AT THE EVENT SAID
Here's the translation Sequoia didn't make, because it wasn't their audience.
For a frontier lab, "not your weights, not your product" means a competitor controls the intelligence your business runs on.
For a founder-led company, the intelligence your business runs on is in your head. Every judgment call, every exception, every "just send it to me and I'll handle it." You own it completely — which sounds like sovereignty and is actually the opposite. Knowledge that exists in exactly one place isn't an asset. It's a single point of failure with a mortgage attached.
Karanam had it right. What's missing is never intelligence. It's experience — and experience only compounds if something catches it.
The frontier labs are building giant brains. Use them. They're the best deal in the history of business inputs.
But the judgment is yours, and right now it isn't written down anywhere. Rent the brain. Own the judgment. Then get the judgment out of your head and into a system, because the day you sell this company, the buyer is not paying for what you know.
They're paying for what runs without you.
Sources: Sonya Huang, "Own Your Intelligence: A How-To Guide," Sequoia Capital, August 19, 2026. MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025." Mercor revenue reporting via Bloomberg, Forbes, and Dealroom, June-July 2026.
