Own the record before you own the model.
AI Under Your Terms, episode 3. Josh Olayemi, Founder, Handvantage. October 2026.
The summaries changed on a Tuesday. Nobody at the firm had changed anything.
The tool looked the same. The login was the same. The answers were shorter, and two of them were wrong in a way the old ones never were. It took a week to learn that the supplier had swapped the model underneath. That is a composite, drawn from more than one firm, with the details changed. The shape of it is common.
The firm did not own that model. It rented it. Renting is fine. Most businesses rent most of their software. The problem was smaller than ownership. When the model changed, no record said who should have approved that, or how the firm would know it had happened.
What YC's "Own Your Intelligence" note says.
On 28 October 2026, Y Combinator published an 89-second video titled "Own Your Intelligence". It is a request for startups, the kind of note YC puts out to say what it wants to fund in its next batch. You can watch it at youtube.com/shorts/DzblM1Mc-J8.
The argument runs like this. Nearly every company rents its intelligence from one of a few large AI labs. Open-weight models, where the trained model is published so anyone can run it, are now good enough to compete. So companies will want to own the models they rely on and improve them on their own data. In YC's words, they will want to "run them where they choose, and decide for themselves when they change." YC says this matters most in regulated industries with strict rules on data.
Then it names the hard part. Someone has to turn each workflow into a test and expert knowledge into training data. Someone has to turn production failures into fixes. All of it has to be carried forward every time a new base model is released.
YC is talking to founders. I read it as a buyer's note. Read the hard part again. Every item on that list is a record.
A test of a workflow is a written statement of what the tool should do, plus proof that it did. Expert knowledge turned into data is a record of how your people decide. A failure turned into a fix is a logged incident with a decision and a name attached. Carrying it forward means someone wrote down what changed and when.
None of that can exist until somebody first writes down where AI touches the data. Which tools. What data each one sees. Who approves a change. What record is kept. The owning that YC describes starts there. So does the renting.
The building and the key register.
Buying a building does not give you a key register. Renting one does not either. A tenant and an owner both need to know who holds keys and who last changed the locks. The owner just has one fewer person to ask.
AI works the same way. If you rent the model, the supplier can change the locks. Your record is what says who should have approved that and how you would find out. If you own the model, you change the locks yourself. Your record is what tells the next person what you did and why.
Three questions to answer before Friday.
Try these on the AI tool your people use most. Ten minutes is enough.
- What data does it see? Client names, bank details, patient notes, contract text, or only your procedure manuals.
- Who approved the last change to it, and how would you know the change had happened? If the supplier swapped the model last month, would anyone have noticed?
- What would you show a client, an auditor or a partner who asked where their data goes? How old is that document?
If any answer is "we do not know", that is the line to fix first. It is also the line the next security questionnaire will land on.
What good looks like.
Good looks like one page. It names each AI tool that touches your data and what that tool sees. It says where each one runs and who hosts it. It names the person who approves a change and how you would know it happened. It says what record you keep, where, and for how long. Someone signs it, dates it, and writes in the next review date.
That page is the first asset. The controls come second: the approval that stops a change, the log that records it. The evidence comes third: a dated copy showing the control held on the day somebody asked.
Handvantage supplies those three. We help regulated firms write the record, put the controls behind it, and keep the evidence that they held. The page itself is free. We publish it as a fillable template with worked examples for a clinic, an accounting practice, a law firm, a carrier, a manufacturer and a retailer, at secvantages.com/AIRecordSheet.
What this is not.
Handvantage does not build models. We do not train them, fine-tune them or host them. Nothing in this piece is advice to buy hardware or to leave your current supplier.
It is not an endorsement from YC, and we have no connection to YC. YC wrote a note about startups it wants to fund. We quoted it because the argument is sound and the gap it names is the one our clients already live in.
It is not a policy. A policy says what should happen. The record says what does happen, with a name and a date on it.
One thing to do this week.
Pick the AI tool your people use most. Fill in the four sections for that one tool. If a line says "we do not know", ask the question that fills it, in writing, and keep the answer.
Whether you rent your intelligence or go on to own it, that page is where the owning starts.
Source: Y Combinator, "Own Your Intelligence", YouTube Short, published 28 October 2026, youtube.com/shorts/DzblM1Mc-J8. Viewed 28 October 2026. The one quoted line is YC's; everything else about the video is paraphrase. Reviewed through 28 October 2026.
