Fifty-two conversations, and two things we had wrong

We spent the autumn in NSF I-Corps asking people questions instead of shipping features. Fifty-two completed interviews, fifty unique people, across buyers, users, decision makers, recommenders — and one saboteur, deliberately, because a person with a reason to say no is the most useful interview you can have.

Here is what changed.

We had the buyer wrong, twice

We started believing the buyer was a security officer: someone who buys air-gap because egress is the risk. That is a real person and they are real customers — but the interview that reframed it came from a business-development lead deploying AI hardware in space. He did not ask us to prove no egress. He said the air-gap gives him no confidence at all. What he needs is a watchdog that reboots in under thirty seconds, because his risk is downtime, not theft.

So air-gap is necessary and insufficient, and our second thesis — "we sell sovereignty" — was half a product.

Then we moved the buyer to AI-native founders, because the pain was loudest there: one founder went from $700 a month to over $3,000 as usage grew, and had built a payment protocol to push AI costs onto his own customers. Real pain. Large. But $9,000 of hardware is trivial against a $36,000 annual bill and impossible for someone who has not measured the bill yet.

Third move, and the one that has produced users: operators with no technical staff. They do not lack interest in running models locally. They lack the procedure. The model is a download; the thing that makes it usable is a pile of software nobody has time to assemble.

The evidence that changed our behaviour

• In pharma, AI tools are approved by a governance council, not one person with a veto. That

lengthens the cycle and changes who we have to convince.

• Corporations are not approving frontier chat tools, so employees buy them personally, out of

pocket, and use them for work. The budget exists. It is on the wrong card.

• Founders already own AI hardware and under-use it, mostly because installing is a project.
• One founder reviewing her own settings found the tool had been **training on her usage by default

on a paid account**. Nobody had told her.

What we shipped because of it

An email address that answers. No install, no login, no hardware, free to claim. More than ten people claimed one and immediately sent real work — manuscripts, Drive links, a video they wanted analysed. That is the cheapest test of the opposite hypothesis, and it passed.

The rule we now hold: lead with the zero-effort surface, and let the hardware be the upgrade the customer asks for once they have felt the cost.

What we still do not know

Whether anyone will pay for the workstation. Not one has sold yet. Whether the operator signs or the governance council does — that single answer changes our sales cost by an order of magnitude. And whether "zero token tax" beats "we already have ChatGPT" when the invoice is not in front of them.

If you are in one of those rooms, we would rather have the conversation than the compliment.