Public relations

Why AI privacy is a must-have for PR agencies under NDA

Sam Polstein works under NDA with defense tech clients. Five conditions have to hold before an agency can put that material into any model.

Enterprise privacy is a precondition for AI in public relations, not a feature. Sam Polstein, who runs communications and content for deep tech startups at HAUS, signs NDAs with clients that include defense tech companies, and says assurance that their information is not training models is "a must have, not a nice to have for the work we do." Five conditions have to hold before an agency can put that material into a model.

Why is this a precondition rather than a preference?

Because the material is contractually and, in some cases, nationally sensitive. Polstein's description of the constraint, given to Artificial Reality, the AI buyer-evidence podcast hosted by Galina Fendikevich:

"We sign NDAs with our clients and we've got information that we don't want to be used to train models and don't want to be in the public domain, because it's incredibly important. We work with defense tech companies — it's important from a national security standpoint."

The question he cannot currently answer for a client is the one that decides whether the work happens at all:

"How do we give them that peace of mind that their information is not being used to train the models or is publicly available?"

What would an agency have to be able to tell a client?

Five things, drawn from what Polstein describes needing:

Verifiable guarantees that client data does not train models
Assurance clear enough for an agency to pass on to a client under NDA
Suitability for defense and national-security-adjacent work
No self-hosting requirement
Communicated in a way a non-technical client can accept

The fifth item is the one vendors underrate. The person who has to be convinced is not the agency's technical lead. It is a founder reading a paragraph in a contract.

Why is self-hosting not the answer?

Because it moves the burden onto a business with no appetite for it. Polstein is direct about this:

"Especially if you're a PR agency — we don't want to deal with [self-hosting]. That's overhead you don't want to have, and shouldn't have to have."

An agency is not going to run infrastructure to earn the right to use a model. Any answer that begins with a deployment topology has already lost this buyer.

Which vendor is closest, in his assessment?

He names one, with a hedge intact. "I think Claude has pretty good safeguards and controls in that regard, but I think they need to do some more work on building that trust." That is his read of the current state, not a benchmark and not an endorsement, and the second clause is doing as much work as the first.

Is this an engineering problem or a communications problem?

His diagnosis is that it is largely the second, and he flags his own bias while making it:

"Maybe I say this because I'm a comms and marketing person, but I think part of that is comms and marketing. They have some work to do, the big AI companies, to make sure that that trust is built up with larger enterprises."

Worth taking seriously, given the source. This is a person whose job is establishing that a technical claim is credible to a skeptical outside party, saying the technical claim is probably fine and the credibility is missing.

Why does this buyer have no switching cost?

Because he engineered it away on purpose. Polstein keeps his context as markdown files, read through Obsidian and backed up in Google Drive, and points whichever model he is using at them. His stated reason: "I want to make sure that the context is mine, it's not Claude's."

His position on vendor loyalty follows from that: "I'm here for the highest bidder. Whoever has the best product and the best compute and is gonna get me the tools and output that I need, I'm gonna switch over to them."

Artificial Reality's analysis: this combination is the uncomfortable one. A buyer with a hard trust requirement and no accumulated lock-in is a buyer who can leave the moment a competitor clears the bar he cares about. Privacy assurance is not a compliance checkbox in this segment. It is the retention mechanism, because nothing else is holding him.

What should an AI company do about it?

Write the paragraph an agency can forward. Not a trust center, not a whitepaper, not a settings toggle — one plain-language statement about training use and data retention that a principal can paste into an email to a client under NDA, and that survives being read by that client's counsel.

Then test it against the hardest case in this account: a defense tech company whose founder wants to know, in one sentence, what happens to the deck. If your answer requires the reader to understand your architecture, it is not an answer. Who at your company owns that sentence?