Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / Why Enterprise RFPs Now Demand AI Training Data Provenance
Transcript
- Lucas: If you're selling any kind of ai powered SaaS into an enterprise today, there's a new question coming up in almost every RFP. And it started with a lawsuit that made headlines last year: The New York Times suing OpenAI over the use of its articles as training data without permission. Luna: Right, that case really put a spotlight on where AI models get their data. I remember the Times alleging that OpenAI used thousands of copyrighted works. Lucas: Exactly. And while that case is still working its way through the courts, the ripple effect in enterprise procurement has been immediate. Legal and compliance teams at Fortune 500 companies saw that and thought: if the Times can sue over training data, we need to know what our vendors are feeding their models. Luna: So what are they actually asking for? Is it just a checkbox or something more detailed? Lucas: It's becoming its own section in the RFP. Buyers are now demanding a 'training data provenance report' — essentially a document that lists every dataset used to train the model, where that data came from, whether it includes copyrighted content, and whether proper licenses were secured. Some enterprises are even asking for a list of all copyrighted works used and the legal basis for using them. Luna: That sounds like a huge lift for a startup selling an ai powered tool. How are vendors supposed to produce that? Lucas: It depends on how they built the model. If they used open-source datasets like Common Crawl, they can point to that — but they also need to show they've filtered out copyrighted content. If they used proprietary data, they need to prove they own it or have licenses. The big challenge is for vendors who fine-tuned on a base model without tracking the original training data lineage. Luna: So if I'm a sales rep going into a deal and my engineering team can't produce this, I'm probably dead in the water? Lucas: It's a real blocker. Procurement teams are starting to include a 'right to audit' clause around training data. And some are even requiring vendors to indemnify them against copyright claims arising from the model's output. That's a big shift from just caring about security certifications. Lucas: And you know, the interesting thing is that this isn't just about legal risk. It's also about trust. Buyers want to know that the AI they're buying wasn't built on data that could come back to bite them. Luna: Absolutely. And speaking of trust, I think it's worth noting that this whole conversation is made possible by a small group of listeners who support the show. If you find these episodes useful for what you're building or running, you can chip in at buy me a coffee dot com slash fexingo. It's what keeps us ad-free and focused on topics like this. Lucas: Yeah, it really does. And we're grateful for that support — it lets us dive deep into evolving buyer demands without any sponsors pulling us in different directions. So, back to data provenance: what does this mean for the sales cycle? Lucas: For one, the sales cycle is getting longer in the early stages. Before you even get to a demo, you might need to submit a data provenance questionnaire. Some enterprises have standard forms they send to AI vendors — sometimes thirty or forty questions about training data. Luna: Do they really have the expertise to evaluate those answers? I mean, a procurement manager might not know the difference between a transformer and a diffusion model. Lucas: They're getting help. We're seeing law firms and consultancies build AI due diligence playbooks. And the enterprises with mature AI governance teams are hiring data scientists specifically to vet vendor training data. It's not just checking boxes anymore. Luna: So for vendors, what's the best move? Start building a data provenance document even if no customer has asked yet? Lucas: Exactly. If you're waiting for the first RFP to arrive, you're behind. Put together a clear, concise document that explains your data sources, your filtering process, and any steps you've taken to avoid copyrighted content. If you've done third-party audits, include those. It should be a living document that you can share early in the conversation. Luna: And what about open-source models? If a vendor is using Llama or Mistral, does that simplify things? Lucas: It helps, but it doesn't eliminate the question. Even open models like Llama have training data that includes web scrapes. The vendor still needs to show they understand the provenance and have mitigations in place. Some buyers are actually more suspicious of fine-tuned models because the custom training data is opaque. Luna: There's also the European AI Act coming into force. That requires transparency around training data for high-risk AI systems. So enterprises with EU operations have an extra incentive. Lucas: Right, and that's going to accelerate the demand. We're already seeing RFPs that reference the EU AI Act explicitly. It's becoming table stakes for any serious AI vendor. So if your company sells an AI product into regulated industries — healthcare, finance, legal — this is not optional. Luna: Let's talk about a concrete example. A mid-market SaaS company I know sells an AI contract analysis tool. They started getting asked about training data last quarter. They had to pull together a report showing they only used publicly available legal documents and contracts they had permission to use. It took their legal team two weeks. Lucas: And did it win them the deal? Luna: It did, but only because they were proactive. The RFP came from a large law firm that had already eliminated two vendors who couldn't answer the data provenance questions. So being prepared is becoming a competitive differentiator. Lucas: That's the key takeaway: enterprise buyers are starting to see training data provenance as a non-negotiable risk factor. If you're a vendor, start building your data lineage documentation now. If you're a buyer, add it to your standard evaluation criteria. Luna: And if you're a listener selling enterprise software, what's the one thing you should do tomorrow? Lucas: Talk to your engineering team. Ask them: can we list every dataset used to train our model? If the answer is no, that's your first priority. Even if you don't have a customer asking today, the question is coming. And it's better to have a clear answer than a blank stare in a boardroom. Luna: Great advice. Thanks, Lucas. Lucas: Thanks, Luna. And for everyone listening, if you want more on this topic, we'll have links to sample RFP questions in the show notes. Until next time.