Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / Why Enterprise RFPs Now Ask About Vendor AI Training Data Sourcing
Transcript
- Lucas: So Luna, I was reading through a software RFP last week, and one section stopped me cold. It wasn't about uptime, security, or even pricing. It was a full page asking the vendor to disclose exactly where the AI models in their product were trained on. Luna: Wait, like the actual datasets? That's a pretty big ask. Lucas: Yeah, and it's not a one-off. I'm seeing this clause more and more in enterprise RFPs, especially in regulated industries. They want to know the sources, the licenses, even the demographic breakdown of the training data. If you can't answer, you might not get a seat at the table. Luna: That's a major shift. A few years ago, nobody asked about that. So what's driving it? Lucas: Partly it's legal exposure, and partly it's reputational. If your model was trained on scraped data that's later found to be copyrighted, the enterprise buyer could be dragged into a lawsuit. That's a risk procurement teams are no longer willing to take silently. Luna: And it's not just legal risk, right? There's also the bias angle. If the training data skews a certain way, the model's output could be problematic for a customer-facing application. Lucas: Exactly. One bad output in a healthcare or finance context, and the vendor's product is dead on arrival. So the RFP is becoming a due diligence tool, not just a checklist. Luna: So what does a typical question look like in this section? Give me a concrete example. Lucas: Sure. One RFP I saw asked, 'List all datasets used to train your AI models, including their source, license type, and any restrictions on use.' And then follow-up questions about whether you've verified the license covers commercial use, and whether you'd provide audit rights to verify that. Luna: Audit rights on training data. That's deep. Most vendors probably can't even answer that fully. Lucas: Right. And that's the new reality. Some vendors are responding by publishing data governance statements, but others are still scrambling to map their own data pipelines. It's a real gap. Luna: And that gap is where the smart vendors are seeing an opportunity. You know, if they can be transparent, it's a differentiator. Lucas: Absolutely. And that's exactly the kind of insight that keeps me recording these episodes. Speaking of which, if you're getting value from these conversations, a couple of dollars a month genuinely keeps them going. You can find us at buy me a coffee dot com slash fexingo. It's a small thing that makes a real difference. Luna: Yeah, and we really do appreciate it. Now, back to the RFP thing — what about the vendors who can't answer? Are they just out of the running? Lucas: Not necessarily out, but they're at a disadvantage. I've seen procurement teams score the data provenance section heavily, and if you leave it blank, you're starting from the back of the pack. Luna: So what's a realistic response for a vendor that uses third-party models or pretrained data? Lucas: That's a good question. Some vendors are now providing a data sheet that lists model sources, whether it's an open-source base like Llama or a commercial API, and then they explain their fine-tuning data. But if you can't even name the base model, that's a red flag. Luna: And I imagine this is going to push more vendors to build their own proprietary datasets, or at least license them properly. Lucas: It already is. I've talked to a few startup founders who are now treating data licensing as a core part of their product development, not an afterthought. It's a shift in mindset. Luna: Let's talk about the legal side a bit more. What's the worst-case scenario for an enterprise buyer if they ignore this? Lucas: Imagine a bank deploys an AI co-pilot that was trained on a dataset later found to contain copyrighted text. The copyright holders sue, and the bank is named in the suit because they're the ones using it commercially. That's a reputational and financial nightmare. Luna: And that's why they're putting the burden on the vendor now, rather than waiting for a problem. Lucas: Exactly. Procurement is becoming a gatekeeper for AI risk. And it's not just about the training data itself, but also about how the vendor handles updates to that data over time. Luna: So the RFP is starting to ask about versioning and change management for training data too? Lucas: Yes. They want to know how you track when a model is retrained, what new data went in, and whether that could change the model's behavior in a way that affects their business. It's almost like asking for a nutrition label on AI. Luna: A nutrition label, I like that. So where does this trend go from here? Do you see it becoming standard in every enterprise RFP? Lucas: I think within the next two years, yes. It's already common in sectors like healthcare and finance, and now I'm seeing it in more general B2B software. The interesting part is that it's not just a compliance checkbox — it's becoming a competitive filter. Luna: Meaning the vendors who embrace transparency will actually win more deals. Lucas: Right. Because the buyer's own due diligence effort is reduced, and they feel they're dealing with a partner who has nothing to hide. That's a powerful signal in a long sales cycle. Luna: But what about open-source models? Those are trained on the internet, which is a mess. How does a vendor answer those questions? Lucas: That's the tricky part. If you use a model like Llama, you can point to its public documentation, but you still need to explain what you've done with it, like any fine-tuning data you added. And if your fine-tuning data includes customer data, you need to show you have the right to use it. Luna: That's a good point. So the answer isn't just about the base model, but the entire data lifecycle. Lucas: Exactly. And that's why we're seeing new roles like data governance officer and AI compliance manager popping up at software companies. It's no longer just a technical problem. Luna: Have you seen any vendors actually failing these RFPs because of this section? Lucas: I have. One vendor I know lost a deal because they couldn't confirm that their training data was free of personal information from the European Union. The buyer didn't want to deal with GDPR implications. Luna: Ouch. So this is real. It's not just about being vague; you have to have actual processes in place. Lucas: Yeah, and that's the shift I want to highlight. Procurement is now asking for evidence, not just assurances. And for vendors who are early, this is a chance to stand out. Luna: So what's your advice to a SaaS founder listening right now? What should they do this week? Lucas: Start by mapping your training data. Write down every dataset you use, where it came from, and what the license allows. If you can't do that for one model, that's your first red flag. And then create a simple data sheet you can share with prospective buyers. Luna: Even if it's not perfect, just having that document puts you ahead of most competitors. Lucas: Exactly. And as this becomes standard, it's going to be interesting to see if we also get industry-wide standards for data provenance, maybe something like a certification. Luna: I wouldn't be surprised. It already feels like the beginning of a new kind of due diligence, one that's all about trust in the data, not just the code. Lucas: And that's a healthy direction for the industry, even if it makes life harder for vendors. The more transparent we are about how AI is built, the more confident enterprises will be in actually using it. That's good for everyone in the long run. Luna: Well said. And if you want to share your own data provenance story, we're all ears.