Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / How Enterprise RFPs Now Demand an AI Data Governance Framework
Transcript
- Lucas: So I was on the phone last week with a head of procurement at a large insurance carrier—company I can't name, but think top five by premium—and she told me something that stopped me cold. Luna: I love cold stops. What did she say? Lucas: She said, 'We now automatically reject any vendor that doesn't provide an AI data governance framework as part of their initial RFP response.' Not after the demo, not during due diligence—day one. Luna: That's a big shift. Even a year ago, most buyers were still asking for a SOC 2 and maybe an AI model card. Now they want the whole data governance playbook upfront. Lucas: Exactly. And it's not just the insurance carrier. I've heard similar from three other enterprise buyers in the last month—one in healthcare, one in financial services, one in energy. They're all converging on this idea that before they even evaluate your product, they need to see how you handle data across the AI lifecycle. Luna: Which raises the question—if these conversations are useful for what you're building or running, we keep this show completely ad-free because we think that's the right way to have a real conversation about enterprise software. If you want to support that, it's buy me a coffee dot com slash fexingo. And we get back to the framework. Lucas: Yeah, so let's get into what a proper AI data governance framework actually looks like from the buyer's perspective. I talked to a couple of analysts and a few vendor-side folks who've been through these RFPs recently. Luna: What's the first thing buyers look for? Lucas: Data classification. They want to know: do you tag every piece of data that flows into your AI system? Is it labeled as PII, financial, confidential, internal use only? And they want to see that classification is automated, not a spreadsheet someone updates once a quarter. Luna: Makes sense. If you can't classify at ingestion, you can't govern anything downstream. Lucas: Right. Second pillar is retention policies. Buyers want a clear, documented schedule for how long you keep training data, inference logs, and user prompts. And they want evidence that you actually enforce deletion—not just policy documents that say 'we delete after 90 days' but audit trails showing the deletion happened. Luna: That's where a lot of vendors slip. I've seen startups with beautiful data governance slide decks that have zero actual enforcement. They present a policy but can't prove it's operational. Lucas: Third pillar is access controls. Who inside your company can touch the data that feeds the model? Buyers want role-based access, least-privilege principles, and—this is the new one—they want to see that your AI engineers don't have unfettered access to production training data. Luna: That's a cultural shift for a lot of AI teams. Engineers used to having full access to everything because 'it's just data.' Now procurement is asking for separation of duties. Lucas: Fourth is lineage tracking. This is the big one. Buyers want to trace any model output back to the specific training data that influenced it. That means versioned datasets, reproducible training runs, and a clear chain of custody. Luna: That's hard. Especially if you're using third-party foundation models or fine-tuning on customer data. Lucas: It is hard. But the buyers I've spoken with say they're willing to accept partial lineage as long as you're transparent about the gaps. The problem is when vendors claim full lineage and can't actually demonstrate it. Luna: Fifth pillar? Lucas: Anonymization and de-identification protocols. Buyers want to know exactly what techniques you use—differential privacy, k-anonymity, synthetic data—and they want to see validation from a third party that those techniques are effective. Luna: So not just 'we use differential privacy' on a slide. They want a report. Lucas: Exactly. And the sixth pillar is third-party subprocessor governance. If your AI model runs on an API from OpenAI or Anthropic or a cloud provider, buyers want to know what data flows to them, what they do with it, and what contractual protections exist. Luna: That's increasingly critical as more enterprise vendors build on top of foundation model APIs. The buyer's risk doesn't stop at your perimeter. Lucas: Right. And here's the interesting thing: one vendor I spoke with—a mid-market HR tech startup—actually turned this into a competitive advantage. They created a public 'data trust page' on their website that documents all six pillars. It's not a PDF you request after signing an NDA. It's just there, for any buyer to see. Luna: How did that affect their sales cycle? Lucas: They told me their average rfp to close time dropped by about 30 percent. Because instead of weeks of back and forth on security questionnaires, the buyer already had the answers. And they started getting invited to demos from companies that previously wouldn't even return their calls. Luna: That's a great example of treating compliance as a product feature rather than a burden. Lucas: Exactly. And the flip side—I also heard about a startup that lost a $4.2 million deal because their data lineage documentation was incomplete. The buyer asked for a trace from a specific model output back to the training data, and the vendor couldn't provide it. So the procurement team just moved on. Luna: Four point two million. That's a painful lesson. And it's not like the vendor was doing anything shady—they just hadn't built the infrastructure to answer that question. Lucas: Which is exactly why buyers are now demanding the framework upfront. They don't want to discover in week eight of due diligence that a vendor can't produce lineage. They want to know before they even schedule the first demo call. Luna: So for any SaaS founder or VP of Sales listening, what's the one thing they should do this week? Lucas: I'd say start by auditing your own data governance against those six pillars. Don't wait for an RFP to force you. Map out your data classification, retention, access controls, lineage, anonymization, and subprocessor governance. Identify the gaps. Then start building a narrative that turns those pillars into a story about trust. Luna: And if you don't have a public data trust page, consider creating one. It signals to buyers that you're not hiding anything. Lucas: Right. Because the vendors who treat this as a checkbox exercise are the ones getting cut from shortlists. The ones who embrace it as a differentiator are closing faster. Luna: I think we're going to see this become standard practice within the next twelve months. Within a couple of years, no serious enterprise buyer will even consider a vendor without a published AI data governance framework. Lucas: I agree. And the vendors who get ahead of it now will have a real advantage. It's not just about avoiding rejection—it's about earning trust before the first handshake. Luna: Great episode. That's all for this one—see you next time. Lucas: Take care.