Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / Why Enterprise Buyers Now Demand a Vendor AI Model Card for Every Deal
Transcript
- Lucas: Every enterprise procurement team now has a new checkbox, and if you can't tick it, you don't get a signature. I'm talking about the model card — not the AI model itself, but the one-page standardized disclosure that describes what the model does, how it was trained, where it works well, and where it fails. Luna: And we're not talking about some Silicon Valley trend. I've seen RFPs from insurance companies, banks, even a large healthcare system — all of them now include a request for a model card in the initial vendor packet. Lucas: Right. And the interesting thing is that this used to be just a best practice, something that responsible AI teams did voluntarily. But in the past 12 to 18 months, it's become a hard requirement. If you don't have a model card ready to go, you're not even in the conversation. Luna: Let's back up — what exactly is a model card? Because I think some of our listeners might know the term but haven't had to produce one yet. Lucas: Sure. The concept was formalized in a 2019 paper by Google researchers — Margaret Mitchell and others — called 'Model Cards for Model Reporting.' The idea is simple: a one- or two-page structured document that covers the model's intended use, its architecture, training data, evaluation metrics, ethical considerations, and known limitations. Think of it like a nutrition label for an AI model. Luna: And now procurement teams are treating it exactly like a nutrition label — they compare two vendors side by side. If your competitor's model card shows 98 percent accuracy on the demographic that matters to the buyer, and yours shows 95 percent, that's a concrete differentiator. Lucas: Exactly. And it goes beyond accuracy. Buyers want to see things like: what was the training data composition? How bias was measured? What are the model's failure modes? I talked to a sales leader at a mid-market AI startup last week — let's call the company 'Spectra AI' — and they lost a seven-figure deal because their model card disclosed a three percent accuracy drop on underrepresented dialects. Luna: Wait — they disclosed that voluntarily? I would think some startups might try to gloss over that kind of detail. Lucas: They did, and that's actually what cost them the deal. The competitor's model card didn't mention any dialect-level breakdowns. But here's the kicker — when the buyer's data science team ran their own tests, the competitor's model had a similar drop. They just didn't put it in the card. But Spectra was transparent, and that honesty was seen as a liability rather than a strength. Luna: That's brutal. But it also suggests that procurement teams are still learning how to interpret model cards. They're looking for red flags, but they don't always have the context to judge whether a disclosed limitation is acceptable or not. Lucas: That's exactly right. And that's part of the challenge for vendors. You want to be transparent, but you also need to frame the limitations in context. For example, if your model is 97 percent accurate overall and drops to 94 percent on one dialect, you should also note that the training data for that dialect was only five percent of the total corpus, so the variance is within expected statistical bounds. That's a more useful disclosure. Luna: So the model card isn't just a compliance document — it's a sales tool. It can either accelerate trust or create friction. Lucas: Yeah. And vendors who treat it as a chore, something to just check the box, are missing the point. A well-constructed model card can actually shorten the sales cycle because it answers the tough questions before they're asked. It preempts the six-week due diligence phase. Luna: Let's talk about what a good model card should include. What are the must-have sections? Lucas: I'd say there are eight core sections. First, model details — name, version, type, developer. Second, intended use — what it's designed for and what it's NOT designed for. Third, training data — sources, size, composition, any known biases. Fourth, evaluation data — how you tested it, metrics used, results. Fifth, performance metrics — accuracy, precision, recall, F1, and importantly, breakdowns by subgroups if relevant. Luna: That subgroup breakdown is where a lot of vendors stumble. It's hard to do well, and if you don't have the data, you might not want to claim it. Lucas: Right. But if you don't include it, buyers will assume you're hiding something. The sixth section is ethical considerations — any known bias, privacy issues, or misuse potential. Seventh is limitations — explicit failure modes, edge cases, out-of-distribution performance. And eighth is maintenance — when was the card last updated, and how often will it be refreshed? Luna: And frequency matters, right? A model card that's six months old for a model that retrains every quarter is almost worse than having none. Lucas: Absolutely. Procurement teams now check the date. If it's more than one quarter old, they flag it. And if your model has been updated since the last card, they'll ask for a new one before giving final approval. I've seen deals stall for weeks just because the vendor's engineering team was too busy to regenerate the card. Luna: So it becomes an operational issue, not just a compliance one. You need a process for keeping the model card in sync with the model lifecycle. Lucas: Exactly. And the startups that do this well treat it as part of their continuous delivery pipeline. They have automated scripts that regenerate the model card whenever a new model version is deployed. That's the gold standard. Luna: I want to come back to something you mentioned earlier — the idea that the model card can be a differentiator. Can you give me an example where a strong model card actually helped close a deal faster? Lucas: Sure. I talked to a company — let's call it 'Veritas AI' — that sells a document extraction model for legal firms. They built a model card that went beyond the basics: they included a section on 'adversarial robustness' showing how the model performed under input perturbations, like OCR noise or skewed documents. That wasn't in the RFP requirements, but it impressed the buyer's technical team so much that they skipped the usual six-week proof of concept and went straight to a pilot. Luna: That's a great example. They anticipated a need the buyer didn't even know they had. Lucas: Right. And that's the strategic play. If you can show that you've thought about failure modes that your competitor hasn't, you build trust. And in enterprise sales, trust is the currency that closes deals. Luna: So for listeners who are building or selling AI, what's the one thing they should do today to get ahead of this requirement? Lucas: Start with a template. The original Google model card paper includes a template you can adapt. But don't just fill it out once — make it a living document. Assign someone on the product team to own it. And if you're in the middle of a sales cycle, proactively send the model card to your prospect before they ask. That signals seriousness. Luna: Yeah, I've seen that work wonders. One sales engineer I know includes the model card in the initial data room, right next to the security whitepaper and the pricing sheet. Lucas: That's the right instinct. And look, I know this feels like yet another compliance burden for small teams. But the cost of not having a model card is losing deals. And the cost of having a bad one is losing credibility. So it's worth investing the time now, before it becomes a crisis. Luna: If these conversations are useful for what you're building or running, and you want to keep them ad-free, listener support is what makes that possible. You can find us at buy me a coffee dot com slash fexingo. Lucas: It really helps. And it means we can keep digging into topics like this without any sponsor constraints. So thank you to everyone who has chipped in. Luna: Alright, back to the procurement trenches. Lucas, you mentioned automated regeneration of model cards — is that something you see becoming standard in the next year or two? Lucas: I think it has to. Because the alternative is manual updates that never happen on time. I've seen startups with 10-person teams try to maintain model cards in Google Docs, and it just doesn't scale. There are now tools that integrate with MLOps platforms to generate cards automatically from model metadata. Luna: Are we likely to see a common standard emerge? Right now everyone uses a slightly different format, which makes comparison hard for buyers. Lucas: That's the next frontier. I know that the ML Commons group is working on a cross-industry standard for model reporting. And some large buyers like JPMorgan have internally standardized on a specific format that their approved vendors must use. So I think we'll see convergence, but it'll take another 12 to 24 months. Luna: In the meantime, the advice is to pick a format and stick with it. Consistency builds trust. Lucas: Exactly. And one more thing — don't forget the intended use section. I see a lot of model cards that are vague about what the model should NOT be used for. That's a huge red flag for procurement. If you don't specify boundaries, buyers assume you haven't thought about misuse. Luna: Good point. It's almost like a product warning label. 'Do not use this model for medical diagnosis' — that kind of clarity is what buyers want to see. Lucas: Right. And if you do include those boundaries, it actually protects you legally as well. So it's a win-win. Luna: Well, I think our listeners have a clear action item: go audit your current model card, or if you don't have one, start building it today. This isn't going away. Lucas: And if you want to see what a good model card looks like, Google's original paper includes a full example for a vision model. It's a great starting point. We'll link to it in the show notes. Luna: Thanks Lucas. And thanks to our listeners. We'll be back next episode with another deep dive into the real-world requirements of enterprise AI procurement.