Latest / B2B SaaS Talks with Fexingo: Enterprise Software, Sales Cycles, and Procurement / Why Enterprise Software Deals Now Include a Vendor AI Model Explainability Mandate
Transcript
- Lucas: Luna, I want to talk about a clause that is quietly rewriting enterprise software contracts this year. It's something we haven't covered yet on the show, but I'm seeing it in almost every deal memo over a certain size. Luna: Okay, I'm intrigued. What is it? Lucas: The AI model explainability mandate. Basically, buyers are now demanding that vendors prove their artificial intelligence can explain its own decisions in a way a human can actually understand. Luna: Right — not just that the model works, but that the vendor can articulate why it gave a specific output on a specific input. Lucas: Exactly. And this is a direct consequence of the EU AI Act, which started enforcement phases earlier this year. But it's not just European companies. us based Fortune 500 firms are adding this to their procurement checklists too. Luna: I've heard about the AI Act requiring transparency, but what does that actually mean in a software contract? Like, what does 'explainability' look like in black and white? Lucas: Good question. Let me give you a concrete example. I spoke with a procurement director at a large financial services firm — let's call them a top-five US bank. They recently walked away from a twelve million dollar deal with a fintech vendor because the vendor couldn't explain why its credit-scoring model flagged certain loan applicants. Luna: Wow. Twelve million. And the model probably performed well on accuracy metrics? Lucas: It did. The AUC scores were great. But when the bank's compliance team asked the vendor to produce a human-readable explanation for a specific declined application — say, a small business owner with a 720 credit score and three years of revenue growth — the vendor came back with a feature importance chart that listed seventeen variables and their weights. Luna: Which is basically a black box with a little window. Lucas: That's exactly what the compliance team said. They needed to be able to tell the applicant, and the regulator, why the model made that decision. The vendor didn't have that capability. So the deal died. Luna: And this is happening in healthcare, insurance, HR tech — any industry where a model's output affects people's lives or regulatory exposure. Lucas: Right. So let's define the terms clearly, because I think they get confused a lot. There's interpretability — which means you understand how the model works internally. And there's explainability — which means for a specific input, you can trace why that output occurred. Luna: The distinction matters because a vendor might say 'our model is a linear regression, so it's inherently interpretable.' But if your linear regression has 200 coefficients interacting in non-obvious ways, good luck explaining a single prediction to a regulator. Lucas: Exactly. So what buyers are now asking for is a written explainability policy. They want to know: for each model, what method do you use to generate explanations — LIME, SHAP, counterfactual examples? And can you generate those explanations on demand, at scale, in real-time? Luna: That last part is the killer. Generating a SHAP explanation for one prediction is computationally cheap. But if your enterprise customer has millions of predictions a day, can your system produce an explanation for each one without slowing down? Lucas: And without tripling your cloud bill, which some vendors have found to be a problem. I've seen a startup pitch where their explainability module added 40 percent latency to each inference. That's a non-starter for real-time applications like fraud detection or loan origination. Luna: So the contract clauses I'm seeing now have three layers. One: the vendor must maintain an explainability framework for all AI models used in the service. Two: the vendor must provide per-prediction explanations upon request within a defined time window, say 72 hours. Three: those explanations must be in a format that a non-technical regulator can understand. Lucas: Right. And there's a fourth layer that I think is emerging: the buyer can audit the vendor's explainability system. Some procurement teams are literally sending in adversarial examples — edge cases designed to break the model — and demanding explanations for those too. Luna: Which is a clever way to stress-test not just the model, but the vendor's commitment to transparency. If the vendor's default response is 'we can't explain that because it's a rare case,' that's a red flag. Lucas: Let's talk about what this means for the sales cycle, because that's where our listeners live. Traditionally, a SaaS demo focused on features, UI, ROI. Now, the sales engineer needs to be able to script a conversation about explainability. Luna: Right — not just 'here's how our AI works,' but 'here's how our AI explains itself.' And if the buyer asks a question the sales engineer can't answer, that can kill a deal that's been six months in the making. Lucas: I've seen procurement teams actually write explainability questions into their RFPs. They'll submit five or six test cases and ask the vendor to provide explanations as part of the evaluation. The vendor that can't deliver — or delivers a vague answer — gets scored down. Luna: It's a whole new dimension of technical due diligence. And it's not just compliance-driven. I think buyers are using this as a proxy for overall engineering rigor. If you can't build explainability into your model, what else is sloppy under the hood? Lucas: That's a smart point. Explainability forces a certain discipline. You have to document feature engineering, data provenance, model versioning, and decision logic. That documentation is valuable for the buyer, but it also makes the vendor's own team better. Luna: And vendors who already have this built in are using it as a competitive differentiator. I've seen a CRM company put 'explainable AI certified' on their homepage. That's a direct signal to procurement teams. Lucas: So if you're a B2B SaaS founder listening, and your product uses any form of AI — even a simple recommendation engine — expect this to come up in your next enterprise deal. Start building the explainability layer now, or risk losing seven-figure contracts. Luna: And it's not just about the technology. You also need internal processes. Who in your company is responsible for generating an explanation when a customer asks? The data science team? The support team? You need a clear ownership chain. Lucas: Absolutely. Some vendors are creating a new role: AI explainability engineer. That person sits between data science and customer success, and their job is to handle these requests and maintain the explanation infrastructure. Luna: It's a whole new career path emerging from a contract clause. That's fascinating. Lucas: Okay, speaking of things that emerge from a commitment to transparency — and I promise this connects — one thing we've always tried to do on this show is keep it free from any outside influence. No sponsors, no ads, no one telling us what to cover. Luna: Yeah, that's intentional. We want these conversations to be purely about what's useful for you, the operator or builder listening. Not what some advertiser wants us to say. Lucas: And the way we keep it that way is simple. If you find value in these episodes — whether it's an explainability clause or a sales cycle insight — and you want to support the show staying ad-free, you can contribute at buy me a coffee dot com slash fexingo. Luna: It's a small way to keep the signal clean. And honestly, it helps us know that this content matters to you. Lucas: Anyway, back to explainability. I want to zoom out a bit and talk about the legal implications. Because once you produce an explanation, that explanation becomes part of the contractual record. Luna: Right — if the vendor says 'the model denied your loan because your debt to income ratio was too high,' and that explanation is incorrect or incomplete, the buyer could potentially hold the vendor liable for misrepresentation. Lucas: That's a huge risk. So some vendors are pushing back on explainability mandates by saying 'we'll give you an explanation, but it's for informational purposes only and not part of the contractual SLA.' And buyers are rejecting that language. Luna: Because the whole point is that the explanation is used for regulatory compliance and customer disputes. If it's not contractual, it's worthless to the compliance team. Lucas: So there's a negotiation happening right now around the liability boundary. Where does the vendor's responsibility end and the buyer's begin? And that's going to play out deal by deal over the next twelve months. Luna: I think we'll see a standard clause emerge eventually, maybe from an industry body like the Enterprise SaaS Association or the AI Governance Consortium. Lucas: Hopefully. Because right now every procurement team is reinventing the wheel. And that slows down deals, which nobody wants. Luna: So for our listeners who are on the buyer side — what should you ask for in your next contract negotiation on explainability? Lucas: Three things. One: define what 'explanation' means. Specify the method — SHAP values, LIME, counterfactual — and the format — plain language, not just a feature importance list. Two: set a service level for explanation generation — for example, within 24 hours for any prediction made in the last 90 days. Three: require an annual audit of the explainability system by an independent third party. Luna: And on the vendor side, I'd add: start building your explainability narrative now. Even if you don't have a full system, have a roadmap. Buyers will forgive a gap if you show you're serious about closing it. Lucas: Especially if you tie that roadmap to a specific date — 'we will have SHAP explanations available for all models by Q1 2027.' That shows you understand the requirement and are investing in it. Luna: And if you're a startup pitching to enterprise, this is table stakes now. I was talking to a seed-stage company last week that had a brilliant AI product but no explainability plan. The enterprise pilot fell through because the buyer's legal team flagged it. Lucas: That's the new reality. It's not enough to be accurate. You have to be transparent. And that's a good thing for the industry overall, even if it's painful for vendors who have to retrofit. Luna: Yeah, I think in the long run, explainability will drive better model design. When you know you have to explain every decision, you naturally build simpler, more robust models. Lucas: It's the same principle as writing code that another human has to read. You comment it better, you structure it cleaner. Explainability is just good engineering hygiene, now enforced by contract. Luna: Alright, I think we've given our listeners plenty to think about. Let's check in next month and see if any of these clauses have already started to standardize. Lucas: That's a great idea. We'll track it. Thanks, Luna.