How Fortune 500s Use Procurement to Negotiate Vendor AI Output Ownership

Lucas: If these conversations are useful for what you're building or running, I want to start with something that's quietly becoming the most contested line item in enterprise software contracts. It's not price. It's not even data residency anymore. It's who owns the output of the AI that's embedded in the software you're licensing. Luna: You mean like when a sales rep uses Salesforce's Einstein GPT to draft a pitch — does the company own that pitch, or does Salesforce? Lucas: Exactly. And it goes way beyond CRM. Microsoft Copilot in Office, ServiceNow's AI Search, Workday's AI recruiting tools — every major SaaS platform is now shipping generative AI features by default. But the terms of service often say something like 'we may use your content to improve our models.' That's a problem if your content includes proprietary business processes or customer data. Luna: I can see why procurement teams are suddenly very interested. And if today's episode helps you navigate that, the way we keep these shows ad-free is listener support — buy me a coffee dot com slash fexingo. A small way to keep the conversation going. Lucas: Appreciate that. So let me give you a concrete example. Early last year, a Fortune 500 manufacturer I'll call Acme Industrial was negotiating a renewal with a major CRM vendor — think Salesforce or a close competitor. The vendor had just launched a generative AI assistant that could auto-generate call scripts, follow-up emails, even pricing proposals. Acme's procurement team flagged the default output ownership clause. Luna: What did the default say? Lucas: It said the customer owned the output, but the vendor had a broad license to use it for model training and improvement. That's pretty standard in consumer AI tools, but in a B2B context it means your sales playbooks, your pricing strategies, your customer insights — they become part of the vendor's training data. Acme's legal team freaked out. Luna: Rightly so. If a competitor uses the same CRM, the AI could essentially learn from Acme's best practices and serve them to the competitor. Lucas: That's the nightmare scenario. So Acme's procurement team came to the table with three specific contractual options they wanted to discuss. I'll walk through them because they're becoming a template in the industry. Luna: Let's hear them. Lucas: Option one: treat the AI output as a 'scratch space' — the customer owns it fully, and the vendor has zero rights to use it for training. This is the most protective for the customer, but vendors push back hard because they lose the data that improves their models. Option two: treat the output as a 'modified work' — the customer owns the output, but the vendor retains rights to use anonymized, aggregated data for training, as long as they don't reproduce customer-specific content. Luna: That's a compromise a lot of vendors are willing to accept. Lucas: Option three: joint ownership with a non-assertion covenant. Both parties own the output, but each agrees not to sue the other for copyright infringement. This is actually what Acme ended up negotiating. They got the vendor to agree that AI outputs generated using Acme's data would be jointly owned, and the vendor would not assert copyright against Acme if Acme reused those outputs elsewhere. Luna: But joint ownership can be messy — who controls licensing to third parties? Lucas: Great point. In Acme's contract, they added a clause that neither party could license the output to a third party without the other's written consent. That prevented the vendor from selling Acme's ai generated sales scripts to a competitor. And they also included a data deletion clause — if Acme terminated the contract, the vendor had to delete all prompts and outputs within 30 days. Luna: So the vendor can't hold onto Acme's data as a training asset after the relationship ends. Lucas: Exactly. Now, the legal backdrop here matters. In March 2025, the US Copyright Office issued new guidance on ai generated works, saying that outputs with 'sufficient human authorship' can be copyrighted. But if the AI generates a sales email with minimal human modification, the copyright status is murky. That uncertainty is spooking legal teams, which is why procurement is stepping in. Luna: It also creates a trade secret exposure issue. If I'm typing a proprietary pricing formula into a prompt, and the vendor's AI uses that to train, I've just disclosed a trade secret. Lucas: That's actually the bigger risk in my view. Copyright is important, but trade secret misappropriation is a whole different liability. Acme's contract specifically prohibited the vendor from using any prompt content that the customer marked as 'confidential' for any purpose beyond delivering the service. And they required the vendor to have technical controls — like prompt isolation — to ensure customer data wasn't leaking into the training set. Luna: Did the vendor agree to that? Lucas: After some back and forth, yes. The vendor argued that prompt isolation would degrade model quality, but Acme pointed out that Microsoft offers a similar feature in its Copilot commercial license — the ability to opt out of training data use. So the vendor couldn't claim it was technically impossible. Luna: That's a great negotiating tactic — using a competitor's existing offering as a benchmark. Lucas: It works because the vendor's legal team knows that if Microsoft can do it, they can too. The key is to have that example ready before you sit down. Now, one more layer: Acme also negotiated an audit right specifically for AI output usage. They wanted the ability to verify that the vendor wasn't using their outputs beyond what the contract allowed. Luna: How does that audit work in practice? You can't just walk into a data center. Lucas: It's typically a third-party security audit, like SOC 2 Type II with a specific AI module — some auditors now have 'AI training data usage' as a control objective. The vendor has to provide a report annually, and if there's a dispute, the customer can bring in their own auditor at their expense. Acme also got a clause that if the audit found misuse, the vendor had to reimburse the audit cost. Luna: So the vendor has skin in the game. Lucas: Exactly. Now, the interesting thing is that this isn't just about large enterprises. Mid-market companies are starting to see these clauses too. I talked to a procurement director at a $200 million logistics firm who said their CRM renewal last month included a similar AI output clause. They weren't as sophisticated as Acme, so they just accepted the default. That's a risk. Luna: What would you recommend for a smaller company without a dedicated procurement team? Lucas: At minimum, ask the vendor to sign a data processing addendum that specifies the AI output is owned by the customer and cannot be used for training without explicit consent. Most vendors will say yes to that if you push. And if they refuse, that's a red flag — it means your data is their product. Luna: Good advice. I also think the trend is moving toward more regulation. The EU AI Act has provisions about training data transparency, and I could see output ownership becoming a statutory right. Lucas: That's a longer-term shift. For now, the contractual options are the main lever. But the landscape is changing fast — I've heard of at least two vendors who are now offering a 'no training use' tier at a premium price. So the market is beginning to segment. Luna: Which means procurement teams need to decide whether that premium is worth it, or whether to negotiate it into the standard price. Lucas: Right. And the calculus depends on how sensitive your data is. For a manufacturer like Acme, the risk of a competitor learning their pricing strategy is huge — so paying a premium might be justified. For a company that uses AI to generate low-stakes content like internal memos, the default might be fine. Luna: So it's a risk-based decision, not a one-size-fits-all. Lucas: Exactly. And that's why procurement professionals are becoming central to AI governance. They're the ones who can assess the specific use cases, data sensitivity, and vendor leverage before signing. Luna: I want to circle back to the audit clause — did Acme actually ever exercise it? Lucas: Not yet. The contract is only about 18 months old. But they did conduct a mock audit internally to test the vendor's reporting, and they found that the vendor's standard usage logs didn't differentiate between ai generated content and user-typed content. So they went back and asked the vendor to add a metadata tag on AI outputs. That's now in the contract as a technical requirement. Luna: So procurement didn't just sign a paper clause — they validated it operationally. Lucas: That's the gold standard. If you're in procurement, get your IT team involved early. Ask the vendor for a demo of their audit logging. If they can't show you how they track AI output usage, that's a problem. Luna: It also highlights how fast this area is evolving. Six months from now, there might be new model architectures that change the game. Lucas: True. But the contractual principles — ownership, training use, audit rights, trade secret protection — those will remain relevant even as the technology changes. That's the part worth getting right now. Luna: Any final takeaways for listeners who are about to negotiate a software contract with AI features? Lucas: Three things. One: know what the default says — read the output ownership clause before you sign. Two: have a range of options ready — from full ownership to joint ownership to anonymized training. Three: tie it to trade secret protection, not just copyright. And if you can, benchmark against a competitor's offering. That gives you leverage. Luna: And if you're not negotiating a contract today, at least start tracking which AI features your vendors are rolling out. It's only going to become more important. Lucas: Absolutely. That's our time for today. Thanks for listening.