We don’t have the luxury to wing it on critical AI product provisions. That’s why understanding the critical issues, the drafting traps, and negotiation goals for these provisions is so important. This lesson provides insights into exactly how to draft and negotiate testing obligation language that hits at the heart of AI product compliance and governance.

Olga Mack and Linsey Krolik spent a recent webinar on AI compliance and governance, including this 17+ minute segment on drafting and negotiating AI product testing provisions.

What do customers want from a testing clause in an AI product contract? Olga highlighted that they want the vendor's promise that someone checked the system for bias before launch, and will keep checking as the system changes. We never had to worry about this with a SaaS deal. Either The software matched the documentation or it didn't. With AI, that match doesn’t determine whether the system complies with applicable regulatory requirements. The AI product’s system can run perfectly and still hand you a biased result. It drifts as the model and the data shifts.

Each webinar drafting segment centers around a sample provision drafted by AI. Here’s the language we reviewed in this segment:

"Vendor shall conduct periodic testing of the AI System for bias and disparate impact in a manner consistent with industry standards and shall take commercially reasonable steps to address any material issues identified. Customer may request a summary of testing activities upon reasonable advance written notice." AI-drafted slop language (do not use)

Here are some of the issues raised with the typical provision reviewed during the segment:

“Periodic” - It could mean monthly. It could mean once, three years ago. Linsey explained that vendors like that word because they can drive a truck through it.

“Testing” - Exactly what tests are performed and who is picking them? Without any guidance, the vendor gets to pick. And that means the vendor will likely pick a test it expects to pass.

"Consistent with industry standards" - Nobody can tell you what the standard is right now, and it may not be where you think.

Instead of the approach in this sample language, use a clause that ties testing to a clear timetable, identifies the specific testing methodology or external standard, and ties to applicable law. After all, as Olga noted, a duty to follow the law is harder to argue against than a demand you came up with yourself.

“Testing of the AI System for bias and disparate impact” - Olga urged us all to stop and focus on this phrase. She explained that “disparate impact” is a heavy standard out of employment law. You can act in good faith, do everything right, and still be at fault for the outcome. Proving disparate impact requires statistical studies. In other words, this clause added a legal standard that sounds good but doesn’t really belong. Even Linsey, representing the vendor’s perspective, agreed it has no business in an ordinary commercial contract. Be deliberate and precise in your standard. Tie it to “bias," say how often and how it gets measured, and leave “disparate impact” out.

“May request a summary” - The sample provision says the customer "may request a summary." We are used to the onus being on customers with other product reports. But this standard doesn’t work with an AI system. First, customers don’t know when the vendor ran a test, so they’ll rarely know when to ask. Second, with no timing requirements, the vendor may share data much too late for it to be useful or in a format that makes it impossible to evaluate. Customers should ask the vendor to send results on a schedule within a set period of time after each test, and with flags of anything that failed.

“Commercially reasonable steps to address any material issues” - Another problematic phrase is "commercially reasonable steps to address any material issues." Customers are left again at the whim of the vendor’s decisions how and when to address compliance problems with the system. Customers should identify what material means and include a more precise timeframe for resolving the issue.

The lesson also comes with a two-page handout on testing obligations. It steps back from this one clause and shows how these provisions work overall. You will see what a complete clause covers, from scope to frequency to methodology. You will see where vendor forms tend to go thin. It explains why "results on request" rarely helps the customer. And it walks through the gaps that leave you with no real recourse when a test fails.

Treat it as a desk reference, not a one-time read. Keep it next to the contract when you review your next vendor agreement. The next time a testing clause crosses your desk, you will know what to look for.

The full webinar is available only to How to Contract members.

Click this graphic to read the article outlining the top 10 takeaways from the full webinar that took place on June 16, 2026:

How to Contract members have access to all 60+ webinars recorded in 2026 (including 25 webinars on AI contracts and 25 on core contract subjects) plus 40 webinars on AI contract drafting from 2025. That’s part of our library with over 240+ hours of training videos from the last five years of contract training programs.