As companies look for practical ways to integrate AI into legal, compliance, and business operations, speed must be balanced against accuracy, security, and human judgment. Peter DiMattia, senior legal counsel at Chevron Phillips Chemical, discusses his "Stitching Framework," a six-step approach for combining AI with trusted information, organizational knowledge, safeguards, and human validation that DiMattia and Justin Pierce, a co-chair of Venable's IP Division, first wrote about last year.
In this conversation—a follow-up to that article—DiMattia explains how the framework can help organizations use AI more efficiently while creating the governance, traceability, and continuous learning needed for responsible innovation.
Justin Pierce: How would you describe or define the Stitching Framework?
Peter DiMattia: I think it's a multi-step process that integrates AI to help effectuate faster output, but with the key validation step of the information that you're outputting. So it helps increase speed and accuracy, because one of the main parts of the Stitching Framework is the validation at the end. That's very important.
It's really strong for AI use cases where you have good existing documents or frameworks or playbooks, and you can use that information to help really breathe life into the AI and achieve a better output in a more efficient way—something that's user-friendly and something you can take with you and use for the next step in your process, or maybe as one of the end outputs.
Justin: The framework has six procedural steps. Step One is issue intake and contextual framing. What does that entail?
Peter: It kind of comes down to what the issue at hand is, what the problem to be solved is, and why you're doing it. That tells you whether you even need to go down this path or not. It's about framing it out, digesting it. If you start with bad, you get bad results. If you start with good, you get good results. And if you start with at least some dialogue around it, you minimize the risk of a bad result and increase the chance of a good result.
You can use an intake form process of some sort. It could be very minimal. It could be filled in by the lawyer, or it could be filled in by the lawyer and the stakeholder, or by the stakeholder alone. That helps frame the issue and challenges it—is it really an issue?
Justin: Step Two is precedent and reference integration. What is that?
Peter: Well, that's where the good stuff is—that's the tried and true stuff. Maybe a good patent for reference, from either a drafting perspective or a related technology perspective that has gone through the process and has really laid things out well. It could be a framework for an office action response for a particular field—using that structural aspect and framework as leverage in your next office action to hopefully speed things up.
It could be contract analysis—making sure that any type of confidential information is assessed, whether you can use it or not, that kind of stuff. But maybe just some general contract perspective of, "Here's a framework, here's a playbook for what we're really looking for."
Justin: Step Three is the stitching process. How would you describe that?
Peter: This is where the soup's made. This is where the bread's kneaded. This is where you're putting all the core feedback together. "Here's the goal, here's the background. This is some additional information," either from input from various stakeholders—maybe including the lawyer, the businesspeople—and I put it in that system where it's going to help breathe life into the answer.
You're stitching all these elements together for a multifactorial answer at the end of the day. That's where the stitching is—all those elements come together.
Justin: Why is it important that the process be auditable?
Peter: You want to have some view into how things are being put together. It may not always be possible—if you have an off-the-shelf system—but I think you can attenuate it. We may not necessarily know the algorithm per se, but you can attenuate it by looking at various responses and outputs from your inputs and challenging yourself: "Is this a good output? Is this a good input? Do I need to put more information in there? Do I need to fix my playbook?"
You're testing outputs, almost to the extent of doing that experimental phase. You always have to validate, but at least you've challenged the system to say, "Am I getting what I should get out of it?"
Justin: Step Four is data classification and safeguards. What's entailed in that step?
Peter: All those inputs we talked about in the previous steps—we have to make sure we audit those. Can we actually use that information? And then it also determines, "Okay, if there's some type of something going on—is this going to be discoverable material? Maybe there's a legal hold on something, or some type of pending litigation." It could be sensitive intellectual property and information.
Sometimes you may have a composition you're making, and for some of those additives if there's an NDA, you make sure you don't have any kind of roadblocks to the information you're putting in there.
Justin: Step Five is AI synthesis and human validation. Where does the human fit into the framework?
Peter: The AI synthesis is AI putting it together. But the human validation is one of the most important steps. You've gained all the efficiencies from this stitching process, but you want to make sure your output—number one—answers the question that you started out with. Number two, it's supported by the information. And then sometimes, even in the sense of the AI synthesis, asking it for an output of, "What were your sources?" and checking your sources.
Sometimes regulations, intellectual property, and contracts are so nuanced. It can pick up the nuance, but maybe there's some type of historical customer-interfacing aspect that it's not accounting for. It's like, "Well, that is a good answer, that's the correct answer, but maybe we want something more tailored and suited." So that's where that human validation comes in—to customize.
Justin: Step Six is comparative validation and traceability. What's the goal?
Peter: I think this is all about traceability. It's essentially making sure you're sourced—you have that documentation stream of where the output came from.
If someone says, "How'd you come to this conclusion? How did you come to this response to a regulator?"—you make sure you have it all documented. That's the final packaging. It's making sure you have good bookkeeping, good record keeping at the end, to support your conclusions.
Justin: What are some of the benefits of implementing something like the Stitching Framework?
Peter: I think it helps with next-level thinking, because it takes a lot of the routine tasks away and allows you to have more deep-think time. Using these tools gives you more time to get out there and maybe dialogue with people you may not have necessarily had time to dialogue with because you're working on the work, per se.
I think it helps with better ideation—or it helps free up the opportunity for better ideation—with the side benefit of being able to do some faster work. So I think it all works together—from an innovation side, helping efficiency number one, innovation number two, and a better result.
Justin: What's next?
Peter: One of the things AI could be helpful with is pulling those weak signals from various data points—that weak signal analysis—and maybe leveraging it to say, "Okay, what should be our next R&D project? What should be our next direction? Do we have a litigation risk? Do we have a compliance issue? Do we have some type of competitor that we don't know about today, but they seem like they could be coming tomorrow?"
I think adoption is still something we have to keep focusing on, and we have to keep challenging ourselves. Sometimes maybe you don't get the result you want, but that doesn't mean you stop there. It's an iterative process, and like anything else, you put a first draft together, and it may not be your best draft.
Justin Pierce hosts Venable's AI and IP: The Legal Frontier, a podcast designed to help your company use AI and IP law to gain a competitive edge. Click here to visit the podcast's website or listen on your favorite podcast player.