Q: A lot of legal ops leaders think about AI as a tool only for their own team. What’s wrong with that framing?
It’s not exactly wrong—it’s just leaving most of the value on the table. I think people overindex on their own personal experience with AI, or what it’s going to do for their department. But if you zoom out and think about how to enable the people who work with you and rely on you, that’s where you can make a huge impact.
I had someone pull an audit of the last six months of AI usage and we’ve had nearly 12,000 interactions. Legal is the heaviest user, but I’ve got finance and accounting, customer success, sales, sales ops, product, and security all using different agents. That told me: this is meaningful for my whole enterprise, not just my team.
Take our security team for example—they’re actually the top users of one of our agents. Before, whenever they needed to confirm a confidentiality framework was in place before sending a SOC 2 report to a customer, that case would land in our legal queue until we picked it up, read the contract, and got back to them. A lot of the time that was happening in the middle of a sales cycle. Now, they just run the agent themselves and quickly confirm the framework is in place.
The question I’d encourage every legal ops leader to ask is: what are the tasks happening across my business that have a connection to contracts—and how do I remove myself as the bottleneck? Nobody likes waiting on us any more than we like having to look up the information.
Q: How do you make the business case for investment—especially when you’re competing with other priorities?
The framing matters a lot. If you go in asking for budget for a redlining tool, the conversation is going to be around one key question: how much faster will it make you?
And honestly, that’s a hard number to produce. Review time isn’t a metric most teams track, all contracts are not created equal, and the impact is largely contained to your own department. So you have to be tactful when you bring up the conversation.
Instead frame it as a solution with tangible value. For example, “Security asks us about confidentiality 50 times a month. That’s at least 250 minutes of legal team time per month and it’s 50 moments where a SOC 2 report or NDA is sitting in a queue instead of going out the door and keeping a sales cycle moving”. That changes the conversation because you’re speaking the language of the business, not the language of legal.
When you can show that an AI investment benefits finance, security, sales ops, and customer success, you’re going to get a very different response than when you’re just asking for something that speeds up your own work.
Q: For someone just getting started, what do you wish you’d known earlier?
- Get familiar with AI in a general sense to understand what it is and how it works.
There’s a lot of content out there to educate yourself. But more importantly, you need to understand the difference between using something off-the-shelf and using something that is intentionally integrated withyour data. - One of the biggest risks with AI is hallucinations or inaccurate outputs.
The way to address that is to ground the AI in a defined, trusted data structure. When the model is answering from a specific set of things you know are good, rather than from the entire internet, the accuracy improves dramatically. And accuracy isn’t optional. If I have to hold everyone’s hand every time, I haven’t actually solved the problem. - The other thing I’d say: AI is like any new thing.
There’s initial excitement, and then people often regress back to what they were doing before. If you build something and someone uses it two or three times and thinks “this doesn’t seem that great,” they’re not going to use it. You only get one shot at that first impression. Build for reliability before you build for breadth.
Q: You’ve built a lot of agents at this point. What does that process actually look like?
It varies a lot—some agents I’ve built in five minutes and some in twenty. My risk scoring agent took about six hours. That one was complex because I had so many key terms I was directing it to interpret.
What you need is to be a scientist and have a hypothesis. If you have it in your head, “I think this prompt will do this thing,” then test it. See what happens and adjust. Precision matters because if you want people to rely on the output, it has to be reliable. Simple as that.
Q: What makes an agent actually get adopted—versus one that people try once and abandon?
I think about adoption in three buckets: mental, emotional, and practical.
The mental piece is: do people understand what this is doing? If they get results they don’t expect, they won’t trust it and they won’t use it. So accuracy and explainability matter a lot.
The emotional piece is: am I OK letting go of the old way? If the output is accurate and it’s less painful than what they were doing before, that barrier falls quickly.
The practical piece is: how easy is it to actually use? If someone has to do ten steps to get to the answer, adoption suffers. The easier you make it—ideally, push a button and get what you need—the faster it spreads.
Once you get all three right, adoption takes care of itself.
Q: You mentioned self-service AI design capabilities as critical. Why does that matter so much?
Because so much of getting AI to perform well is rapid iteration and you can’t iterate rapidly if you have to go back to your vendor for every small change.
You need to be able to go in, tweak your prompt, tweak your agent, and run it yourself. That’s how you get from “interesting but unreliable” to “this thing actually works.” If you’re filing support tickets to make adjustments, you will never move fast enough.
Also: make sure your tool shows you how the AI is reasoning. That sounds like a minor feature, but it’s actually essential. When I was building my product dashboard agent, I could look at the thought process and say, “OK, that’s not quite how I want it to think about this,” and then adjust the prompt with that specific context. Without that visibility, you’re just experimenting blindly. You’ll get there eventually, but with a lot less precision.
Q: What does “contract intelligence” actually mean to you in a practical sense?
Honestly, I think it’s largely a relabeling of what people who use AI well are already doing: the more context you give AI, the better its output.
When you use a random LLM, it’s drawing from everything. When you use something intentionally integrated with your contract data—a defined data structure you know and trust—it’s drawing from that. That accuracy difference is significant.
Context has always been the key. When I first started experimenting with AI, just giving it a role and defining an audience got me better outputs. That’s context. Prompt engineering is context.
What “contract intelligence” is really describing is what happens when you layer that kind of intentionality on top of a system that itself provides structured, domain-specific context. The accuracy gets high enough that you can actually rely on it.
And when people can rely on it, they use it. When they use it, they love it!
Q: Any final advice for legal ops leaders who are about to jump in?
Trust yourself. I see a lot of people who think this is beyond them and that it requires some technical background they don’t have. It doesn’t.
You already know what a good answer looks like. You know what people ask you. You know your contracts, your playbooks, your business. That’s your prompt engineering superpower.
My best advice: start with one thing. Build it. Test it. Get it working reliably. Then share it. The first time a colleague says “I can’t believe how fast that was,” that’s when this stops being an experiment and starts being a practice.
Legal Ops AI Best Practices
Foster’s journey from curious early adopter to 12,000 AI interactions across six departments didn’t happen because he had an “AI prompting” background or pressure from leadership. It happened because he asked a simple question: who else is waiting on me, and what would it mean if they didn’t have to?
Start with one problem. Build for reliability. Share the win. The rest tends to follow.
Ready to get started with AI? Schedule a demo of Pramata today.