Latest / The Operator's Manual with Fexingo: Daily Decisions, Tactical Wins, and Business Execution / How One Law Firm Cut Discovery Costs 65 Percent With One Prompt
Transcript
- Lucas: Let me tell you about a mid-sized litigation firm in Chicago that cut its e-discovery costs by 65 percent and reduced document review time by 70 percent with one change. Not a new AI platform. Not a bigger budget. One structured prompt. Luna: One prompt — as in, they typed a better instruction into the AI tool they already had? Lucas: Exactly. The firm was spending roughly $2.5 million a year on e-discovery for about 50 active cases. Most of that went to contract reviewers — paralegals and junior associates — loading documents into a review platform and tagging them for relevance, privilege, responsiveness. Standard stuff. They had already deployed an AI model to pre-tag documents, but the model was giving them about 70 percent precision, meaning reviewers still had to re-check a lot of flags. Luna: So the output was noisy, and the reviewers didn't fully trust it. That's pretty common. Lucas: Right. And here's where the one-prompt change comes in. Instead of using the default instruction the AI model came with — something generic like 'identify documents responsive to the case' — the firm's head of legal ops spent a day writing a very specific prompt. It included the exact legal standard for responsiveness, a list of excluded file types, a date range, a list of key players and their role descriptions, and an instruction to flag any document where a key player appeared within five lines of a financial term. That was the key constraint. Luna: That is so much more structured than what most teams do. Most people just type 'find relevant docs' and hope the model reads their mind. Lucas: Exactly. And the result was striking. Precision jumped from 70 percent to 94 percent. Recall actually improved slightly too, because the model stopped over-tagging noise and focused on the signal. Reviewers went from checking every pre-tagged document to spot-checking about one in ten. The cost per document fell from about 40 cents to 14 cents. Over a year, that saved the firm about $1.6 million. Luna: And how long did it take to see those improvements? Did they have to retrain the model? Lucas: No retraining at all. Same model, same architecture. They just swapped the prompt and ran a test set of 5,000 already-reviewed documents. The new prompt outperformed the old one on every metric within one afternoon. They rolled it out to all active cases the next week. Luna: That's the kind of speed that makes operational people smile. But I'm curious — did they have any pushback from the reviewers? I mean, if you're a paralegal whose job just got a lot more efficient, maybe you start worrying about your role. Lucas: Great question. The firm actually handled this well. Instead of cutting headcount, they reassigned the senior paralegals to higher-value work — deposition prep, trial exhibit organization, expert witness coordination. The contract reviewers, who were mostly temps, were gradually reduced as cases closed. The permanent team felt their work got more interesting, not less secure. Luna: So the prompt change didn't just save money — it reshaped the workflow in a way that actually improved job quality for the people who stayed. Lucas: That's the ideal outcome. And I think it points to a broader principle that's relevant far beyond legal. Most of the operational gains from AI right now aren't coming from flashy new models. They're coming from people who take the time to write a really good instruction — a prompt — that constrains the model's output to exactly what the business needs. Luna: It's almost like the prompt is the new process document. In the old world, you'd write a standard operating procedure. Now you write a prompt that enforces that procedure automatically. Lucas: That's exactly how the legal ops head described it. She said, 'I basically wrote a one-page SOP and turned it into a prompt.' And it worked because she understood both the legal domain and the model's capabilities. She knew which terms the model would interpret reliably and which ones it would confuse. She tested different phrasings for the financial-term proximity rule until she got consistent results. Luna: That kind of domain knowledge is what separates a good prompt from a great one. You can't just copy a template from a blog post. Lucas: Right. And the firm has since built a small library of prompts for different case types — employment litigation, contract disputes, regulatory investigations. Each prompt has about eight to twelve specific constraints. They share them across practice groups. It's become a reusable asset. Luna: So the one prompt turned into a playbook. That's the kind of scaling story I love. But let me push back a little — does this work for really complex litigation? Like, fraud cases with hundreds of custodians and multiple languages? Lucas: Fair point. The firm's cases are mostly mid-market. They don't do the billion-dollar mega-discoveries. But the logic holds: a more structured prompt outperforms a generic one at any scale. The head of legal ops told me they're now experimenting with multi-prompt workflows — one prompt for first-pass relevance, a second for privilege, a third for confidentiality. That way each prompt stays narrow and precise. Luna: That's smart. Modular prompts. You can tune each one independently without breaking the whole system. Lucas: Right. And the cost savings are real. The firm's outside counsel bills for document review at an average of $85 an hour. With the old workflow, a typical 100,000-document case required about 2,500 hours of review. At $85, that's over $200,000 just for review labor. After the prompt change, the same case took about 750 hours. That's a saving of nearly $150,000 per case. Luna: And that's just direct cost. There's also the time to answer benefit. If you can get through discovery in three weeks instead of ten, you can start building your case strategy earlier. Lucas: Exactly. And that changes the dynamics of settlement, motion practice, everything. The firm's general counsel said that faster discovery has already helped them win two early summary judgments because they had the evidence organized and ready before the other side even finished their review. Luna: That's a competitive advantage that goes way beyond cost savings. It's strategic leverage. Lucas: And it came from one focused day of work by someone who understood the problem deeply. Not a six-month IT project. Not a new software vendor. Just a person with domain expertise, a text editor, and a willingness to iterate. Luna: You know, it's stories like this that make me think the most underrated skill in business right now is prompt engineering — not in the hype sense, but in the craft sense. Writing an instruction that a machine can follow reliably. Lucas: I think that's exactly right. And if today's episode moved your thinking forward on that, even in a small way, you know where to find the link. It's buy me a coffee dot com slash fexingo. That's the only way this show stays ad-free — listener coffee. No pressure, just a real option if the show has paid for its own coffee in your day. Luna: Yeah, it's a straightforward way to keep the conversation going without any sponsors. Appreciate everyone who has chipped in. Lucas: So back to the prompt library idea. The firm is now exploring whether they can package some of these prompts as templates for their smaller clients who handle their own discovery in-house. That would be a new revenue stream built entirely on the prompt work. Luna: That's a fascinating business model. Sell the instruction, not the hours. It flips the whole billing structure. Lucas: Exactly. And it's still early. But the head of legal ops said they're planning to offer a 'prompt audit' service where they review a client's existing e-discovery prompts and suggest improvements. One lawyer I spoke with called it 'prompt consulting' — which might be the most 2026 job title I've heard. Luna: I love that. It's the next evolution of legal process improvement. First you automate the review, then you automate the automation. Lucas: That's a great way to put it. And the takeaway for anyone managing a knowledge-intensive operation is simple: before you buy a new tool, ask yourself whether the tool you already have is being used with the best possible instruction. The answer might save you a million dollars. Luna: Alright, worth thinking about next time I open a new SaaS subscription page. Lucas: Exactly. That's all for today. See you tomorrow.