How In-House Lawyers Get More From AI by Asking a Bigger Question

When I was a white-collar investigations attorney, I learned that the most useful question in an interview is usually the broadest one: "What do you want to tell me that I haven't already asked?" "What else should I know about this?" Open-ended questions are the ones that surface the real, meaty answers. A checkbox or a drop-down only gives you back the answer you already expected.

Throughout my career, I’ve forgotten that lesson when I delegate to team members: “Here's the problem, please go find the answer.” And more recently, when I hand a task to AI, I tend to make the same mistake. The prescriptive reflex hasn't changed; only the recipient has.

Tell an AI tool (or a human!) exactly what to look for, and it will (hopefully) find it. Give it room, and it can tell you something you didn't know to ask. The size and scope of the answer is usually set by the size and scope of the question.

This kept coming up during the final session of the Claude for Legal: In-House Edition series. One panelist after another described a build that started with a broad question rather than a narrow task, and in every case, that openness turned up something they hadn't set out to find.

Ask What Happened, Not What's Wrong

Abbey Martin, General Counsel at ID5, treats her own assumptions as the first thing to interrogate. "Am I even solving the right problem?" she asked during the recent webinar. "Or am I holding the tool back with my assumptions or prescriptions of what I think the problem might be?"

She had a specific problem in mind: deals were taking too long, and she was sure she knew why. So she tested that hunch. She connected her source systems, pointed Claude at a single deal that had dragged on, and asked it to read the Slack channel and the emails and simply tell her what happened. She didn't ask it to find the delay or summarize the redlines, just to tell her what happened.

About an hour and a half of back-and-forth later, she'd built two skills: a deal audit and a deal readiness check of roughly 300 lines of instructions which run in about eight minutes. The audit gave her a forensic retrospective of a deal she thought she already understood. What she found was that a structural question that should have been a day-one intake item didn't surface until day 23, and when it did, it "triggered rework across every downstream document." That one late question cost 22 days.

It also put a number on something "that nobody measures, but every lawyer feels": context switching. On one complex deal, a single team member had made 48 separate context switches. Across all the deals she audited, 30 to 40% of legal deal time was preventable. In her words, these were figures that were "previously invisible and are now reportable metrics" — the kind that turn a vague sense of being overworked into an actual case for headcount.

"The gap between [finding] a 2-hour time saving on redlining and a 22-day saving on the deal cycle was not a better tool," she said. "It was a better, more open-ended question." And she built the whole thing on her terrace, glass of wine in hand.

Take Two Assumptions Off Before You Ask

Abbey also shared advice she borrowed from Coco Chanel, who advised “Before you leave the house, look in the mirror and take one thing off.” Write down your problem statement, Abbey says, then take out two of the assumptions buried inside it about what's actually wrong before you hand it to the tool.

She named three ways we sabotage ourselves without noticing:

1. Fixing a problem we misdiagnosed. She was "absolutely sure that the bottleneck was legal speed and bandwidth," and the tool showed her "unequivocally, it was intake quality."

2. Solving only the problems we already know about.

3. Boxing the tool into a task so narrow it can't surface anything we didn't already suspect.

Her rule is evidence first, diagnosis second: let the tool pull the record and do the reading before it reaches a conclusion. That's the opposite of how most of us prompt, which is to hand over our conclusion and ask the tool to back it up.

A bigger question is also rarely a one-shot. Michael Berrini, Assistant General Counsel at Ironclad, put it plainly: "iterating and challenging and refining is a very important part of this process. You don't just ask one question and get an output and you're done."

His own build pointed Claude at a full quarter of enterprise sales agreements, through Ironclad's MCP connector, to find which clauses drew the most pushback and how they actually got resolved.

That didn't come from one prompt. It came from a running conversation that got sharper with each round, until the tool understood what counted as a high-risk clause and where his team actually tended to land.

Once a conversation like that produces something worth keeping, Claude lets you save it as a skill (a procedure it runs the same way every time) so you can point it at fresh data next quarter and run the whole analysis again.

Then Make It Argue With You

Dina Segal, Chief Legal Officer at Gusto, has a habit worth stealing: before she considers a build finished, she turns the tool on itself. "What did I miss? What else can I do?" she asks it. "If you are a really skeptical audience looking at this, what wouldn't you like? Where could you, Claude, get this wrong? Where might you hallucinate?" The iteration she gets back from those questions, she said, has been unbelievable.

It's the open-ended question again, just pointed in reverse. First you give the tool room to tell you what's really going on. Then you give it room to tell you where it might be fooling you. Both work for the same reason: you've stopped supplying the answer and let the tool do the thinking you brought it in to do.

Dina's other reminder is the one that makes any of this safe to rely on: always tell it to link back to the source. "We as lawyers and compliance folks, we like to actually read the law," she said.

There's a companion move to that: telling it what to do when there's no source to point to.

Smita Rajmohan, Assistant General Counsel for AI at Intuit, instructed Claude that if it couldn't find an answer in her vetted documents, it shouldn't invent one, it should flag the gap and point the person back to her. The quality of the output, as she put it, depends on the accuracy of the sources behind it.

Where This Leaves Us

Abbey's homework is a good place to start, so I'll pass it along exactly as she gave it. Sit down later today, give the tool everything on your most challenging deal from last quarter, make sure your connectors are on, and ask it one thing: where did the time go?

You already know how to spot the issue and drive toward the answer. The newer skill is resisting that instinct for a minute, asking something bigger than feels comfortable, and seeing what the tool turns up when you stop telling it where to look.

This post draws on a live practitioner session from The L Suite's Claude for Legal: In-House Edition webinar (Part 3), featuring Michael Berrini (Assistant General Counsel, Ironclad), Dina Segal (Chief Legal Officer, Gusto), Abbey Martin (General Counsel, ID5), and Smita Rajmohan (Assistant General Counsel for AI, Intuit), moderated by Tommy Taveras-Verreira (Chief Strategy Officer, Law Trades).