How In-House Lawyers Build AI Workflows That Keep Them in Control

In-house lawyers struggle with two sides of the human-in-the-loop coin: either they are out of the loop or too looped in. The ones building the best workflows have figured out how to strike a balance.

We’ve heard the term for years now: human-in-the-loop. Some of us hate the term. I’m in that camp, though I wholly support human oversight and accountability of AI tools. But I digress…

On one side of the HITL coin is the fear of being taken out of the loop entirely. AI is fast and confident and sometimes just wrong. Agentic AI ups the game. A wrong answer can be hard to catch; a wrong action can be catastrophic.

The other side is this: If I have to verify everything, and verifying takes as long as doing the work myself, why bother? It's a fair question. If AI hands you a fast answer and you spend the time you saved checking whether it's true, you haven't gained anything. You've just moved the work around.

At last week's Claude for Legal: In-House Edition (Part 2) webinar, the thing that stood out to me was how deliberately each of our panelists designed their AI workflows to stop, to surface their own reasoning, and to put the decision back in their hands all while maintaining efficiency and velocity. Each built a workflow they actually use on real matters, where the tool moves fast but not without them.

My take-home point is that the loop doesn't have to be slow. You do a lot of the work once, up front: in how you build the workflow, in the instructions you write, in the language you pre-approve for the tasks you repeat. Then the tool runs fast inside the lines you drew. Verification stops being a tax on every output and becomes a design decision you made.

The Workflow Should Pause for You

Natalie Kim, a strategic advisor at Anthropic who until recently led legal and AI at a Series C company, walked through a scenario every commercial lawyer knows. A big buyer sends over their paper, not yours, and sales wants it marked up by end of day so the deal closes this quarter.

Natalie broke the work into what she called reading it, grading it, fixing it, and then shipping it, and she chained several Claude skills together to move through each step.

A skill, in her words, is like a spell you can invoke. You write the procedure once, and Claude runs it the same way every time instead of you re-typing a long prompt. She created various skills to summarize a contract, to flag what the other side's paper was conspicuously silent on, to grade clauses against her playbook, to drop in pre-approved fallback language, and to clean and prepare the document for sending.

She didn't run the tasks together from start to finish as one skill. She intentionally built the workflow with discrete skills that allowed a pause in between each step.

"Watch how every step puts a decision back into the human lawyer's hands," she said. She built the workflow to stop and show its work so she can review it before moving on. She was honest about the temptation to eliminate the pauses: "I could close my eyes and click accept at every single one, but I would not be doing my job." Maybe a flagged issue doesn't apply to a deal. Maybe there are ten suggested changes and the deal only has room to negotiate two. That triage is the lawyering, and she structured the workflow so it couldn't happen on autopilot.

The same logic showed up at the very end, before the document leaves her hands. Her / clean-up skill drafts the final version and strips the internal comments, but as she put it, "it does not email people on your behalf. It produces a document, it hands it back to me, and then I hit send." It doesn’t ship without her final approval and that is a design choice she intentionally chose.

Build It to Distrust Itself

Tamara Momirov, SVP of Legal and Privacy at mobile game publisher Tilting Point Media and the company’s sole in-house lawyer, used Claude on two matters where the stakes were high and the timeline was tight: a majority-stake acquisition and a regulatory report tied to a settlement.

Her core instruction: tell the tool to distrust itself. Left alone, she said, it was an eager assistant that wanted to tell her what she wanted to hear. With the right prompts, "I turned it into kind of a skeptical junior" that knew the limits of its own judgment and wasn't afraid to say so.

In practice, that meant putting controls on every finding rather than trusting the finding itself:

• A confidence rating on each output, with anything below her threshold flagged as speculative and worth a second look.

• A handful of adversarial inputs for each conclusion, so she could see what might break it and judge, with the full context she had, whether it applied.

• A flag whenever the tool was leaning on its general training rather than the documents she'd actually given it, because that wasn't verified information.

• A requirement that every flag point to the exact document and clause it came from, so she could check the reasoning step by step.

"The credibility definitely comes from the controls that we are putting on the finding," she said, "and not necessarily trusting the output itself." When using Claude, you can build those instructions and controls into each Project (and associated Skill) or, if they truly relate to all of your work, within Claude Global Instructions.

She also made a point that goes beyond diligence: don't just check whether the findings are right, check for what isn't there at all. "You should be looking for what's absent," she said, because it's easy to get so focused on grading the answers in front of you that you miss the question nobody asked.

And when something did come back wrong, she was clear that it usually wasn't a hallucination. "They're just structural gaps," she said, the result of the tool not having the context that she, as the lawyer, carried in her head. The tool handled the volume so she could spend her time on the risk assessment

Decide Where the AI Doesn't Go

Laura Jeffords Greenberg, General Counsel at workforce platform WorkSome, is building what she calls an AI-native legal team with a headcount of three. She showed two builds, a legal command center and an automated workflow that answers sales and customer-success questions with legally-approved language. Both run on a principle she stated plainly: a human should never have to answer the same question twice.

What I want to flag from her builds is the opposite instinct from the one we usually celebrate. Much of the excitement about AI tools is about what they can develop and generate on their own. Here, Laura spent her energy deciding where she didn't want them to be creative at all.

For the workflow that responds to go-to-market team queries, she told the tool to pull her approved language word-for-word. "I actually don't want you to use your creativity to rewrite something," she said. "I want you to use this language verbatim." When she had let the tool paraphrase, it started drifting from the wording she'd approved. In a context constrained by employment law across jurisdictions, this drift presented real risk. Knowing where AI can't go, as she put it, was as important to the design as knowing what it could do.

That extends to verification. Laura monitors the channels where her AI tools answer legal and compliance questions. Responses are labeled so everyone knows what came from the tool and what she's personally verified, and her team is trained to question an answer that looks off. The automation does the repetitive retrieval, and the judgment, along with the accountability for it, stays with her.

As she described it, the knowledge base where her approved positions live is the centerpiece of the whole thing, and filling it is the part only she can do.

Where This Leaves You

The pauses, the confidence checks, the rule about pulling language word-for-word, the lawyer who still hits send at the end: none of that should be considered a tax you pay to use AI on serious work. It's the thing that makes the AI output usable.

And look at what those controls actually are. Deciding which battles are worth fighting, spotting what's missing rather than only what's wrong, tying a claim back to its source, refusing to send a draft you haven't read. This is the judgment you already use every day, now aimed at a faster set of tools.

This post draws on a live practitioner session from The L Suite's Claude for Legal: In-House Edition webinar (Part 2), featuring Natalie Kim (Strategic Advisor, Anthropic), Tamara Momirov (SVP Legal and Privacy, Tilting Point), and Laura Jeffords-Greenberg (General Counsel, WorkSome), moderated by Tommy Taveras-Verreira (Chief Strategy Officer, Law Trades).