Does AI really save time for lawyers and legal counsel?

The short answer

Yes, for some tasks. Much less than the vendor numbers suggest. And the time saved rarely turns into free time.

Across practitioner blogs, controlled studies and interviews with in-house teams, one pattern keeps coming back. AI pays off when checking its output is much faster than producing it yourself. When checking means redoing the work, the gain disappears.

What separates the lawyers who get hours back from the rest is rarely the tool. It is whether the method behind the task has been written down.

Where AI genuinely saves time

The best controlled evidence is a randomized trial led by Daniel Schwarcz. Upper-level law students completed six legal tasks with a legal research tool, a reasoning model or no AI. AI raised productivity in five of the six tasks, by roughly 34% to 140% depending on the tool, and quality improved too Schwarcz et al. · AI-Powered Lawyering, 2026 (opens in a new tab). Two caveats: the participants were students, and the tools date from 2024.

In practice, the gains cluster around a few kinds of work:

• First drafts of documents you already know how to write. You recognise a bad clause on sight.
• Summarising and extracting from documents you will check against anyway.
• Reviewing against a defined playbook. The standard is written down, so deviations are easy to confirm.
• Polishing finished work. Even one of the strongest skeptics we read accepts AI as a final-review redline tool.

The common thread: you can judge the output quickly, because you already know what good looks like.

Where it doesn't: the verification cost

The sharpest critique comes from academic Joshua Yuvaraj, who calls it the verification-value paradox arXiv · The Verification-Value Paradox, Oct 2025 (opens in a new tab). Because lawyers answer for every word, he argues, the net value of AI is often close to zero.

Critics reply that lawyers already verify everything, from junior memos to expert reports. Both sides are partly right. It depends on the task.

Research in an unfamiliar area is where the paradox bites hardest. Peter Winders explains why his firm still bans generative AI for research and briefs Carlton Fields · Why our firm still prohibits generative AI, Mar 2026 (opens in a new tab). Verifying properly means doing the traditional research anyway, so it takes more time, not less. Removing fake citations does not leave a competent brief behind.

Skipping verification is not an option either. Damien Charlotin's public database of court decisions involving AI hallucinations grew from about 200 cases in mid-2025 to 1,598 by June 2026 Haqq · AI hallucination audit, Jun 2026 (opens in a new tab). And the hardest errors to catch are not invented cases. They are real cases with a slightly wrong holding or quotation, as lesson 2.1 shows.

Why the saved time rarely shows up

Even where AI works, most lawyers do not get a shorter week. In a 2026 poll of 85 UK in-house lawyers, 93% said they use AI in their legal work. Only 38% said they work fewer hours since adopting it Nomio · Ready or Not?, Jul 2026 (opens in a new tab).

The interviews behind that poll explain why. One group general counsel estimates AI removed 10% to 15% of her team's low-value work, yet she works as hard as before. Another lawyer put it in five words: “Work just fills up the void.”

Research on knowledge workers points the same way. AI speeds up tasks, which raises expectations, which expands scope. Because prompting feels like chatting, work slips into evenings without registering as effort Thomson Reuters Institute · AI makes lawyers work more, not less (opens in a new tab).

In law firms there is a further catch. Under the billable hour, a lawyer cannot bill time not spent, so efficiency becomes a pricing problem rather than a gift of time.

The in-house twist

Legal counsel face a cost law firms rarely see: AI output arriving from everyone else, from executives to counterparties.

In the same interviews, a group GC described executives sending her contracts drafted in ChatGPT that were badly inadequate. Another GC tested Copilot on a drafting task. The result looked impressive until she asked it to cross-check, and it admitted errors. On a hundred-page wind turbine supply contract, one team found the AI's risk comparison useful but its advice crude: it recommended rejecting nearly everything.

Counterparties now send AI-generated terms too. That leaves counsel with a new question: is it worth their time to fix the other side's draft?

So the hours saved on your own drafting can be spent correcting someone else's AI.

What separates those who get time back

The lawyers who do get hours back have one thing in common: they wrote their method down.

Open-source lawyer Luis Villa makes the point clearly Luis Villa · Of Monsters, Men, and Lawyers, Mar 2026 (opens in a new tab). Ask a general-purpose model to simply do a legal task and it often fails. Give it detailed definitions and instructions first, and it often works very well. The catch is cost: for a task you rarely do, writing those instructions takes longer than doing the work.

The in-house evidence matches. The one interviewee who cut his hours had built his own agents for specific tasks, and avoided a hire. Most others never found the time to build anything.

The gain also grows with practice. Early on, you use AI task by task and check everything by hand. With experience, you learn which methods generalise and which checks can be automated, and the share of work that qualifies grows.

Two honest limits. A borrowed method is a starting point: your positions and risk appetite still have to go in. And structure does not remove verification. Automated checks can catch mechanical errors: a quotation that is not in the source, a citation that does not exist, a rule the agent skipped. They cannot catch errors of judgment, such as a real case cited for a point it does not support. Automation makes the review shorter, but the responsibility stays with you. On open-ended research, where most of the work is judgment, we think the skeptics are mostly right.

Measure it, or you are guessing

Almost nobody measures whether AI actually works on their legal tasks. The headline numbers are mostly self-reported: vendor surveys of their own customers, or polls asking lawyers how they feel. Few teams time the review, and fewer still check the output against a standard.

That gap matters, because every question in this article is a measurement question. Does it save time once review is counted? How often is it wrong, and on which points? Did the last change to the method make it better or worse?

Three measures answer most of it:

• Time end to end, including the lawyer's review, not just the time to a first draft.
• Accuracy against a reference: the output scored point by point against work your team has already approved.
• Errors by type: which checks get missed, which rules are misread, which sources go unused.

Verification becomes measurable once you have that reference. Instead of a vague sense that the tool is pretty good, you know where it fails. And once you know where it fails, you can optimise: change the method, run it again on the same cases, and keep the change only if the score goes up.

We will go deeper into metrics, verification and optimisation in the next chapters. Here is what it looks like in practice.

An example from our own work

A company's legal team analysed tender documents with an internal method that existed only in its jurists' heads. We wrote it down as a plugin: a set of skills for each part of the analysis, with the points to check, the team's positions and its house format, as described in lesson 1.3.

We added the references that the jurists use: external documents, past cases and case law. Then we built a dataset from the team's past analyses, each one approved by the team. This dataset became the standard for the plugin.

Then the optimisation started. We ran each version of the skills on the full dataset and scored every result against the team's own analysis, point by point. Each error showed us a gap in the method: a missing check, an unclear rule, a reference that the agent did not use. Our optimisation methods rewrote the skills to close these gaps, ran them again on the full dataset, and kept a change only if the score went up. Version after version, the method became more precise, and each improvement was measured, not guessed.

The workflow runs in our harness, which we optimise for legal tasks. We also built an interface for the jurists. In it, they start an analysis, work with the AI agent, review each result, export it, and return to correct and run it again. At the end, the workflow cut the time for each analysis by 70%, review by a jurist included. The gain did not come from a better model. It came from a written method, a dataset to measure it, and a workflow built around the review.

Today, this work needs our team. We are now building general methods to make the same steps easy for every team on Lawve: write down your method, measure it against your own past work, and improve it, without technical knowledge.

Bottom line

AI saves lawyers time when three conditions hold: the task repeats, the method is written down and measured, and you can check the result faster than you could produce it. Without them, it mostly moves work around.

You do not need a vendor study to find out which side your work falls on. Run this test:

1. Pick one task you do every week.
2. Time it by hand three times.
3. Time it three times with AI and a written procedure, review included.
4. Note every error you had to catch along the way.

If the gap survives the review time, you have found real time. If not, you have found where your judgment is the work.