How do I get started with AI?

Legal

Speed up first-pass contract reviews, flag high-risk terms automatically, and keep outside legal spend down.

Updated 4 days ago · 9 sources

Proof over hype

  • Mid-market

    DarrowEverett

    A $30 million divorce case turned on one question: was the $8.7 million the other side called separate property actually part of the shared marital estate?

    More than $4 million in additional value for the client, after opposing counsel conceded on seeing the traced records.

    The takeaway The win wasn't drafting speed. It was five years of transaction records traced by a machine in minutes instead of by an associate over days — the unglamorous, document-heavy work that decides cases and that nobody wants to bill hours to.

    Reported by Harvey · checked Feb 2026

    How they did it

    DarrowEverett's lawyers uploaded five years of the couple's financial records into Harvey and asked it to trace commingling — the mixing of separate money into joint accounts, which is what turns separate property into shared marital property under the law. Harvey worked through five years of statements, wires and joint tax filings and returned a source-linked analysis: every dollar it traced, with a citation to the exact statement or transaction it came from. The trail accounted for $2.2 million that opposing counsel had claimed was untouchable separate property. DarrowEverett's partners put the traced records in front of opposing counsel before the hearing rather than saving them for cross-examination. Opposing counsel conceded the point rather than contest a paper trail with citations attached to every line, which reopened the rest of the estate calculation and added more than $4 million in value for the client.

    Tool used Harvey

  • Public sector

    Los Angeles County Superior Court

    Six civil judges were drowning in paper — hundreds of pages of motions and declarations to read before writing a single ruling, in a court system already straining under a wave of self-represented litigants filing their own AI-drafted paperwork.

    The takeaway The safeguard is the story, not the software: adoption was conditioned on a judge reading and editing every single draft. That's the bar for any pilot where a wrong answer has real consequences, not just an inconvenient one.

    Governing · Mar 2026

    How they did it

    The court signed a roughly year-long, $300,000 contract with Learned Hand, a company built specifically for judicial work that was already running in trial courts across ten states. Its founder calls it a "judicial sous chef" — meant to prep the ingredients, not decide the menu. Six Los Angeles civil judges got access starting in February 2026, pointed at the exact filings that had been piling up: summary judgment motions and class-action settlement approvals. The tool reads the filings, produces a summary, and drafts a tentative ruling — a preliminary decision written to match that judge's own past writing style — with every sentence run through a fact-check step called Deep Verify, which checks each claim against the case law it cites and links back to the source so a judge can click through rather than take it on faith. The one hard rule the court built in: a judge has to personally read and edit every draft before adopting it as an actual ruling. One participating judge, speaking anonymously, added a caution worth keeping: even an AI suggestion he ultimately rejected became a reference point that shaped the ruling he wrote instead — a reminder that a draft is never a neutral starting point, rejected or not.

    Tool used Learned Hand

  • Enterprise

    Linklaters

    The firm was already running Harvey, Legora and Copilot firmwide, but its messiest, highest-value client data sets — the ones a generic tool chokes on — had nowhere to go.

    The takeaway If your team keeps hitting a wall with off-the-shelf AI on its messiest, highest-value work, the fix other firms are reaching for isn't a better tool. It's putting a technologist in the room with the people doing the work, on a standing basis.

    Legal IT Insider · May 2026

    How they did it

    Linklaters set up a standing unit called Applied Intelligence that pairs lawyers directly with data scientists, rather than routing requests through an IT department. Its job is narrow on purpose: build one-off AI tools for the specific matters where an off-the-shelf product can't do the job, working across several underlying AI models instead of committing to one vendor. Co-founder Tom Quoroll said the team was built "not just for technical capability, but for the collaboration and judgment needed to turn complexity into impact" — the pairing of lawyer and data scientist in the same room is the point, not a detail. It runs alongside the firm's existing Harvey, Legora and Copilot deployments rather than replacing them: those handle the daily, repeatable work, and Applied Intelligence exists for the matters that don't fit a template.

  • Learn from thisSmall business

    Mike Singh Sethi and William Rounds, immigration lawyers

    An immigration appeal to the Ninth Circuit needed real case law behind it. What the court got instead was fiction.

    $2,500 fine and a six-month suspension from practicing before the Ninth Circuit for each attorney, referral to the State Bar of California, and a two-year court-ordered rule requiring every filing from their firm to carry a signed statement, under penalty of perjury, naming any AI tool used and certifying the attorney personally checked every citation.

    So: Put a named person on record checking every citation before filing — the exact discipline the court itself now requires of this firm, in writing, on every brief, for two years.

    The takeaway The lawyers who signed this brief never read the cases it rested on. If nobody at your firm is required to personally verify every citation an AI tool hands them, you are one motion away from this exact story.

    The Volokh Conspiracy, Reason · Jun 2026

    How they did it

    The opening brief in the appeal cited two cases that do not exist, "Eduardo v. Garland" and "Lay v. Holder," and attached quotations to two real cases, Kamalthas v. INS and Avendano-Hernandez v. Lynch, that neither decision actually contains. An unlicensed writer had drafted the brief using a generative AI tool, and neither Sethi nor Rounds — the licensed attorneys who signed it — had personally read the cases before it was filed. At oral argument, the panel asked Rounds directly whether AI had produced the errors; he denied it more than once before eventually conceding it was "possible." Sethi and Rounds had earlier told the court the fake citations were typos. The Ninth Circuit's opinion draws a sharp line: it did not sanction the lawyers for using AI. It sanctioned them for letting fabricated authority reach a federal appeals court and then calling it a typo instead of owning it.

  • Learn from thisEnterprise

    Deloitte Australia

    Australia's Department of Employment and Workplace Relations paid Deloitte roughly A$440,000 for an independent review of the IT system that decides who gets flagged and fined under the country's welfare compliance rules.

    Refunded roughly A$97,000 of the A$440,000 contract, and added an AI-use disclosure to the corrected report that the original never had.

    So: Ask every adviser in writing whether generative AI touched the report, name the tool, and name the person who checked every citation and quote against the original source before it shipped.

    The takeaway This wasn't a startup cutting corners — it was Deloitte, one of the most process-heavy firms in the world. If their review didn't catch fabricated citations before publishing, yours won't either without a named human checking each one.

    CFO Dive · Oct 2025

    How they did it

    Deloitte used generative AI, later confirmed to be Microsoft's Azure OpenAI service running GPT-4o, to help produce the review. University of Sydney health and welfare law researcher Chris Rudge read the published report and found the consultants had invented the references: nonexistent academic papers attributed to real researchers, a quote attributed to a Federal Court judgment that was never written, and even a misspelled name for the judge who supposedly wrote it. Once the errors became public, Deloitte published a corrected version months later, in October 2025, that added a disclosure stating generative AI had been used in producing the report — a disclosure the original version never carried — while telling reporters that "the substance" of the review's findings and recommendations remained unchanged.

Try this today

  • Turn a contract into a dated obligations list

    15 min

    You see every deadline, notice period and renewal date in one table, not buried in pages.

    Copy the prompt

    You are a contracts analyst. Below is one agreement. Build a table with these columns: obligation, who owes it, trigger or date, section number, exact quoted wording. Cover payment terms, notice periods, renewal and termination dates, service levels, audit rights, and any reporting duty. Use only text that appears in the document. Quote the wording word for word. Do not paraphrase, and do not add anything that is not written there. If a date depends on another event, name that event. If a section is unclear, list it separately under CHECK THIS rather than guessing. Agreement: [PASTE THE FULL TEXT OF THE CONTRACT]

    One check first. A first pass for a lawyer, never a citation or a legal conclusion. Check every quote yourself. Keep privileged files out of consumer chatbots.

  • Compare a supplier draft against your standard positions

    20 min

    You get a clause-by-clause gap list before the call, so you argue about the three that matter.

    Copy the prompt

    You are a contracts analyst. First I paste our standard position on each clause. Then I paste a supplier's draft. Work through our positions in order. For each one: name the clause, quote the supplier's exact wording, mark it BETTER, WORSE or SAME against our position, then give one sentence a non-lawyer could act on. If the draft says nothing on that point, write MISSING and move on. Do not guess at wording, do not summarise anything I have not pasted, and do not draft replacement language. End with the three worst gaps, ranked, and say why each one is worst. Our positions: [PASTE YOUR STANDARD POSITIONS] Supplier draft: [PASTE THE DRAFT]

    One check first. Treat the ranking as a first pass for a lawyer, not a citation or a legal conclusion. Never paste privileged drafts into a consumer chatbot.

  • Rewrite a legal position in plain English

    10 min

    Your business team understands the answer on first read, so the same question stops coming back.

    Copy the prompt

    You are a translator between lawyers and business teams. Below is a legal position I have already decided. Rewrite it for [ROLE, FOR EXAMPLE A SALES DIRECTOR] in under 200 words. Use these headings: What you can do. What you cannot do. What to do if a customer pushes back. Who to ask. Write the way a national newspaper writes. No Latin, no clause numbers, and no hedging words such as generally or typically. Keep my decision exactly as I wrote it. Do not soften it, widen it, or add exceptions I did not give you. If my position leaves a question open, list it at the end under OPEN. My position: [PASTE YOUR POSITION]

    One check first. A first pass for a lawyer, not a citation or a legal conclusion. Your position may be privileged, so keep it off consumer chatbots.

Specialized tools, and what to ask vendors

Specialized tools for this function
ToolWhat it doesSetupBest fit
HarveyenterpriseAnswers legal research questions and drafts documents against your own files, with links back to the source it used.MonthsIT sign-offYou run repeat, document-heavy work like due diligence (checking contracts and records before a deal).Skip it ifSkip it if your team is five people. Seat prices and setup only pay back across hundreds of lawyers.
Spellbookper seat, small team to enterpriseHandles contract redlining (marking up suggested changes) inside Microsoft Word, against the positions you normally accept.Under a weekIT sign-offYour team lives in Word and reviews a steady stream of routine commercial contracts.Skip it ifBespoke, high-stakes drafting where every clause gets negotiated. You will fight the suggestions.
Luminancemid-market to enterpriseReads a large pile of contracts and flags the odd ones: missing clauses, unusual terms, obligations you forgot.WeeksIT sign-offA deal, an audit, or a new rule forces you to read thousands of old agreements.Skip it ifYou have a few hundred contracts and a decent filing system. Read them.
Ironcladmid-market to enterpriseRuns your contract process end to end: request, drafting, approval, signature and storage, with AI review built in.MonthsIT sign-offContract volume is growing, your team is not, and the business complains about how long you take.Skip it ifYour templates are a mess. A CLM (contract lifecycle management system) will only speed up the mess.

Questions to ask before you buy

Harvey — 7 questions to ask them

Also used by. DLA Piper, Slaughter and May

  1. When your tool gives me a citation, is it checked against a real case database before I see it?
  2. Has any court tested whether using you affects privilege (the protection that keeps lawyer-client advice confidential)?
  3. Where are my documents processed and stored, and do they ever leave my country?
  4. Is my client work used to train your models or anyone else's, and will you put that in the contract?
  5. What happens to my documents and my query history on the day we terminate?
  6. Which customer is closest to my size and practice mix, and can I call them without your team listening?
  7. Show me a firm that ran a pilot and did not buy. What went wrong there?
Luminance — 6 questions to ask them
  1. What accuracy do you get on my document types, and who measured it, you or a customer?
  2. Where are the documents processed, and can everything stay inside my own environment?
  3. When the system is unsure about a clause, what does a human see, and when?
  4. What happens to the document set after the review ends, and can you show me it was deleted?
  5. How do you keep privileged material separate from everything else in a review?
  6. Which customer ran a review my size, and what did the tool miss on that job?
Ironclad — 6 questions to ask them

Also used by. Dropbox, Rivian, Zoom

  1. How much template and playbook (your standard position on each clause) clean-up do we finish first?
  2. What does the export look like on termination: every contract, every version, every approval?
  3. Which AI features are generally available today, and which are still in early access?
  4. How does the system decide a clause is off-policy, and can my team see and change that rule?
  5. What is the total first-year cost including setup, and who on my team is needed full time?
  6. Which customers moved off you, and what did they say broke?

What everyone is asking

  • NewJul 2026

    Best AI for contract review right now?

    That depends on where your team works. Spellbook and tools like it sit inside Word for routine commercial contracts. Ironclad runs the whole contract process. Luminance is for large piles of old agreements.

    What to watch for

    Almost every public comparison is written by a vendor. Ask two customers of each to run the same ten contracts and tell you what got missed.

    Also worth a look. Spellbook, Ironclad, Luminance, Harvey

  • NewJul 2026

    Is it malpractice to use AI for legal research?

    Courts have not banned AI research. They have punished lawyers who filed citations they never read. The heaviest orders went to the lawyers who then hid what happened.

    What to watch for

    Court guidance now says one thing over and over. Read the case before you cite it. Nobody gets excused because a tool wrote the brief.

    Also worth a look. Check every citation in a primary source, Ban AI research outright, Require a written AI disclosure on every filing

  • NewJul 2026

    Does using AI waive attorney-client privilege (your right to keep legal advice confidential)?

    Courts disagree so far. Some have treated AI tools as tools rather than third parties. One court found privilege waived where a client used a public tool without a lawyer directing it.

    What to watch for

    The rule your team can follow today: privileged material only goes into a tool you have a signed contract with.

    Also worth a look. Enterprise tool under contract, no training on your data, Public consumer chatbot, No AI on privileged material at all

  • NewJul 2026

    Will AI actually make my legal team faster?

    Drafting speeds up and checking eats some of that back. Teams that measure tool usage, rather than how long a job takes end to end, cannot answer this.

    What to watch for

    Lawyers are more doubtful than vendors. The usual complaint is that checking time simply replaces drafting time.

    Also worth a look. Measure how long a piece of work takes end to end, Measure tool usage, Measure outside counsel spend

Worth following

  • Artificial Lawyer

    Legal technology news site covering AI in law firms and legal departments.

    daily · publication

    Why them

    First to report most legal AI deals, and openly sceptical when the claims get loud.

  • LawSites — Bob Ambrogi

    Robert Ambrogi, a lawyer and journalist who has covered legal technology for two decades.

    several times a week · blog

    Why them

    He tests products, reads the court orders, and writes plainly about what actually shipped.

  • Legal IT Insider

    Trade publication covering legal technology buying, vendors, and law firm IT.

    daily · publication

    Why them

    Strong on what firms actually bought, including the deals nobody announces loudly.

  • Legal Evolution — Bill Henderson

    A law professor's education project on how the legal services market is changing.

    monthly · publication

    Why them

    Long, careful essays. Read it when you need to think, not when you need news.

  • Legaltech Hub

    A directory and research site listing legal technology products by what they do.

    weekly · directory

    Why them

    Worth reading before a buying cycle, so you know who exists beyond the loudest three.

  • The Volokh Conspiracy

    A law professor blog hosted at Reason, running since 2002.

    daily · blog

    Why them

    It posts the AI sanctions orders fast, with the judge's actual words attached.