The new game of legal work with AI
AI commoditizes the mechanical part of legal work, but what gains value is the judgment and the accountability of whoever signs off. That's why, in this module, auditing every citation and every fact is non-negotiable.
You throw a 60-page contract at AI and ask for an opinion on the risk of the termination clause. In 30 seconds back comes flawless text, with the tone of someone certain, citing a legal article and a precedent. You almost sign off on it. Except that precedent doesn't exist, and the cited article talks about something else.
You ask AI to summarize the résumés for the senior opening and point out the three strongest candidates to call in for interviews. It hands back a tidy list, a score for each one, and the line "so-and-so has 8 years of experience in the field," confident, ready for you to pass along to the hiring manager. You almost send the list without opening anything else. The candidate's actual résumé shows 5 years, AI rounded the number up because it seemed reasonable for the role, and nobody compared the summary against the original document before deciding who moves to the next round. It's the same risk from legal knocking on HR's door: AI describes things with the same confidence whether it's right or making it up, and you can't tell the difference just by reading the summary. Any data point that becomes a decision about who to interview needs to be checked against the source résumé, not just the pretty summary AI handed you.
You throw the user research spreadsheet at AI and ask for the most common pain points, to decide what moves up on the quarter's roadmap. It hands back "73% of users ask for calendar integration," clean, with the look of closed data, ready to become the justification in the PRD. You almost prioritize the feature on top of that number. That percentage isn't anywhere in the original research, AI filled the gap with a number that sounds plausible for this kind of product, and nobody went back to the raw responses to check. It's the same trap as a legal opinion: invented data looks exactly like true data, and only someone who checks the source notices the difference. Before any research statistic becomes a priority line on the roadmap, it needs to match the source spreadsheet, response by response.
You ask AI to put together the sales proposal and cite the case studies that prove results for that prospect. It writes "client X grew 40% in three months with us," convincing, tidy, a great argument for closing. You almost send it straight to the client. That number doesn't exist in your CRM, AI invented the case study to make the proposal stronger, and nobody opened the account's real history before the proposal went out. It's the same confidently-worded hallucination that brings down a legal opinion: the case study sounds true until the prospect asks for the reference, and then whoever signed the proposal answers for a number that doesn't exist. Every case study cited in a proposal needs to come from your real CRM, checked before the proposal leaves your desk.
You ask AI to analyze the month's indicators and say whether the delivery SLA hit target. It answers "SLA met on 98% of orders," tidy, reassuring, ready to go into the report for the COO. You almost close the report just like that. The real percentage, in the source dashboard, is 91%, AI rounded it up on its own because the number seemed reasonable for a good month, and nobody went back to the source spreadsheet before reporting. It's the same risk that makes a legal opinion collapse under audit: a convincing number is no guarantee of a true number. Every indicator that goes into the SLA report needs to be checked against the source dashboard before it becomes an official number for the committee.
You ask AI to map the new process's risks and say whether it complies with LGPD, citing the article that backs the analysis. It comes back with "compliant with Article 7, no relevant concerns," confident, with the look of a closed opinion, ready to go to the risk committee. You almost attach it straight to the controls report. Article 7 covers a different legal basis for processing, not the one in your process, AI cited it with the same confidence it would use to cite it correctly, and nobody opened the text of the law to check before reporting. It's the same legal hallucination that gets lawyers sanctioned for citing invented precedent: the wrong citation carries the same confident tone as the right one. Every article cited in a compliance opinion needs to be opened and checked against the law, not just repeated because AI said it existed.
You ask AI to review the authentication snippet before it ships to production and say whether it's secure. It answers "all good, the function validates the token correctly before granting access," confident, even naming the function that does the protecting. You almost approve the deploy on that answer alone. That function doesn't exist in your repository, AI described a common authentication pattern as if it were your actual code, and nobody opened the source file to check before the merge. It's the same mistake that brings down a legal opinion citing a law that doesn't exist: the technical answer sounds right until someone opens the real file. Every claim about your code needs to be checked line by line in the repository before any deploy.
You throw the research sessions at AI and ask where users get stuck in the checkout flow, to justify the redesign to the product team. It hands back "abandonment happens on the payment screen because of the coupon field," tidy, with a user quote that sounds real. You almost take it straight to the prioritization meeting. That quote doesn't appear in any of the recordings, AI filled the gap with a typical checkout pain point, and nobody went back to the original transcripts before presenting the insight. It's the same trap as a legal opinion anchored in phantom case law: the cited insight looks exactly like a real insight. Any user pain point that becomes justification for a redesign needs to be checked against the original transcript, sentence by sentence, before it goes into the meeting.
You ask AI to put together the market sizing slide for the board meeting, with the TAM for the new market and the source backing the number. In seconds it hands back a tidy slide, "the TAM is 2.3 billion dollars," citing a heavyweight report right next to it, and everything looks like closed, impressive data. You almost take that straight to the board. The cited report doesn't exist, AI invented the source with the same confidence it would use to cite a real one, and nobody on the team checked the number against real research before locking the deck. It's the same mistake that makes a legal opinion collapse the first time the other side checks it: a pretty citation is no guarantee of a true citation. Before any market number becomes a board slide, it needs a source you can actually open and check, not one that just sounds plausible.
Whoa, notice something: the danger of AI in legal work isn't that it's dumb. It's that it's convincing. It writes with the confidence of a senior partner, cites law, cites case numbers, and sometimes it's inventing all of it with the look of absolute truth. In almost every other field, an invention like that costs some rework. In legal work, it costs a lost case, a fine, a signature you can't take back. Think with me: what changes in the game isn't AI doing the lawyer's job, it's AI driving down the price of the mechanical part and throwing all the value onto the side it doesn't do.
The core idea of this lesson. AI commoditizes mechanical legal work (the first read of a contract, the draft of a filing, searching through a mountain of documents), and that gets cheap and fast. What gets more expensive is what it doesn't deliver: judgment (the real risk of a clause, strategy, interpretation) and accountability (whoever signs the opinion answers for it, always). And there's a deadly catch: AI invents law, case law, and precedent with the look of certainty. That's why this module's rule is strict: every citation and every fact go through audit before becoming a deliverable.
01What AI commoditizes in legal work
Commoditizing is what happens when something that used to be expensive and slow becomes cheap and instant. That's what AI did to the mechanical part of legal work, and it's better to face that head-on than pretend nothing changed.
Think about the tasks that used to eat hours of a junior lawyer's time. The first read of a big contract, to understand what it's about. The draft of a standard filing, one of those that barely changes from case to case. The search for a specific passage in the middle of hundreds of documents in a case file. AI does all of that in minutes, for a cost that trends to zero.
This isn't a threat, it's leverage. The economic frame is simple: what becomes a commodity loses price. Whoever used to bill by the hour for mechanical reading watches that price collapse, the same way Harvard shows AI driving down the price of mechanical expertise in general. The right question isn't "how do I protect this work," it's "where did the value that left here go." Fair?
02What gets more expensive: judgment and accountability
If the mechanical side became a commodity, the value didn't evaporate, it migrated. It migrated to two things AI doesn't deliver and that get more expensive precisely because they've become rarer to do well.
The first is judgment. AI reads the termination clause and tells you what's written. It doesn't tell you, with accountability, what the real risk of that clause is for your client, in that sector, at that moment in the negotiation. That's the work of interpreting, of weighing strategy, of seeing what the clause triggers down the road. It's the kind of thing that's worth more once mechanical reading becomes free.
The second is accountability, and here there's no middle ground. Whoever signs the opinion answers for it. Always. AI has no bar registration, no name on the filing, doesn't sit in the hearing. If the opinion is wrong, the bill goes to the human who signed it, not to the model. That means AI can produce, but it can't be the last to speak.
Think of it as a brilliant, fast intern, who delivers quickly and sometimes makes things up. You use their work, and you're still the one who signs and answers for it. The intern doesn't become a partner just because they type fast. Fair?
03The deadly catch: legal hallucination
Now the part that makes legal work different from almost every other domain. AI hallucinates. It invents information with the look of absolute certainty, and in law that's especially poisonous.
It invents a law article that doesn't exist, or cites one that exists but talks about something else. It invents case law, with a court name and case number, tidy, plausible, false. It invents precedent that was never ruled on. And it does all of this in the same confident tone it uses when it's right, so you can't tell truth from invention by reading alone.
Why this destroys a case rather than just being an inconvenience. Because in legal work the citation is the base of the argument. An opinion anchored in phantom case law falls apart at the first check by the other side or the judge. Lawyers have already been sanctioned for filing a brief with a precedent invented by AI. The cost of the error here isn't rewriting a text, it's losing credibility, losing the case, and sometimes answering for it.
04The golden rule: audit before delivering
Out of everything we've seen comes a single practical rule, and it's worth the whole module: in legal work, every AI output goes through audit before becoming a deliverable. No exceptions.
Auditing here means something concrete. Every citation of law, checked against the text of the law. Every case law reference, checked against the official source, case by case. Every fact stated about the client's document, checked against the document. If AI says clause 12 says X, you open clause 12 and confirm it says X. What can't be checked doesn't go into the deliverable.
The economic frame closes the reasoning: AI drives down the cost of producing the draft, but the cost of auditing is the new expensive work, and it's what protects the signature. You're not paying for the typing, you're paying for the check that guarantees the piece is true. Whoever skips the audit is outsourcing their own accountability to a model that answers for nothing.
Think of a bridge: AI delivers the beams ready and fast, but it's the engineer who signs the load report before opening it to traffic. Nobody crosses on the factory's word alone. Fair?
05The map of the Legal module
This lesson is the gateway. From here on, the module goes deep into each concrete piece of legal work with AI, always under the same golden rule.
You'll learn to connect your documents to AI while respecting confidentiality, because legal material is confidential and can't leak outside. You'll see contract review, using AI for the first read and your judgment for the risk. You'll build the anchored opinion, where every citation has a checkable source. You'll use AI in due diligence, sweeping through lots of documents without blindly trusting the summary. You'll generate drafts faster, keeping control of what's standard and what's specific. And you'll structure the audit, the step that turns a draft into a safe deliverable, all the way to designing the OS, the legal operating system that ties all of this into one flow.
Do it now
Pick a real, mechanical legal task from your day that you'd be willing to delegate to AI: your real task.
- Classify: what in this task is mechanical (first read, draft, search) and what is judgment (real risk, strategy, interpretation)? List in two columns.
- Point to the signature: who answers if the final deliverable is wrong? Write the responsible human's name.
- List what would need to be audited before becoming a deliverable: every law cited, every case law reference, every fact stated about the client's document.
- Define the gatekeeper: write in one sentence the audit rule you're going to apply before any AI output leaves your hands.
You've just separated what AI commoditizes from what stays yours, and designed the audit gate that protects your signature.
Practice
1. In the new game of legal work with AI, what best describes what AI does and what gets more expensive?
2. Why is legal hallucination especially dangerous compared to AI errors in other contexts?
3. What is this module's golden rule for using AI in legal work safely?
For the board
On the dangerit is not that it is stupid, it is that it is convincing. It writes with the assurance of a senior partner and sometimes invents the whole thing.
On what got expensivethe price falls on mechanical execution. The value moves to interpreting the risk and answering for the delivery.
On the citationit is the foundation of the argument. Anchoring on a phantom precedent sinks the filing and can bring sanctions.
Thanks for the feedback. It helps sharpen the next lesson.