It is late on a Thursday and the opposition brief is due Friday at noon. The associate asks the firm's research assistant for authority on a narrow procedural point, and it comes back with exactly what she needed: a decision from the right circuit, four years old, with a paragraph that reads as if it had been written for her motion. She drops the quote into section three. The partner reads the draft over breakfast, likes the argument, and signs. At ten the next morning a paralegal doing a last cite check goes to pull the opinion and cannot find it. Not in the reporter, not on the court's docket, not anywhere. The case has a caption, a citation, a judge, and a holding. Nobody ever decided it.
That firm got lucky, because one person still did the boring part by hand. Plenty of others have not. Since 2023, courts in the US and abroad have sanctioned lawyers for filing briefs built on invented authority, and the list has grown every year since. Those stories usually get told as a warning about the technology. I read them as something more useful for anyone running a firm: a clear picture of where the value in legal AI sits, and where it does not.
Cheap words, expensive signatures
A law firm does not sell paragraphs. It sells a licensed person's willingness to stand behind them. Every brief, opinion letter, diligence memo, and contract markup carries an implied promise that someone qualified read the authority, understood the client's position, and is accountable if it is wrong. Clients pay for that promise. The words are how it gets delivered.
Generative AI pushes the cost of producing those words toward zero. A first draft of a research memo that took six hours can take forty minutes. What it does not touch is the cost of the promise. The partner who signs still has to know the cases exist, still has to know they say what the draft says they say, still has to know they have not been overruled. If anything the work of vouching gets harder, because the draft now arrives fluent and confident whether or not anything behind it is real.
So the firms measuring their AI programs by drafting hours saved are tracking the cheap half of the business. The number that matters is how much AI-assisted work the firm can put its name on without adding risk, and that is set almost entirely by the controls around the tool. The economics push the same way. ABA Formal Opinion 512, issued in 2024, said plainly that a lawyer billing hourly cannot bill for time the tool saved, and generally cannot pass along the time spent learning it. When six hours of research becomes forty minutes, the firm either reprices the work or loses the margin. The firms that come out ahead will be pricing judgment. Judgment is only worth a premium if clients believe someone exercised it.
Competence now includes knowing how the tool fails
The duty of competence has, for more than a decade, included keeping up with the benefits and risks of relevant technology. With generative AI that duty gets specific. A lawyer does not need to know how a language model works inside. She does need to know how it fails: that it will produce a citation with the same confidence whether the case exists or not, that a summary can drop the one exception that decides the issue, and that it has no idea which jurisdiction's rule governs unless someone tells it and checks.
For a firm, competence turns into a set of operating decisions. Which tasks can the tool draft, and which ones should it only search? How much verification does each kind of output need before it reaches a client or a court? Who trains the second-year associates, who are now the heaviest users and the least able to spot a plausible error? A firm that rolls out a research assistant with a lunch-and-learn and a usage policy nobody reads has handed its least experienced people its most convincing source of mistakes.
Privilege does not survive a careless prompt
Everything a firm does sits on confidentiality. Client facts, deal terms, litigation strategy, and the substance of privileged communications are the raw material of legal work, and they are exactly what people paste into an AI tool to get a useful answer. Where that text goes next is a professional-responsibility question, not an IT one. A tool that trains on inputs, keeps them longer than the matter requires, or lets a vendor's staff read them can put privilege at risk and the client relationship along with it. For some uses, the ABA's guidance points toward getting the client's informed consent before their information goes into the tool at all.
Inside the firm, the harder problem is the ethical wall. Firms spend real effort keeping one matter's information away from lawyers working the other side of a related one. An AI layer that indexes the whole document management system, caches answers, and remembers past conversations can walk straight through that wall without anyone noticing. We wrote recently about the answer that belonged to someone else, and how an index drops the permissions on the content it copies. In a law firm the cost of that failure is specific: a conflicts problem, a disqualification motion, and a client who wants to know why their strategy memo informed an answer given to the lawyer across the hall.
Candor is a provenance problem
A lawyer owes the court candor. She cannot make a false statement of law, and she has to correct one she made. An invented citation is the bluntest way to break that rule, but it is not the most common one. The more ordinary failure is a real case cited for a proposition it does not support, or one that was good law when the model learned it and has since been reversed or distinguished out of usefulness.
Treat that as a provenance problem and it becomes manageable. Every authority in a draft should trace back to a source someone can open, read in the original, and confirm is still current. We have argued before that the answer was never the hard part, and that a system has to be able to say where an answer came from. Legal work adds a freshness requirement on top, much like the retired document that keeps answering: an overruled case still sitting in the index will keep getting cited with total confidence. Many judges now have standing orders that require lawyers to disclose AI use or certify that a human checked every citation. Those orders are a floor. A firm that can show the trail behind each authority in a filing never has to worry about meeting them.
Supervising something that is not a person
The professional rules already know how to handle work that a lawyer did not personally produce. Partners are responsible for supervising associates, and lawyers are responsible for the work of the nonlawyers who assist them. The sensible reading, and the one the ABA took, is that AI output belongs in that second category. It is assistance. The lawyer who relies on it owns the result, exactly as she would own a paralegal's memo she forwarded without reading.
That makes the person reviewing the output the control that matters most. Their review has to be real. A partner who skims a forty-page draft at the speed it was generated is approving it, not supervising it, and a disciplinary board will see the difference even if the billing system does not. We have written about agentic systems that need a hand on the wheel, and in legal work that hand belongs to someone with a license on the line. The design job is to make checking easy: show the sources next to the claims, flag what the tool was unsure of, and send the high-stakes passages to someone senior enough to catch what a junior reviewer would miss.
Fairness shows up before the client does
Most of the public conversation about legal AI is about briefs and research. Some of the less visible uses carry more ethical weight. Firms and legal aid organizations are using models to screen intake, predict likely outcomes, prioritize which matters get attention, and set early settlement ranges. Each of those decides who gets help and on what terms, often before a lawyer has spoken to the person involved.
A model trained on a firm's own history learns that firm's habits. If certain kinds of claims, from certain places, have historically settled low or been turned away, an intake model will learn to turn them away faster and with a cleaner-looking score. Nobody chose that outcome, and the overall accuracy figure will not show it. Testing outcomes by the groups a matter affects, before deployment and after, and keeping a person on any decision that declines representation, is the minimum. For a profession that describes itself in terms of access to justice, it is also a question of whether the firm's tools match what it says about itself.
The bill has to tell the truth too
General counsel have noticed all of this. Outside counsel guidelines increasingly ask how firms use AI, where client data goes, and how the bill reflects work the tool did. Some clients want to approve AI use matter by matter. Some want the savings passed through. Nearly all of them want to know that a person with the right experience reviewed what they are paying for.
A firm that can answer those questions in specifics wins work. One that answers with a policy PDF and a reassurance does not. Trust here is commercial, not abstract: it decides whether the firm gets to use AI on the client's matters at all, and at what price. Transparency about how the work was produced, a bill that matches the effort spent, and a clean record of who verified what turn into a competitive advantage that a firm with a faster tool and a vaguer story cannot easily copy.
How we build for the people who sign
When we build for legal and professional services, we start from the signature and work backward. Lumen, our knowledge solution, is built around the idea that an answer is only useful to a professional if she can see exactly what it rests on and whether that source is still current. Where the work is an investigation or a regulatory matter, Throughline carries the case with its evidence and its trail attached, so the person who has to defend the conclusion can see how it was reached. Both run on Neura-Cortex, the shared substrate that holds the guardrails, the identity and access controls that keep matter walls intact, the evaluations, and the audit log. Confidentiality, provenance, and a record of who reviewed what live in the floor beneath the solutions, so they do not depend on each practice group remembering to set them up.
We are not going to describe how those systems are put together. The point is the stance: in a profession where the signature is the product, the responsible way to use AI and the profitable way to use it end up being the same. Both come down to work a lawyer can stand behind and a client is willing to pay for.
Hours saved on a draft are worth nothing if no one can vouch for what the draft says. In law, the tool writes faster. The lawyer still has to be able to swear to it.
A spot check for the managing partner
Pick three documents the firm sent out last month where AI helped with the drafting: a brief, a client memo, and something transactional. For each one, take every authority it cites and see how long it takes to show who confirmed the source exists, says what the document claims, and is still good law. Then check what client information went into the tool to produce it, where that information is now, and whether any of it came from a matter the drafter was walled off from. Finally, put the bill next to the work and see whether a client reading both would think they were treated fairly. If all of that takes an afternoon, your program is in decent shape. If it takes a week, or nobody can say, you have a faster drafting process sitting on a liability you have not priced.
The paralegal who could not find the case did the most valuable work on that brief, and it took her about four minutes. If you take one rule back to the firm, make it this: nothing the tool wrote leaves the building until a named lawyer has read the authority it depends on, in the original, and is willing to say so. The rest of what we have described here is how you make that rule cheap enough that people actually follow it.
MTekLabs designs, deploys, and governs production-grade agentic AI platforms for government and commercial enterprises. Our legal and professional-services work is built human-in-the-decision-loop, with provenance and confidentiality designed in from the start. Explore our solutions at mteklabs.com.
Could your firm trace every authority in last month's filings?
We would be glad to talk through what it takes to put AI-assisted work in front of a court or a client with confidence, from provenance and matter walls to the review that keeps a licensed name meaningful.