How law firms can use AI to automate workflows

The fastest way to understand legal AI is to look at who got burned. In Mata v. Avianca, Judge P. Kevin Castel found that attorneys “abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations,” and imposed a $5,000 penalty. The model had invented cases. The lawyers filed them unread. The court was plain that there’s nothing wrong with using “a reliable artificial intelligence tool for assistance,” but that existing rules “impose a gatekeeping role on attorneys to ensure the accuracy of their filings.”

So here’s the short version for a managing partner. The firms pulling real value out of AI aren’t the ones with the fanciest model. They’re the ones with a disciplined intake-to-closing process the model plugs into, with a human review step the work can’t route around. The model triages, classifies, and drafts. A competent lawyer reviews anything that reaches a client or a court. The tool changed; the duty didn’t.

This is the evergreen vertical playbook, so a scope note up front. The deep treatment of what is actually working with legal AI after the hallucination crisis lives in its own post, and so does the service-delivery angle for online legal work. This one maps the whole firm and where a model safely fits.

Where AI fits in a law firm

On the work that’s repeatable, document-heavy, and reviewable before it counts. A law firm runs on the same hidden truth as any operation: a sharp tool laid over a sloppy process just produces sloppy work faster. Point a capable model at an undefined intake-to-closing workflow and the mess only moves quicker. Point it at a disciplined one and the model takes the rote reading while the lawyer keeps the judgment.

The discipline is older than the tool, and it’s the part that travels.

That distinction matters more in law than almost anywhere, because the billable model punishes wasted time. Clio’s Legal Trends data puts the average law-firm utilization rate at 38%, meaning lawyers capture only about 3 billable hours of an 8-hour day. The rest leaks into intake, admin, and coordination. That leak is exactly the reading-and-drafting work a model can take, which is why the question isn’t whether to use AI but where to put it inside how AI is reshaping who does the work and who checks it.

The model belongs on the proposing steps: finding candidate cases, summarizing a long record, drafting first-pass language, classifying documents. It doesn’t belong on the committing steps: clearing a conflict, giving legal advice, or signing a filing. A wrong proposal costs a reviewer a few minutes. A wrong commit costs a sanction with your name on it.

Put a model behind these five workflows

Start with the matter lifecycle you already run. Each step below has reading a model can take and a sign-off a lawyer must keep.

Client intake, conflict check, matter open. The model triages the inquiry, captures the structured data, and drafts the conflict search across your systems. A lawyer signs off on the conflict result, because clearing a conflict is a committing step. The matter-opening checklist then runs on rails.

Document and discovery review. The model takes first-pass classification and privilege flagging across a large set. A lawyer reviews the calls, especially the privilege ones, because a missed privilege flag is its own kind of malpractice.

Billing and invoice review. The model cleans up time narratives and drafts the pre-bill. A partner reviews before it goes out. Given the utilization math above, tightening this loop is found money.

Matter closing. The model assembles the closing checklist, routes files for retention, and hands off the trust-account reconciliation. A person owns the final sign-off.

Engagement-letter drafting. The model drafts from the matter facts; a lawyer reviews and signs. Fast to draft, never auto-sent.

These templates already hold the right bones: a defined entry, checks along the way, and a sign-off a named person owns. Take a new matter through the intake one with AI in the reading seats. The model reads the inquiry, drafts a structured summary, and runs the conflict search across the firm’s prior clients and adverse parties, surfacing the possible hits with the context a lawyer needs to judge them. A lawyer reads those hits and signs the conflict result, because clearing a conflict is the committing step. The matter-opening checklist then runs, the engagement letter drafts from the captured facts, and a partner reviews before it goes out. Dropping an AI step into the reading parts, while keeping the sign-off human, is most of the work, and it’s the part that gives an associate their afternoon back.

Now name where AI doesn’t belong, because a buyer’s guide that hides the limits isn’t one a careful firm will trust. Legal advice to a client stays with a lawyer. A filing never leaves unread. A conflict is never cleared without a human signature. And nothing that implicates the duty of competence or supervision gets handed to a model and forgotten. The model drafts and triages right up to those lines, and a person owns everything past them.

Draw that line clearly and the model finally gets useful.

The orange gate in that diagram is the whole game. The AI’s output parks there, a lawyer reviews it with the source material attached, and only a recorded pass moves it forward while a fail routes it back. That gate is what turns “use AI carefully” from a poster on the wall into a step the draft can’t skip.

Who answers when the AI invents a case?

The lawyer who filed it. That’s what Mata settled, and it’s why the review step isn’t optional. The base rate makes the risk concrete: Stanford researchers benchmarked the purpose-built legal AI tools and found Lexis+ AI and Ask Practical Law AI wrong more than 17% of the time, and Westlaw’s AI-Assisted Research wrong more than 34%, even after vendors rebuilt them on real case databases. General chatbots fabricated on 58% to 82% of legal queries. The tools are a real improvement. They aren’t something you file unread, and the wrong sixth looks exactly like the right five. The sibling post on what is working digs into the verification step itself; the point here’s that the duty to run one is already written down.

The rule is older than the technology, which is what makes it bite. ABA Model Rule 1.1, Comment 8 asks a lawyer to keep abreast of changes in the law and its practice, “including the benefits and risks associated with relevant technology.” Read next to a documented one-in-six error rate, that competence duty turns a skipped review into a professional-conduct problem. Stack on Rules 5.1 and 5.3, which put partners on the hook for supervising the lawyers and non-lawyers, and now the AI tool, under them, and the verification gate stops reading as good hygiene and starts reading as the floor.

The question we hear from managing partners is whether they can let the model file routine, low-stakes matters to save associate time. For anything that reaches a court or a client, no, the review is the control, and a tool that’s right most of the time is exactly the trait that talks a careful person out of looking. In Tallyfy terms, the model’s draft parks at a blocking approval step with a named reviewer, and the run history records every step as it happens, so privilege and supervision both have a trail. Worth saying plainly, since I run a workflow company and not a law firm: this is the shape of the duty, not legal advice on your matter.

When to do it yourself, and when to get help

Be straight about the split. The drafting and classification pilots are yours to start now. Pick one workflow, intake-and-conflict or document review is the usual first choice, put a model on the reading step, keep your existing sign-off, and watch what it catches. Small, contained, easy to reverse if the model disappoints.

Firm-wide rollout is the harder problem. Once AI runs across matters, you’re into supervision policy, privilege handling, malpractice exposure, and a defensible record you would show a disciplinary panel. Supervision is the part that catches partners out: under the duty to oversee the work of others, a partner is answerable for what a model produces under them the same way they answer for an associate’s draft, so “the AI did it” isn’t a defense anyone has won with. The mistake we watch firms make when they roll AI out is treating the firm-wide version like the pilot, scaling before the review gate and the audit trail are built to carry it. That’s where a vendor-neutral view of what to automate, and in what order, is money well spent, because the sequencing is the risk.

The crisis didn’t end legal AI. It ended the version that files unread. The firms that handle it well won’t be the ones whose models never erred, because every model errs. They will be the ones who can show, for anything consequential the AI touched, that a lawyer checked it before it counted, the way the wider move to workflow automation already records work that has no AI in it at all.