Small firms hear about AI mostly from vendors. The pitch is a widget bolted onto the website, or a subscription that promises to read everything in the matter file. Neither one answers the question a managing partner actually has, which is where the hours are going and whether software can take some of them back.
We work with law firms. We manage the website of California Tenant Law, one of the top tenant law firms in California, with a team of over ten lawyers and paralegals. We built LOMan with the Aesopware team, a legal-operations SaaS platform covering workflow automation and case management. Sites for two more legal practices are in development. What follows is where our AI services have paid off inside firms that size, and where they have not.
Price the task before you pick a tool
A 20-minute task that runs 40 times a week is 13 hours a week. That is two full days of somebody's time, spent on one repeated motion.
Run that multiplication across a normal week and the candidates sort themselves. Frequent and short beats rare and large, even when the rare one is more painful to sit through. A quarterly report that eats a full day costs the firm about 32 hours a year; the 20-minute task above costs about 675.
The same number tells you what you can spend. A workflow burning 13 hours a week can justify custom engineering. A workflow burning 40 minutes a month cannot, and no demo should talk you into it.
Four places AI earns its keep
Intake triage. Firms get more inquiries than they can take, and sorting them is repetitive reading. A model can pull the facts that decide fit (jurisdiction, key dates, claim type, prior representation) and route each inquiry into a queue with a short summary attached. A lawyer still makes the call. They make it from a structured summary rather than a voicemail and a half-filled web form.
Document processing at volume. Leases, notices, discovery productions, records requests: high page counts, narrow questions. Pulling dates, classifying documents, flagging the pages a person needs to read are the tasks current models handle well, and they scale in a way that hiring does not. The output belongs in a review queue.
First-draft assistance with lawyer review. Standard letters, demand drafts, a summary of a file for a colleague picking it up. The model produces a first pass grounded in the firm's own templates and the lawyer edits. What you save is the blank page. The judgment stays where it was.
Internal knowledge Q&A. A ten-person firm accumulates years of precedent, forms, local practice quirks. Most of it lives in one person's head or in a folder nobody browses. An assistant grounded in the firm's own documents answers "how do we usually handle this" and shows the document it came from, which also tells you when the answer is out of date.
Where it does not belong
Legal judgment stays with lawyers. A model can tell you what a document says; deciding what to do about it is a different act, and nothing in current systems makes them accountable for it.
Nothing files unsupervised. Automation can prepare a filing, assemble the exhibits, calendar the deadline. A person presses send. Deadlines and signatures carry consequences that no rollback undoes.
Privileged data needs architecture before it needs AI. Where the documents sit, which model sees them, what gets logged, how long the logs are retained, who at a vendor could read them. Answer those questions before a client file goes anywhere near a third-party tool.
How we build it, and what comes first
Our AI work is Anthropic-first. We use Claude models, prompt caching to hold down latency and cost, and evaluation suites so regressions get caught before users find them.
Caching matters more in legal work than in most fields. A firm assistant re-sends the same background on every question: templates, policy documents, statute excerpts, the matter file itself. Caching that context means it is not reprocessed and re-billed on every turn. That is the difference between an assistant that answers in a couple of seconds and one nobody opens twice.
An evaluation suite is a fixed set of questions with known-good answers, run against every change. Without one, a prompt edit that fixes a single behavior quietly breaks three others and the firm hears about it from a client. With one, you see the break before it ships.
None of that helps if the underlying records are a mess, which is the more common blocker. LOMan automates the workflows firms used to handle with spreadsheets and paper. Case records live in one system, deadlines derive from dates instead of being retyped, status is visible without asking anyone. That is software work, and for most firms it has to come first. A model pointed at five spreadsheets and a shared drive will answer confidently from stale data.
Once the records are in one place, workflow automation has something to hook into. An intake creates a matter. A signed document moves a status. A deadline schedules its own reminders. The AI layer then sits on top of a system that knows what is true, which is the only condition under which its answers are worth reading.
What to do this month
Pick one repeated task and time it. Count how often it runs. If the product of those two numbers is more than a few hours a week, it is worth a conversation; if it is not, keep the spreadsheet and spend the money elsewhere. Our AI services work with firms usually starts with that one number rather than a platform decision.
Send us the task and how often it runs, and we will tell you whether it is worth building. You can start a project with a paragraph.