Most firms that bought an AI tool in the last two years are not using it. The licence renews, a few people log in occasionally, and nobody can say whether it saved any time. Meanwhile the actual bottlenecks, intake that goes cold, documents nobody has indexed, billing narratives written from memory on a Friday afternoon, are unchanged.

That gap is not a technology problem. It is a problem of matching a tool to a task that is genuinely repetitive, verifiable, and low-stakes enough to survive being wrong occasionally. Most conversations about ai for law firms skip that matching step entirely and go straight to demos. This article works through where the technology holds up in a small or mid-size practice, where it does not, and how to tell one from the other before you sign anything.

The tasks where it reliably works

The pattern is consistent. AI performs well when the input is unstructured text, the output is a first draft rather than a final answer, and a human who already knows the matter reviews it in under a minute.

Intake and triage is the clearest case. A prospective client fills out a web form or leaves a voicemail. The text or transcript gets summarised into a structured record: practice area, jurisdiction, apparent statute of limitations exposure, conflict-check names, and a plain-language description of what happened. That record lands in Lawmatics, Clio Grow, or whatever your CRM is, with a suggested priority. Nobody is deciding whether to take the case. Somebody is deciding it faster because the file is already organised.

Document summarisation works for the same reason. Medical records, deposition transcripts, discovery productions, and long email chains all reduce well to timelines and issue lists. The summary is not evidence and it does not go in a brief. It tells an associate which of the four hundred pages to actually read.

Billing narrative cleanup is underrated. Attorneys write terse, inconsistent time entries. A model can expand them into compliant, client-readable narratives that follow your firm's conventions, flag entries that look block-billed, and catch entries that violate an insurer's guidelines. The attorney still approves every line. The work of rewriting eighty entries by hand disappears.

Email and correspondence drafting works when the model has the matter context. Status updates, scheduling, document requests, and routine client check-ins are formulaic. A draft that is eighty percent right and sitting in the attorney's queue is more useful than a blank window.

Internal knowledge retrieval works if your documents are already organised. Ask a question and get an answer drawn from your own precedent bank, prior briefs, and firm memos, with citations back to the source file. Note the condition: if your document management is a mess of shared drives and inconsistent naming, this does not work, and no vendor will fix that for you.

A worked example: personal injury intake

Consider a six-attorney plaintiff-side personal injury firm running Filevine, with a website form and a tracked phone number. Roughly forty inbound leads a month, and the partners know some fraction go cold because nobody called back within a day.

The automation is not glamorous. A web form submission or call transcript hits a workflow. The model extracts date of incident, incident type, injuries described, whether the caller has already spoken to an insurer, whether they have counsel, and the names of any other parties mentioned. It calculates days elapsed since the incident and flags anything approaching the relevant state limitations period as urgent.

It runs the extracted names against the existing Filevine matter list for obvious conflicts, then writes a three-sentence summary in plain English. All of that goes into a new Filevine project record and a Slack message to the intake coordinator, ordered by urgency.

Simultaneously, the caller gets a text within two minutes confirming receipt and offering three specific consultation slots pulled from live calendar availability. If they book, the intake coordinator gets a calendar hold with the summary attached. If they do not respond in four hours, a second text goes out. If they do not respond by the next morning, the lead is escalated to a human phone call rather than being left in a queue.

What the system does not do: decide whether to take the case, tell the caller anything about the merits, estimate value, or send anything to the prospective client beyond scheduling logistics. Every substantive touch is a person. The automation removed the lag and the manual data entry, which is where the leads were being lost.

That is what workable ai for law firms looks like in practice. Boring, narrow, and measurable.

Where it does not work

Be direct about this, because vendors will not be.

  • Legal research and citation. Models still fabricate case citations and misstate holdings with total confidence. Purpose-built research platforms with retrieval grounded in real databases are better than general chatbots, but every citation still needs verification against the actual reporter. Treat AI research output as a lead, never as authority.
  • Anything adversarial or client-facing without review. Do not let a model send correspondence to opposing counsel, respond to a client question about their matter, or generate anything that goes to a court without an attorney reading every word.
  • Judgment calls dressed as data tasks. Case valuation, settlement posture, whether to take a matter, whether a witness is credible. Models produce confident-sounding output on all of these and it is not grounded in anything.
  • Complex document drafting from scratch. Models produce plausible contract and pleading language that contains subtle errors: wrong jurisdiction conventions, omitted defined terms, provisions that contradict each other. Assembly from your own vetted clause library is a different and safer approach than generation.
  • Anything where the input data is bad. If your matter records are incomplete or your files are unlabelled, automation propagates the mess at speed.

There is also a confidentiality dimension that deserves a real answer, not a checkbox. Before any client data touches a model, know where it is processed, whether it is retained, whether it trains a future model, and whether your engagement letters and applicable state ethics guidance permit it. Many general-purpose consumer tools are not appropriate for privileged material. Enterprise agreements with zero-retention terms exist and are worth insisting on.

How to tell the difference before you buy

Three questions filter out most bad ideas.

First, can you verify the output in less time than it would take to do the task yourself? If checking the AI's work requires the same expertise and nearly the same time as doing it, the tool is not saving anything. Summarising a deposition passes. Drafting a summary judgment motion does not.

Second, what happens when it is wrong? If the answer is that somebody catches it in review and edits a draft, fine. If the answer is that a client gets bad information, a deadline is missed, or a filing goes out with a fabricated citation, the task is outside the safe zone regardless of how good the demo looked.

Third, is this task actually costing you? Firms automate the thing that is easy to automate rather than the thing that hurts. Track where hours go for two weeks before deciding. The answer is often intake response time, document organisation, or billing preparation rather than anything that looks like legal work.

A useful rule: the best candidates for ai for law firms are tasks currently done by a person who is not an attorney, done the same way every time, and done from written or spoken input. If a paralegal or an assistant does it repeatedly and the output is a document or a data record, it is worth examining.

What to do first

Pick one workflow. Not a platform, not a firm-wide strategy, one workflow that a specific person does every day and complains about. Intake response is the usual candidate for litigation firms; document intake and indexing for transactional work; billing narratives for anyone with a collections problem.

Write down how it works today, including who touches it and what breaks. Then decide what a human must still do, and build the automation around that constraint rather than trying to eliminate it. Run it in parallel with the manual process for a few weeks and compare the output honestly. If it is not obviously better, stop and try a different workflow rather than adding features.

Firms that do this find the second and third automations easier, because the underlying work of cleaning up data and clarifying the process is already done. Firms that start with a platform purchase usually end up where they started, with a licence and a login nobody uses. Alphovia builds this kind of operational automation for law firms, but the sequencing matters more than who builds it.