What is the first pass costing your firm? Run your numbers
All briefings Primer

What legal AI actually is, and what it is not.

A plain account of what these systems do, where they are genuinely reliable, where they are not, and what a firm should expect in the first six months. Written for lawyers deciding whether to look properly.

20 September 2026 · 12 min read · The Solicia team

The short definition

Legal AI means software built on large language models that reads legal material and produces legal work product. That is narrower than it sounds. The underlying models are general (the same technology answers questions about cooking), and almost all of the difference between a useful legal product and a chatbot lies in what the vendor builds around the model: which sources it is allowed to read, how its output is checked, what format the work comes back in, and what happens to the file afterwards.

This matters when you evaluate. A demo is mostly showing you the model, and every vendor has access to models of similar quality. The parts that differ are the parts a demo does not show.

What it does reliably today

Four kinds of work are done well enough to change how a team operates.

  • Reading at volume. Extracting the same fields from 200 contracts, building a comparison table, finding every clause that deviates from a position. This is the strongest use case in the profession and the least glamorous. It works because the task is bounded and verifiable.
  • First drafts from precedent. Producing a draft that follows the firm’s own template and defined terms, for a lawyer to take apart. The saving is in the blank page, not the final document.
  • Chronologies and bundles. Dating events across hundreds of emails and attachments, each entry carrying the document it came from. Mechanical, tedious, and error-prone when done by hand late in the day.
  • Finding the authority. Locating relevant legislation and judgments, with citations to check. It locates authority and gives citations to check. It does not decide.

What it does not do

Being specific about the limits is more useful than any capability list, and vendors are generally too vague about them.

  • It does not exercise judgment. It can tell you a clause is unusual. Whether unusual is acceptable for this client on this deal is not a question it is equipped to answer, and a system that sounds confident about it is more dangerous than one that declines.
  • It does not know what it has not been given. Most bad output traces to missing source material rather than faulty reasoning. If the side letter was not in the folder, no amount of model quality recovers it.
  • It is not uniformly reliable across jurisdictions. Coverage of primary material varies enormously. A system fluent in Delaware corporate law may be noticeably weaker on Australian state legislation. Test in your jurisdiction, on your subject matter.
  • It does not remove the review obligation. Nor should a vendor suggest it does. The professional duty sits with the person whose name goes on the advice, and no product changes that.

It does the work a capable paralegal does, faster than any paralegal, without the judgment a lawyer brings.

Hallucination

The failure everyone has heard about, invented cases cited to a court, is real and well documented. In serious products it is largely controlled by design. It happens when a model is asked to recall from memory rather than read from a source.

A properly built legal system does not work from memory. It retrieves the actual statute or judgment, reads it, and cites what it read. The relevant question to a vendor is therefore not “does your system hallucinate”, since everyone says no, but “show me how a citation in this output is tied to a document you retrieved”. If they can show the chain, the failure mode is contained. If the answer is about model quality, it is not.

The residual risk is subtler and worth knowing: a system can cite a real case correctly and still characterise its holding loosely. That is imprecision rather than hallucination, and the review a junior’s work gets will catch it.

The first six months

Firms tend to expect a step change and get something slower that lasts.

In the first month, one or two people use it properly and everyone else forgets. In the second and third, a pattern emerges: a particular task, usually diligence tables, chronologies or first-pass contract review, starts going to the tool by default because someone proved it works. By month six, the realistic outcome is that one class of work has moved, and the firm has a much better sense of what else might.

The firms that get least from it are the ones that buy broadly and train nobody. The firms that get most pick a single painful, repetitive task, run it through properly for a month, and only then widen. This is unglamorous advice and it is the strongest predictor we have seen.

The questions worth asking before you buy

  • What comes back: an answer to read, or a working file I could send?
  • Who has to change their day for this to work?
  • How is a citation tied to a document the system actually retrieved?
  • Is the no-training commitment in the contract, and where is the subprocessor list published?
  • What does it do when it does not know? A system that guesses is worse than one that stops.

We have written a fuller guide to the categories of tool and how to compare them, and a separate piece on whether the security case actually holds.

Where we sit

Solicia is one answer to the first question on that list. Rather than a workspace you operate, it works by email: you send the instruction and the files to an address, and the finished document, a .docx with tracked changes intact or an .xlsx with live formulas, comes back on the same thread. That design exists because the most common reason legal AI goes unused is that it asks a fee-earner to open something new.

It suits delegable, bounded work. It is not the right shape for a lawyer who wants to iterate live on a clause, and we would point that lawyer at a drafting copilot instead. Judged against the questions above, we would rather you apply them to us than take our word for any of it.

If that fits, you can request access. Questions go to admin@solicia.ai. A person reads them.

Writing your firm’s AI policy? Get the AI use policy template for Australian law firms, free in Word.

Get the next briefing by email.

New briefings by email, a few times a year.

One email to confirm, then the briefings. Unsubscribe in one click, any time. Privacy Policy.