What is the first pass costing your firm? Run your numbers
All briefings Buyer's guide

The best legal AI tools, and how to tell them apart.

There is no single best legal AI tool, because the leading products are not competing for the same job. Four shapes have emerged. Knowing which one you are buying matters more than any feature list.

20 September 2026 · 11 min read · The Solicia team

We build one of these products, so read this with that in mind. We have tried to write the guide we wanted when we started: what the categories actually are, who is strong in each, and the questions that separate a tool that survives a year at a firm from one that quietly stops being opened. Where Solicia belongs, we say so plainly and leave you to judge it against the same criteria as everyone else.

Four kinds of product

“Legal AI” is now a label over four quite different things. Most disappointment we hear about comes from a firm buying one shape while expecting another.

  • Research assistants. Ask a legal question, get an answer grounded in case law and legislation, with citations. CoCounsel (Thomson Reuters) and Lexis+ AI are the obvious examples, and their advantage is the proprietary corpus behind them. If your firm already pays for the underlying research platform, the AI layer is usually the cheapest good decision available.
  • Drafting copilots. These live where you already write. Spellbook is the clearest example: a Word add-in that reviews and redlines contracts against your playbook, in the document, while you work. Adoption tends to be strong because there is nothing new to open.
  • Enterprise platforms. Harvey and Legora sit here: broad workspaces covering research, drafting, review and firm knowledge, sold to large firms with the implementation budget and the internal champion to match. A platform that lands well becomes the firm’s AI layer. Not every one does.
  • Delegation services. The newest shape, and the one we build. You do not operate a tool. You send an instruction and receive finished work, and the measure is whether the returned file is one you could send.

None of these is better in the abstract. A litigation boutique that wants faster authority-checking should buy research. A corporate team drowning in NDAs should buy a drafting copilot. A firm trying to move a whole class of work off the associate roster is buying something else again.

Adoption

Firms usually open a procurement conversation with security or accuracy. Both matter, and we come to both. But the question that best predicts whether a tool is still in use after twelve months is a plainer one.

Who has to change their day for this to work?

Most legal AI that goes unused was dropped because it asked a fee-earner to add a step. A platform that requires opening a new window, re-uploading documents that already exist in the DMS, and re-explaining a matter the firm already knows about is competing against the path of least resistance, and the path of least resistance usually wins.

This is why the in-document copilots enjoy such strong adoption relative to their capability, and why the broadest platforms often show the widest gap between licences bought and licences used. When you run a pilot, instrument that gap. Ask the vendor for weekly active users as a percentage of seats, not total queries. Total queries flatter everyone.

Evaluating the output

Vendor demos are built from matters the vendor has already run, so they work. Three tests help once the demo is over.

  • Bring your own worst document. Use the 2014 scanned deed with the handwritten amendment, or the 200-page agreement whose defined terms contradict each other, rather than the clean template. Weaknesses show at the edges.
  • Check the citations by hand, once. Pick ten. Follow every one to source. You are learning what its failures look like, so your people know what to watch for. A vendor confident in its grounding will encourage this.
  • Ask what comes back. A summary you have to rebuild into a document is a draft of a draft. Ask whether you receive a working file: a .docx with tracked changes and numbering intact, or an .xlsx with live formulas rather than pasted values. Rebuilding an answer into a file is most of the actual work.

Security questions for vendors

Most legal AI runs on the same hyperscale infrastructure your firm already trusts for email and document management. That is the reason the category is viable at all. The meaningful differences lie in the fencing built on top of the cloud, not in which cloud it is. We have written about this at more length elsewhere, but in a procurement conversation, five questions do most of the work:

  • Is the no-training commitment in the contract? Not the FAQ, not the sales deck. A clause, in the agreement you sign.
  • Whose model endpoint do our matters reach, and who holds the contract with the model provider? Enterprise cloud AI deployments carry data-protection terms that consumer endpoints do not.
  • Where is data stored, and what is processed elsewhere? These are different questions and vendors sometimes answer only the first.
  • Is the subprocessor list published? A vendor that will not name who touches your data before you sign will not be quicker about it afterwards.
  • Can we set retention, and can you show a purge ran? Deletion is engineering, not policy.

These are answerable in writing by any serious vendor. Ask all five of everyone you shortlist, including us. The pattern of hesitation is more informative than any single answer.

Which to choose

If you are a large firm with a dedicated innovation function and budget for a real implementation, the enterprise platforms are built for you, and Harvey and Legora are the serious names. If your pain is contract review and your people live in Word, a drafting copilot will show value in a fortnight. If you already pay for a major research platform, turn its AI layer on before buying anything new.

If work is stuck behind capacity, such as the diligence spreadsheet nobody has started or the chronology that will take a week, you need capacity rather than a tool your people operate. That is the shape Solicia is. You send the instruction and the attachments to an email address, and the finished file comes back on the same thread. There is nothing to install and nothing to learn. That is deliberate, because a tool that asks people to change how they work tends to go unused.

That shape has real limits, and you should weigh them. It is not an interactive workspace. It suits delegable, bounded work far better than exploratory thinking. A lawyer who wants to iterate live on a clause is better served by a copilot in the document.

Buy the shape that matches the work. Then hold every vendor, including the one you like, to the same five questions in writing.

If the delegation shape fits, you can request access, or read how the work is checked and sourced before it comes back. 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.