The best tool is the one that already holds the file and shows you the source beside the summary. Everything else is a distant second.
Key takeaways
For a Canadian practice, the best AI for summarising legal documents is normally the one built into the system that already holds the file, because it produces the fewest new copies of privileged material and can show the source passage beside the summary. Fluency is the least useful thing to compare; anchoring and data handling are the two that decide whether the output is usable.
There is no single product answer, and any page that gives you one without asking what the documents are and where they live is selling something.
Does it show its source? A summary line you cannot trace to a page is not reviewable, and an unreviewable summary of a legal document is worse than no summary, because it invites reliance. Tools increasingly build this in: Clio describes its Manage AI feature set as including summarisation of lengthy documents into actionable takeaways, and, for deadlines extracted from court documents, a side-by-side view of the document source so the user can review and confirm accuracy. That side-by-side pattern is the one to insist on generally.
Where does the document go, and what happens to it? This is a contract question with a technical answer, and it should be settled before a file is uploaded. Vendors publish their positions — Clio's AI principles page sets out its stated approach, including transparency and security commitments — but a published principle is a starting point for the terms, not a substitute for them.
Does it fit the document type? Summarising a 40-page commercial lease, a 900-page document production and a transcript are three different jobs. A tool that is excellent at one may be indifferent at another, and the only way to know is to test it on your own documents.
Putting a privileged document into a third-party service is a disclosure decision, and it needs to be made deliberately rather than by whoever had a browser open. The categories are worth being precise about: our sister firm covers litigation privilege and how privileged documents are treated in Ontario litigation.
In litigation there is a second constraint that catches firms out. Documents obtained through discovery are subject to the implied undertaking rule, which limits the use that may be made of them. Feeding an opposing party's production into a general-purpose consumer service is not obviously consistent with that, and "it was only for summarising" is a poor answer to give afterwards.
The pressure point in most litigation practices is documentary discovery, not the lease review. Documentary discovery and the affidavit of documents and electronic discovery in Ontario civil litigation set out the framework; the practical effect is thousands of pages that someone must characterise.
Summarisation helps there in a narrow, useful way: producing a per-document one-line description and a date, so that a person triaging a production reads a table rather than the documents. It does not decide relevance, it does not decide privilege, and the affidavit is sworn by a person who has to be able to stand behind it.
A firm receives a production of 900 pages, mostly email with attachments. The associate's job for the week is to know what is in it.
The run produces one row per document: date, author, recipients, one-line description, page range, and a flag for anything that mentions three named topics the pleadings turn on. It does not produce a narrative of the case, because a narrative would have to choose which documents matter, and that is the associate's job.
The associate sorts by date, reads the flagged 40 documents in full, and skims the rest. Two documents contradict the defence's chronology — found because the table put them in date order, not because a model understood the case. That is the realistic shape of the benefit.
Ontario's position is not ambiguous. Under the Law Society of Ontario's Rules of Professional Conduct, rule 3.1-2 requires a lawyer to perform any legal services undertaken on a client's behalf to the standard of a competent lawyer, and the rules define a competent lawyer as one who applies relevant knowledge, skills and attributes to each matter — including investigating facts and identifying issues. Delegating the reading does not delegate that.
The Law Society's Technology practice management guideline is the companion piece: it invites lawyers to consider electronic document management and analysis support technologies to serve clients conscientiously and cost-effectively, while reminding them to address security and to know the legal and regulatory provisions governing the technologies they use. Read together, the message is that using technology well is part of competence, and using it carelessly is not a defence.
Data protection is treated as its own topic in the Law Society's technology resource centre, and it is the section to read before choosing where client documents are processed. Lawyers outside Ontario should be reading their own law society; the duties are similar in substance and different in detail.
Never rely on a summary for anything you would cite. Authorities get checked in the source, clauses get read in the document, and figures get confirmed on the page. A summary tells you where to look; it is not evidence of what is there.
The same discipline applies to retention. Once litigation is anticipated, what you keep and what you generate both matter — see document retention once litigation is anticipated.
Which tool is best?
The one inside the system your files already live in, if it anchors to source and its terms are acceptable. Adding a second system means a second copy of privileged material, and that cost is usually larger than the difference in summary quality.
Can I use a consumer chatbot for a quick read?
Not on client material, unless you have settled the terms and the privilege position first. The convenience is real and so is the disclosure.
Does a good summary let a junior skip reading the document?
No, and treating it that way is how a firm trains people who cannot do the work. Use summaries to decide what to read, never as a substitute for reading what matters.
A 30-minute call is enough to tell you whether AI pays for itself here.
