There is no single best tool. There is a best fit for your document types, your Word workflow and your confidentiality obligations — and four questions that settle it.
Key takeaways
The honest answer is that the best tool is the one that fits your document mix, your drafting environment and your confidentiality obligations — and you can only establish that by testing on your own files. A firm doing high-volume commercial leases, a firm doing complex M&A, and a firm doing employment agreements are three different buyers with three different answers.
What is common to all three is the evaluation. Four questions decide it, and none of them is about which model sits underneath.
Before capability, settle data handling. Where is the document processed, where is it stored, who at the vendor can access it, how long is it retained, and is prompt content used to train or refine the model? The Office of the Privacy Commissioner of Canada’s principles for generative AI say providers should inform organisations about secondary purposes such as using information from prompts to further train a model, and that organisations should only use tools that respect privacy laws. Ask for these answers in the contract, not the sales deck.
For Canadian firms the follow-on question is residency: if client data leaves the country, you should be able to say so and say where it went. That is a procurement decision with professional-conduct consequences, given the strict confidence obligation in rule 3.3-1 of the Model Code of Professional Conduct.
Contract lawyers work in Microsoft Word. A tool that requires exporting a document, reviewing it in a browser and re-importing the result will be abandoned within a quarter regardless of how good the analysis is. The products that stick are the ones that put review and redlining inside the document. Spellbook describes itself as reviewing and redlining contracts against your standards inside Word, and drafting clauses from your preferred precedents. Luminance likewise describes AI-powered review within Microsoft Word alongside contract generation and analysis across a contract portfolio.
The alternative shape is a platform built for volume rather than for the drafting seat. Kira, now part of Litera, positions itself as contract intelligence for high-volume review with governance controls that can be toggled on or off — the shape that suits due diligence and portfolio extraction more than clause negotiation. Alexi describes a single platform for legal research, drafting and analysis with single-tenant deployment, which is the shape firms ask for when isolation of their data is the deciding factor.
A tool that reviews against a generic notion of market terms produces generic output. A tool that reviews against your written standard positions produces something your clients cannot get elsewhere. Ask specifically: can we load our own standard and fallback positions, per clause, and does the review report against them by name?
This is also the honest reason some firms get no value from these products. If the firm has never written its positions down, there is nothing to configure, and the tool degrades into a summariser.
The single most useful feature in practice is unglamorous: does every finding point at the clause it came from, with quoted text, so a lawyer can verify it in seconds? Output you cannot trace is output you have to re-derive, which removes the saving entirely.
What to ask every vendor, in writing
Vendors publish accuracy figures. Those figures are self-reported, measured on the vendor’s own benchmark set, and tell you nothing about performance on your scanned lease amendments. Treat them as a claim to be tested, not a specification.
The test that settles it takes a day. Pick twenty agreements you have already reviewed and where you still have the issues list. Run each tool over them without showing it your conclusions. Then count three things: issues on your list the tool missed, issues it raised that were not real, and issues it raised that you had missed. The third number is the one that justifies the licence; the first is the one that sets how much checking you will always have to do.
Do this with real documents from your own practice, under whatever confidentiality terms the trial requires. A demonstration on the vendor’s sample contract measures the sample contract.
Whichever tool wins, rule 3.1-2 of the Model Code still requires services to be performed to the standard of a competent lawyer, and its commentary still expects you to understand the benefits and risks of the technology you have adopted. Your own regulator governs — in Ontario the Rules of Professional Conduct — and the analysis a tool produces is your advice once it leaves the building. Where the drafting question is substantive, the underlying law is unchanged too: Treadstone Law’s note on the duty of good faith in Ontario contract performance is a reminder that a clause-by-clause comparison does not capture everything a court will.
Is a general-purpose chatbot good enough?
For a one-off summary of a document that contains nothing confidential, sometimes. For client work at volume, the purpose-built products earn their keep on the parts that are not the model: playbook configuration, Word integration, clause-level citation, and enterprise data terms.
Should a small firm buy anything at all?
Buy the workflow you actually repeat. A two-lawyer firm reviewing the same three agreement types will get more from a written playbook plus a modest tool than from an enterprise platform nobody configures.
What about drafting from scratch?
Every product in this category will produce a first draft. The value is in drafting from your precedents rather than from the open internet, so weigh precedent handling more heavily than raw generation quality.
A 30-minute call is enough to tell you whether AI pays for itself here.
