AI Answer Summary

An AI legal assistant can mean four unrelated products: AI-native platforms built for firm-wide matter work, AI tiers layered onto an existing research subscription, assistants embedded inside a document or practice management system, and consumer chatbots that are not sold to law firms at all. Establishing which type you need does more for a shortlist than comparing brands.

None of the leading platforms publishes a price. Third-party estimates vary widely enough to be unreliable on their own, and at any figure within that range, licensing every lawyer at a hundred-lawyer firm costs somewhere between six and seven figures a year.

That points to a different question from the one vendors ask. For a mid-market firm the decision is which lawyers get a seat, what everyone else uses, and which workflows to build rather than license.

What an AI Legal Assistant Is: Four Products, One Name

Four unrelated products share this name, and one of them is not sold to law firms.

You can check this yourself. Search the term and a single page returns firm-grade platforms, a document management add-on, an Australian product, a free legal chat service aimed at the public, and an app-store product selling contract templates to consumers. A buyer working down that list is comparing things that have almost nothing in common.

Pick the row before you pick the brand. Rows one and two compete directly with each other, since both want to be where research and drafting happen. Rows one and three often sit side by side in the same firm without conflict.

What AI Legal Assistants Cost, and Why No Vendor Publishes a Price

None of the leading platforms publishes a price, and the figures circulating come mostly from companies selling alternatives.

Pricing and funding checked as of 21.09.2026.

Treat every figure in the third column as unverified. They come from competitor blogs, analyst models, customer accounts and a widely circulated forum thread, and several of them contradict each other. Paxton's numbers are different in kind, because they come from Paxton's own help centre.

Now the arithmetic, which is why the disagreement matters less than it looks.

Take the lowest credible estimate for a top-tier platform and apply it to a hundred lawyers. Firm-wide licensing lands in six figures a year. Take the higher estimates and it passes a million. Every number in that range produces the same conclusion, so you do not need to resolve the disagreement to make the decision. Firm-wide standardisation is not the choice in front of a mid-market firm. Seat allocation is.

Seat minimums are the more reliable number, and for smaller firms they decide everything. A three-lawyer firm facing a ten-seat floor is buying seven empty chairs.

Why prices sit where they do: this market is capitalised for large-firm budgets. PitchBook reports Harvey's most recent raise accounting for 16% of the $3.35 billion invested in legal technology globally in 2026, with three deals taking 40.3% of the year's capital between them. Products get built for the customers funding them.

How We Compared These Tools, and What We Did Not Test

We have not run these products inside a law firm. Neither have most of the articles ranking for this search, which is worth saying out loud.

What this comparison uses:

• Published independent benchmarks, named and dated
• Peer-reviewed accuracy testing, where it exists for that specific product
• Structural facts a buyer can verify without a sales call: product type, integration requirements, whether a price is published, seat minimums
• Vendor positioning in the vendor's own words, identified as such

What it does not use: star ratings from review platforms that vendors pay to appear on, our own impression of product quality, and any claim one vendor makes about a competitor.

The evidence below is thinner and older than anyone would like. Saying so is more useful than implying a testing programme we never ran.

What Independent Testing Shows About Legal AI Accuracy

Three independent studies exist now. All are rigorous, and each describes a market that keeps moving.

The pattern in the first study is that reading beats amending. Every task the tools won involved reading a document that already existed. The one clear loss involved changing one.

The second study matters for a different reason. LexisNexis was marketing hallucination-free citations at the time, and the researchers tested what that claim survived. It is the first preregistered evaluation of these products, which means the method was fixed before anyone saw results.

Here is the limitation, and it is a large one. The Lexis+ AI that Magesh tested no longer exists. LexisNexis replaced it in February 2026 with Lexis+ with Protégé, describing the tested product as its first-generation AI experience, then expanded the replacement twice more during 2026. So the only peer-reviewed accuracy evidence in this market describes one product that has been retired and another that has been rebuilt.

That is not a reason to ignore the numbers. It is a reason to ask any vendor for current independent testing, and to notice how rarely one exists.

The Best AI Legal Assistants for Law Firms, by Type

These are grouped by type rather than ranked. A single ranking across types would compare products doing different jobs, which is how buyers end up with two tools that overlap and no tool for the thing that was hurting.

Harvey: Best for Firms Running End-to-End Matter Work

Harvey is the most heavily capitalised product in the category and positions around multi-step matter work rather than single tasks. It participated in the Vals benchmark. The company says it partners with most of the hundred largest US firms.

Best for: firms with the budget and the matter volume to run diligence or litigation workflows end to end.

Trade-off: the highest reported pricing and the highest reported seat minimum in this list, and a roadmap written for AmLaw100 requirements. A hundred-lawyer firm is not the customer this product is designed around.

Legora: Best for Cross-Border Teams Working in Microsoft 365

Legora positions around a collaborative workspace where a team works together, rather than an assistant that works alone. Stockholm-based, with a strong Microsoft 365 orientation, and reported seat minimums roughly half of Harvey's.

Best for: teams collaborating on documents across jurisdictions and offices.

Trade-off: less independent testing exists for it than for the research incumbents, so you are relying on your own pilot rather than published evidence.

CoCounsel: Best for Firms Already on Westlaw

Thomson Reuters. CoCounsel participated in the Vals benchmark, and its research sibling was tested in the Magesh study. It is tiered, and Westlaw is frequently required underneath it.

Best for: firms whose research subscription is already in place and who want the AI layer where their research already lives.

Trade-off: the AI price is not the bill. Ask what the total looks like with the research subscription included, since that is where comparisons against AI-native platforms usually go wrong.

Lexis+ with Protégé: Best for Firms Already on Lexis

The same logic from the other research incumbent, though this product has moved fast. LexisNexis replaced Lexis+ AI with Lexis+ with Protégé in February 2026, expanded it in May with a skills and orchestration layer, document vaults and shared workrooms, and rebuilt the orchestration again in August.

Best for: firms already committed to the Lexis content library.

Trade-off: no independent testing covers the current product. The published hallucination figures apply to the predecessor, and we would not attribute them to Protégé. Pace of change is itself worth weighing, since it means the product you evaluate in a pilot may differ from the one you deploy.

Platform-Embedded Assistants: Best for Keeping Work Inside the System of Record

NetDocuments ndMAX and Clio Duo are examples of AI built into the system where the documents already sit. The argument for them is that content never leaves the platform that already holds it, which removes a category of security and confidentiality questions before anyone asks them.

Best for: firms that have already standardised on the platform and want AI where the files are.

Trade-off: capability is bounded by the platform's own roadmap, and you inherit whatever priorities that vendor sets.

Paxton: Best for Testing Before You Commit

Paxton is the one product here that publishes a price on its own site: $499 per user per month, or $2,999 per user per year on annual billing, with a seven-day trial and no seat minimum on the individual plan.

That transparency is the point of including it. A firm can form an evidence-based opinion on its own matters before entering an enterprise procurement cycle, and can walk in knowing what a published legal AI seat costs.

Best for: a firm that wants its own data before committing to a minimum and a twelve-month term.

Trade-off: thinner integrations than the platforms above. Worth noting that third-party sites list Paxton's price at $199, $250, $299 and $500, none of which matches what the vendor publishes, which tells you something about the accuracy of legal AI pricing content generally.

How Many Lawyers Need a Seat

Fewer than the vendor proposes, and the number should come from your own matter data rather than your headcount.

This is the question the market avoids, because every vendor's revenue depends on the answer being everyone.

A workable approach:

• Identify the practice groups whose daily work matches the tasks the benchmarks scored well: document-heavy review, summarization, chronology building from records
• Count the lawyers who do that work weekly rather than occasionally
• Buy for that group, clearing the seat minimum if one applies, with room to expand
• Set the measurement before the pilot starts: hours on the target task per matter, before and after

One caution worth planning for. Partial deployment creates a division inside a partnership between the lawyers who have the tool and the lawyers who do not, and that conversation goes better when someone has thought about it in advance rather than discovering it at a partners' meeting.

For everyone without a seat, a general-purpose assistant under a firm policy covers a good deal of everyday drafting and summarising at a fraction of the cost, provided the policy defines what client data may go into it and who reviews the output.

What to Build Instead of Buying Another Platform

A seat licence does not touch the firm-specific workflows where mid-market money leaks.

Three examples, each covered properly elsewhere: intake file completeness and conflicts screening before substantive questioning, billing narrative quality before the pre-bill reaches a partner, and inbound document triage wired into the systems your firm already runs.

The distinction is worth holding onto. An assistant helps a lawyer do their own work faster. A workflow build changes what reaches the lawyer in the first place. Firms tend to buy the first and stay stuck on the second, because the first is easier to demonstrate and easier to approve.

And the sequencing rule that applies across all of it: automate the work that was already being written off before paying per seat to make billable work faster. On hourly matters, faster billable work produces a smaller invoice.

How Codebridge Works with Mid-Market Law Firms

We build workflows rather than sell seats. Document review triage, intake and conflicts screening, and billing narrative cleanup are the three we are asked for most often by firms in this band.

One workflow goes live in three weeks, wired into the systems the firm already runs, with the approval checkpoint designed in during the first conversation and a record an auditor can read later. Your firm owns the repository, the prompts and the configuration from day one.

We should be straightforward about our position in this comparison. We do not sell an assistant platform and have no seat licence to protect, which is why this article can tell you that fewer of your lawyers need a seat than any vendor on this page will. Our founding team spent more than a decade at KPMG.

If you want to work out which workflow to automate before adding licences, book a 20-minute call.