If you run a law firm of any size in 2026, you have been pitched an AI paralegal. Probably several. The demos look identical: a clean chat box, a medical chronology built in seconds, a demand letter drafted before your coffee cools. The pricing pages are less clean, and the sales calls are where the real differences hide. This piece is a buyer’s guide for cutting through that. I will name products, quote prices where they are public, and give you a selection lens that survives contact with a sales rep.
One disclosure up front, so you can weigh everything that follows: my company builds a managed AI paralegal service. I have written this the way I would want a competitor to write it, which means naming their strengths and my category’s limits. You can decide at the end whether I managed it.
The question is not “which tool.” It is “who reviews the work.”
Start with the fact that determines everything else. Under ABA Formal Opinion 512, the ABA’s first formal ethics opinion on generative AI (July 2024), and under the California State Bar’s practical guidance, the lawyer stays on the hook. Opinion 512 says a lawyer who relies on AI output “without an appropriate degree of independent verification” may violate the duty of competence. That boilerplate consent buried in an engagement letter is not enough to put client information into a self-learning tool. California’s guidance is blunter still, in the 2026 revision approved on 14 May 2026: “A lawyer’s professional judgment cannot be delegated to AI and remains the lawyer’s responsibility at all times.”
Read that as a buyer, not a bar examiner. It means no software you license will ever carry the review burden for you. The AI drafts; a qualified human still has to check every citation and every factual claim before it reaches a client or a court. So the real question when you evaluate an “AI paralegal” is not how good the draft looks in the demo. It is who does the checking, and whether your firm has the capacity to do it well at the volume the tool will produce.
That question splits the market cleanly in two, and most buyers never notice the split.
What you are actually buying: a tool, or a service
The larger camp is self-serve software. You buy seats, the AI produces work product, and your people review it all. Paxton is a good example and one of the few with pricing you can read off a page: 499 dollars per user per month, or 2,999 per year, positioned for small and midsize firms. Thomson Reuters CoCounsel has an “Essentials” tier aimed explicitly at firms of one to ten attorneys, which makes it more small-firm-accessible than its enterprise reputation suggests. However, access still runs through a sales-led pilot rather than a trial you can start yourself. LawDroid runs as low as 25 dollars a month. At the top end, Harvey sells almost exclusively to large firms; it does not publish pricing, and firms have reported quotes around 1,200 dollars per seat per month with multi-seat, annual minimums (reported figures, not published, so treat them as directional). Eve and Supio specialize in plaintiff and personal-injury work and gate their pricing behind a demo. Spellbook and Gavel focus on contracts inside Microsoft Word.
The smaller, newer camp is managed services, where the vendor pairs its AI with its own trained humans who review and finalize the work before it reaches you, priced per matter rather than per seat. This barely existed a year ago. It exists now because a serious competitor decided the review burden was the product. In May 2026, EvenUp launched “Pre-Litigation as a Service,” pairing its AI with its own case managers who work inside the firm’s system and charge a flat fee per case. Their CEO framed it as a deliberate move away from adding more people and manual processes, toward pairing AI with managed staff. When a well-funded incumbent reorganizes around the human-review layer, that tells you where the friction actually is.
Here is why the split matters more than any feature comparison. A self-serve tool multiplies your output and hands the entire verification duty back to you. If you are a four-lawyer firm and the tool now generates the draft volume of a twelve-lawyer firm, you have not solved your capacity problem. You have moved it downstream to review, where the ethics rules say it has to live anyway. A managed service absorbs the first layer of that review before the work reaches your desk. Neither model removes your final sign-off. They differ in how much unreviewed volume they dump on you to get there.
Match the model to your firm, not the hype to your fear
For a solo or small firm, the workable options are the ones that publish a price and do not need a long setup: Paxton, CoCounsel Essentials, LawDroid. Be careful with security posture at the budget end. The pricier names in this tier claim SOC 2 Type II; the cheaper ones sometimes don't, and that gap is easy to miss when you are comparing monthly fees. It matters the moment you handle medical records. Ask any vendor for the attestation itself rather than inferring it from a privacy policy, which is the wrong document to check.
For personal-injury and plaintiff firms, the field is deeper because the money is: Eve raised 103 million dollars at a billion-plus valuation in September 2025, Supio is built for mass torts, and EvenUp now offers both software and the managed PLAAS option. This is the one segment where you can genuinely choose between “give my paralegals a faster tool” and “hire a service that delivers reviewed demand packages.” Price both against what a demand backlog actually costs you.
For large firms, Harvey and Legora dominate, and both are enterprise animals. Harvey’s onboarding runs for months with vendor engineers embedded in your firm; it is not something you switch on over a weekend. Legora’s pricing is reported, though not published or confirmed, at around 3,000 dollars per seat per year with a ten-seat minimum; treat that as directional. Neither is realistic for a small practice, and both know it.
Five things to check before you sign
You do not need to be technical to run a good diligence pass. Ask five questions and make the vendor answer in writing.
First, security you can verify. Ask for the SOC 2 Type II report and its date, not a logo on the website. StrongSuit, the platform formerly called Callidus, claims on its site to be “100% compliant with every U.S. state’s ethical rules as of March 2025.” Ask which bar issued that certification. A claim like that says something about marketing, not controls. A real attestation has an auditor and a date.
Second, whether it connects to your systems: Paxton is a capable assistant, but as of mid-2026, it integrates with no practice-management system at all, which means its work product does not flow into Clio or Filevine. That is fine if you paste output by hand, but it's a real constraint if you expect matter management. Eve shows you the subtler version of the trap: a G2 reviewer wanted Eve’s output to carry into Litify, and Litify is not on Eve’s published integration list at all. The lesson: ask for the specific integration in writing, not the category.
Third, pricing you can actually see. If a vendor will not give you a number without a demo, assume seat minimums and annual lock-in until proven otherwise, and get the renewal terms before you sign the first term.
Fourth, whether the vendor will still exist is not paranoia. Robin AI was a well-known contract-review company. Per trade-press reporting: layoffs through the autumn of 2025, a winding-up petition filed by HMRC at the start of November, Scissero taking the managed arm and roughly 75 staff in December, and Microsoft hiring about 18 of the remaining engineers in January 2026. The assets were split up and nobody bought the software product. The site still markets it. Check funding, check recent trade-press coverage, and do not take a polished homepage as evidence the company is healthy.
Fifth, the supervision fit, which loops back to where we started. Get the vendor to state plainly who reviews the output and where your sign-off sits. If the answer is “your staff reviews everything,” price in that labor realistically. The Stanford RegLab’s 2024 benchmark, still the most-cited number in this argument, found leading legal-research tools returning incorrect information more than 17% of the time for some tools and more than 34% for others. The sanctions are real and much more recent: in United States v. Farris, decided by the Sixth Circuit in April 2026, a brief drafted with CoCounsel’s help reached the court with fabricated quotations attributed to real cases. The attorney was denied his court-appointed compensation, referred for discipline, and removed from the case. The tool was not the villain there. Insufficient review was. Whatever you buy, budget for the checking.
Current view, subject to change
My view today is that for most firms with about twenty lawyers or fewer, the managed model is the better fit, because those firms have the least slack to absorb a surge of unreviewed drafts, and review is exactly what the ethics rules will not let them skip. I hold that loosely. What would change my mind is evidence that self-serve tools have gotten reliable enough that sampling-based review (which Opinion 512 does allow for previously tested tools) is genuinely safe at volume. If hallucination rates on routine paralegal tasks fall into the low single digits and stay there, the calculus shifts back toward buying a tool and reviewing a sample. We are not there in August 2026, but it is a measurable question, not a matter of faith.
What moved while I was writing this
Four things landed in the month between drafting and publishing, and they all point the same way. Harvey is reported to be raising at a $15.5 billion valuation, up from $11 billion in March, and acquired Benchmark in July, its third acquisition of the year. Legora acquired Wexler at the end of July, its fifth. CoCounsel’s rebuilt agentic version reaches general availability in the United States this month. None of that changes the advice above, but it tells you the top of this market is consolidating fast, which is worth knowing before you sign a multi-year enterprise term with any of them.
Final thoughts
The AI paralegal market looks like a feature race and is actually an accountability question wearing a feature race as a costume. Every vendor can build a medical chronology now. Very few will tell you clearly who is responsible for it being right. That is the line worth watching over the next year, because it is the line the bar regulators are already drawing, and it is the one your malpractice carrier will care about long after the demo ends.
For what it is worth: my company, Coworkers.Global, builds a managed AI paralegal, which is the model I have argued for above, so read my conclusion with that thumb on the scale. If you want to compare a managed option against the tools in this guide for your own firm, email me at cs@coworkers.global and I will walk you through where it fits and, just as usefully, where it does not.
Regards,
Charles Stack
Founder, Coworkers.Global
This article discusses general concepts, not your situation. For that, talk to a lawyer.