Paxton AI Review

Legal AI has a credibility problem it earned honestly, and everyone in the profession knows the shape of it.

Lawyers have been sanctioned for filing briefs containing citations to cases that do not exist, produced by tools that generated something plausible and were believed. That happened repeatedly, publicly, and it set the standard by which every legal AI product is now judged.

Which makes the first question about any of them the same: where do the citations come from, and how would you know if one were invented. Paxton's site does not answer it, and that absence matters more than any feature.

Paxton AI For Small Firms, In Brief

• An all-in-one legal assistant covering research, document analysis, drafting and what it calls thought partnership.
• Medical chronologies and billing summaries — the strongest feature, aimed squarely at personal injury practice.
• Five practice areas: personal injury, family, employment, defence and corporate.
• It compares itself to Westlaw and LexisNexis directly, alongside Harvey, CaseText and vLex.
• SOC 2, ISO adherence and HIPAA stated, which the medical records work makes necessary rather than optional.

Which Sources Does A Legal AI Actually Search?

Paxton describes contextual research drawing on a vast knowledge database, and does not name a single legal source, database or jurisdiction coverage on the page we read.

That omission is significant rather than pedantic, because in legal research the corpus is the product. A tool searching a licensed, updated case-law database is doing something categorically different from a tool searching whatever was in a model's training data.

The distinction shows up in two ways that matter. Currency: case law changes, and a decision from last month either exists in your corpus or does not. And completeness: knowing that a search covered all federal appellate decisions is what lets you rely on the absence of a case.

It also determines the fabrication risk. A tool retrieving from a real database and citing what it found can be wrong about relevance; a tool generating a citation from a model's memory can be wrong about existence, and those are not comparable failures.

None of this means Paxton fabricates anything — we have no evidence either way, and the site does not claim a fabrication rate to test. It means the question is unanswered on the marketing pages, and it is the first question a firm should put to the vendor before anything else.

Contrast that with Blue J, which built its reputation on grounding answers in primary tax authority and says so plainly. Naming the corpus is itself a form of evidence.

A tool that retrieves from a real database can be wrong about relevance. A tool generating a citation from memory can be wrong about existence.

Medical Chronologies Are The Sharpest Feature

Paxton lists Medical Chronologies and Medical Billing Summaries, and these are the most defensible things in the product.

A medical chronology is a dated, ordered account of a plaintiff's treatment, assembled from records that arrive as hundreds of pages of scanned notes from multiple providers in no order, in handwriting, with duplicates.

Building one is a paralegal's least favourite task and it is entirely mechanical: extract dates, providers, complaints, findings and treatments, then sort. That is close to an ideal application for a language model, and the time saving is enormous rather than marginal.

It is also lower-risk than research. A chronology is verifiable against source pages, and an error is a wrong date you can check, not a case that never existed. The work product is checkable in a way a research memo is not.

Billing summaries follow the same logic and touch damages directly. Both explain why HIPAA compliance appears in the certification list — this feature involves protected health information by definition, and any firm using it inherits that obligation.

Comparing Itself To Westlaw Is A Statement

Paxton's footer includes a Compare section naming Westlaw Precision, LexisNexis, Harvey, CaseText and vLex. That set is worth reading carefully.

Westlaw and LexisNexis are the incumbent research platforms, with decades of editorial infrastructure — headnotes, citators, classification systems — behind them. Naming them is a confident claim about substitutability from a young product.

Harvey and CaseText are the AI-native comparison, and vLex sits between the two. Listing all five says the product wants to be evaluated both as a research replacement and as an AI assistant, which are different purchases with different buyers.

The realistic read is about firm size. A large firm is not dropping Westlaw, because its citator is a professional-liability control as much as a research tool. A small firm paying for one platform is exactly who a challenger can win.

That is also who benefits most. A five-lawyer personal injury practice cannot afford enterprise legal research, and an assistant that handles research, drafting and chronologies for one subscription changes what that practice can take on.

For plaintiff-side work the overlap with Eve is worth mapping, and for drafting inside Word rather than researching, Spellbook is the narrower comparison.

• Ask which databases are searched and how current they are, before anything else.
• Verify every citation independently — no legal AI removes that obligation.
• Start with medical chronologies, where output is checkable against source pages.
• Treat HIPAA as your obligation too, not only the vendor's, once records are uploaded.
• Test on matters you already closed, where you know what the research should find.

The Citation Question Nobody Should Skip

This deserves its own section because it is the single professional risk that matters, and it is not hypothetical.

Lawyers in multiple jurisdictions have faced sanctions for filing documents citing fabricated authority. In every reported instance the pattern was identical: the tool produced a citation that looked entirely correct, and nobody checked it before filing.

The uncomfortable part is that fabricated citations are convincing. They have plausible case names, plausible reporters, plausible years and plausible pin cites, because the model has learned the shape of a citation extremely well. Nothing about the format signals the problem.

The only reliable defence is verification in a source you trust, every time, without exception. A citator check is not optional diligence — it is the control that stands between an AI research tool and a bar complaint.

That obligation exists regardless of vendor, and no product removes it. What a vendor can do is make it easier by linking directly to the retrieved source, and whether Paxton does that is something to test in a trial rather than assume from the marketing.

This is true of every vendor in the category, Harvey AI included. The obligation follows the lawyer, not the software licence.

Testing Legal Research Against Work You Already Did

The right evaluation for legal AI is retrospective, because you already know the answers.

Run it on matters you closed, where the research memo exists and the outcome is known.

1. Take five closed matters across your actual practice areas, including one that turned on an unusual authority.
2. Ask the same research questions you asked at the time, without hinting at what you found.
3. Verify every citation it returns in a source you trust, and record any that do not exist.
4. Check what it missed as carefully as what it found, since omission is the quieter failure.
5. Run a medical chronology against one you built by hand, and diff the dates and providers.
6. Ask the vendor directly which databases are searched and how often they update, before signing anything.