AI Legal Assistant: What a Legal AI Assistant and an AI Paralegal Actually Do

An AI legal assistant is software that does defined pieces of legal work on instruction: finding authority, summarizing documents, extracting terms, drafting first passes and answering questions about a matter file. It is not a licensed professional and it does not carry the responsibility a lawyer or paralegal carries. The useful way to judge one is by the task it completes and whether its output can be checked, not by the label on the product page.

The phrase has become a category name without ever getting a definition, which is a problem when you are trying to buy one. Vendors apply it to a chat box on a research platform, to a document review engine, to an intake chatbot on a firm website, and to an agent that runs a multi-step workflow across a document set. Those are four different products with four different failure modes and four very different prices.

This is a practical guide for a US firm working out what the term means, what these tools genuinely do today, where they break, and what to ask before you sign anything. It is informational research guidance, not legal advice.

What is an AI legal assistant?

An AI legal assistant is a software tool that uses large language models to perform legal support tasks you would otherwise assign to a person: locating relevant case law or statutes, summarizing an opinion or a contract, pulling dates and obligations out of a document set, drafting a first version of a routine document, and answering questions about material you have given it. It works from a prompt in ordinary language rather than from a query syntax, which is the main thing that separates it from the research software firms have used for thirty years.

What it is not is a substitute for professional judgment or professional responsibility. Under ABA Formal Opinion 512, issued 29 July 2024, the existing Model Rules already govern generative AI use. Nothing in that opinion creates a new regime, and nothing in it shifts responsibility onto a vendor. Competence under Rule 1.1, confidentiality under Rule 1.6, supervision under Rules 5.1 and 5.3 and candor under Rule 3.3 all stay exactly where they were. The assistant does the task. You own the output.

Is there an AI legal assistant that actually works?

Yes, for narrow, well-defined tasks with checkable output. The tools that hold up in practice share one trait: they retrieve from a real body of material and hand you the source, so verification takes seconds. The tools that fail share the opposite trait: they compose an answer that reads like research and give you nothing to open.

That distinction matters more than any feature list, and the evidence is not subtle. Damien Charlotin's AI Hallucination Cases database, which only records decisions where a court explicitly found or clearly implied reliance on fabricated material, listed 1,890 cases as of its 15 August 2026 update. In January 2026 it held 719. That is not a rounding error in an emerging technology, it is a curve, and every case on it involves a filing that went out without anyone opening the authority it cited.

So the working test for any assistant is simple. Ask it something you already know the answer to, then try to verify what it tells you. If checking the answer takes longer than doing the task yourself, the tool is not saving you time, it is moving the work downstream and disguising it.

What can an AI legal assistant do?

Sorted roughly by how reliable the output is today.

• Find on-point authority. The most mature use. Ask a legal question in plain English and get cases with the holding and a citation you can open. Reliability depends entirely on whether the tool searches an actual corpus of decisions or generates plausible-looking citations.
• Summarize a document. Opinions, contracts, deposition transcripts, complaints. Strong, because the source text is right there and a summary is checkable against it in a minute.
• Extract structured data. Dates, parties, obligations, renewal terms, governing law. Strong for the same reason, and the highest-volume time saver in transactional practice.
• Answer questions about your own files. Retrieval over a matter file or document set. Good, and the closest thing to what people picture when they hear the word assistant.
• Draft a first pass. A routine clause, a research memo skeleton, a client update. Useful as a starting point, and never as a finished product.
• Check whether a case is still good law. Weakest. This is what a citator does, and citators are curated editorial products rather than model outputs. Very few AI assistants have one, and most do not tell you that.

Notice how much of that list is document work rather than research. That reflects where the technology is genuinely strong, and it is also why the market has split between research tools and platform tools. Buying the wrong side of that split is the most expensive mistake in this category.

AI legal assistant vs AI paralegal: is there a difference?

Commercially, mostly not. The two phrases are used interchangeably by vendors, and no product meaningfully described as an AI paralegal does something an AI legal assistant does not. If a vendor uses the paralegal framing, treat it as positioning rather than as a capability claim, and ask which specific tasks the software completes end to end.

The comparison is worth making for a different reason, though, because it tells you where the savings really are. Ask any paralegal what eats their week and research is rarely the top answer. It is chasing documents, tracking deadlines, coordinating filings and moving work between people, and the tooling for that is workflow software rather than a language model. Getting a new matter to the right person on the right day is a task routing problem, not a research one, and no assistant that reads case law will touch it. Firms that buy an AI assistant expecting it to solve the coordination problem are disappointed for reasons that have nothing to do with the model.

Will AI replace paralegals and legal assistants?

Not on current evidence, and the reason is structural rather than sentimental. These tools produce output that a responsible person has to check, and checking is itself skilled work. Someone still has to know that the summary missed the carve-out, that the case retrieved was overruled in part, or that the extracted date is the notice date rather than the effective date. That someone is usually a paralegal.

What is changing is the mix. The mechanical first pass compresses, and the review and judgment portion expands to fill it. The firms getting real value are the ones treating an assistant as leverage on a person rather than as a replacement for one, because the alternative concentrates all the verification burden on whoever signs the filing.

How to use AI as a legal assistant without creating a problem

Four things separate the firms doing this well from the ones generating incident reports.

• Verify every citation, every time. Open the case. Read the passage the tool relied on. This is not optional and it is not slow once it is habit. In Whiting v. City of Athens, Tennessee, the Sixth Circuit imposed appellees' fees, double costs and 15,000 dollars in punitive sanctions each under Rule 38 over a brief with more than two dozen fabricated citations. Courts have not sanctioned anyone for using AI. They have sanctioned filings nobody read.
• Get client confidentiality right first. ABA Opinion 512 says lawyers should generally obtain informed client consent before entering client information into a generative tool, and specifically that boilerplate consent buried in an engagement letter is not adequate. Ask any vendor whether your inputs train their models, where data is processed, and what retention applies. Get it in writing.
• Check disclosure rules in your courts. There is no blanket federal rule on disclosing AI use in filings. It lives at the chambers level in individual judges' standing orders, which means the answer is different from courtroom to courtroom and you have to look it up for the judge you are in front of.
• Supervise it like a junior. Rules 5.1 and 5.3 cover supervision of both lawyers and non-lawyer assistance, and the profession has settled on reading them to cover AI output too. Nobody would file a first-year associate's memo unread. The same standard applies here, and for the same reason.

What to ask before you buy one

Most of the disappointment in this category comes from buying a platform to solve a task, so start by naming the task. If the bottleneck is finding on-point authority and reading the holding, that is one job, and it is the corner of this market where per-seat pricing is actually published. If the bottleneck is document review across thousands of files, that is a different product at a different price. If the bottleneck is coordination, no assistant is the answer.

Then ask five questions. Does it retrieve from a real corpus, or generate? Can I open the source in one click? Does it include a citator, and if not, does the vendor say so plainly? What is the price per seat per month, in writing, including renewal terms? And can I test it on my own questions before committing? A vendor that answers all five without a discovery call is telling you something useful about how the product is built.

We compare the platforms that market themselves this way on legal AI for lawyers, sorted by task rather than by vendor category, with what each one's output can actually be used for. The buying guidance for firms without a procurement department sits in AI for small law firms, the ethics detail in can lawyers use AI for legal research, and the sanctions record in AI hallucinations in legal cases. If what you actually need is the research piece, Cases answers a plain English legal question with on-point US federal and state precedents, the holding and a real citation, at 29 dollars per user per month with the price on the page.

Hallucination case count read from Damien Charlotin's AI Hallucination Cases database on 16 August 2026, showing 1,890 cases and a last-updated date of 15 August 2026. Informational research, not legal advice, and always verify the citation.