Spend an afternoon at a legal-tech trade show and you will see the label "AI legal assistant" printed on almost every booth. It sits on research platforms, contract tools, practice-management add-ons, and consumer apps that promise to explain your lease. Some of those products are carefully engineered systems built around verified sources. Others are a general chatbot with a gavel icon on top. From the outside, they look the same.
This piece is an attempt to separate the two. It explains what an AI legal assistant actually is, what happens inside one when you type a question, what a realistic working day with one looks like, where these tools break, and what to ask before you pay for one. You don't need a law degree or an engineering background to follow along.
TL;DR: An AI legal assistant is software that combines a large language model with a search step over legal sources, so its answers can point back to real cases, statutes, and documents. It is good at drafting, summarizing, and finding starting points; it is bad at being trusted blindly. Products such as LegesGPT are one example of the category: they answer legal questions with citations you can open, review contracts, and draft documents.
The short definition, and what the label hides
An AI legal assistant is a piece of software that takes requests written in ordinary language ("does this indemnity clause cover third-party claims?") and returns legal work product: an answer with sources, a summary, a draft, a marked-up contract, or a list of risks.
That definition is broad on purpose, because the market is broad. Under the single label you will find at least three different kinds of product:
- Research-first assistants that answer questions about the law and cite the authorities they relied on.
- Document-first assistants that read, review, compare, and draft contracts and other files you upload.
- Workflow assistants bolted onto practice-management or e-discovery software, which handle things like time entries, intake summaries, or batch document tagging.
Many products now do some of each. The more useful question is not "is this an AI legal assistant?" but "what is this one grounded in, and what is it allowed to touch?" The rest of this article is about answering that.
How an AI legal assistant works under the hood
The engine inside nearly every AI legal assistant is a large language model, the same family of technology that powers general chatbots. A language model is extremely good at producing fluent text that fits a pattern. On its own, it has no built-in sense of whether a sentence is true. That single fact explains most of the good and most of the bad in this category.
What separates a purpose-built assistant from a plain chatbot is what happens around the model. A typical request moves through three stages.
Stage one: understanding the question
When you type a question, the system first works out what you're asking for. Is it a question about a rule, a request to draft, or a request to review an attached file? Which jurisdiction is implied? Some tools ask you to pick a country or state; better ones infer it and show you the assumption so you can correct it.
Stage two: fetching sources before answering
This is the step that matters most. Before the model writes a single word of the answer, the system runs a search over a body of material: case law, statutes, regulations, the contract you uploaded, or a firm's own precedent files. Engineers call this retrieval-augmented generation, usually shortened to RAG. The plain-English version: the assistant looks things up first, then writes.
A useful comparison is an open-book exam. A plain chatbot sits the test from memory. A retrieval-based assistant is allowed to bring the textbook, and it is expected to show which page it used.
Stage three: writing an answer tied to what it found
The model then writes a response using the retrieved passages as its raw material, and attaches citations that link back to those passages. In a well-built tool, clicking a citation opens the actual case or statute, ideally scrolled to the relevant paragraph.
Retrieval reduces errors but does not eliminate them. The search can pull the wrong case. The model can summarize a correct source incorrectly, or stitch two sources into a claim neither one makes. Researchers at Stanford's RegLab and HAI tested several commercial legal research tools in 2024 and found that even the retrieval-based products produced incorrect or misgrounded answers on at least one in six benchmark queries. That is far better than a general chatbot answering from memory, and still nowhere near good enough to skip checking.
So why does a plain chatbot behave differently?
A general-purpose chatbot, used without any connected sources, answers from patterns it absorbed during training. Ask it for a case on point and it will produce something that looks like a case name, with a citation format that looks correct, because that is what case citations look like. Whether the case exists is a separate matter the model has no way of checking. This is how fabricated authorities end up in court filings, which we'll come back to.
Some general chatbots now browse the web, which helps a little. But the open web is a poor substitute for a curated legal corpus: it mixes blog posts, outdated statutes, and foreign law, and the chatbot is not built to tell you which jurisdiction a source belongs to.
A day with an AI legal assistant
Abstract descriptions only go so far, so here is a composite workday for Dana, an in-house counsel at a regional logistics company with a legal team of two. Nothing here is exotic; this is roughly what the tools are used for today.
8:30 a.m. A carrier has sent back the company's master services agreement with forty tracked changes. Dana uploads both versions and asks the assistant to list every change that shifts risk toward her company. She gets back nine flagged edits, including a quiet expansion of the carrier's limitation of liability. Two of the nine turn out to be harmless on a closer read. The tool saved her the first pass, not the judgment.
10:00 a.m. The operations director asks whether drivers waiting at a customer's dock count as working time under state wage rules. Dana asks the assistant, which returns a short answer with links to the relevant regulation and two decisions. She opens both decisions. One is directly on point; the other is from a different state and gets discarded.
11:15 a.m. She needs a mutual NDA for a software vendor demo this afternoon. The assistant drafts one from a template, adjusted for a two-year term and her state's governing law. She edits the definition of confidential information and sends it out.
1:30 p.m. A 60-page warehouse lease renewal arrives. Dana asks for a one-page summary of rent escalations, renewal options, and termination rights. The summary is accurate except that it misses a rent abatement clause buried in an exhibit, which she catches because she knows that landlord likes exhibits.
3:00 p.m. For a board memo, she runs a longer research request on recent changes to freight broker liability. The assistant takes a few minutes, returns a structured memo with sources, and she spends forty minutes verifying and rewriting it in her own voice.
4:30 p.m. She asks the assistant to turn her rough notes from a vendor call into a follow-up email. It's a small task, but it adds up across a week.
The pattern across the day: the assistant produces fast first drafts and fast first reads. Dana still makes every decision, checks every authority she relies on, and catches the things that require knowing the people and the history.
AI legal assistant vs. chatbot vs. paralegal vs. research database
People often ask which of these an AI legal assistant replaces. The honest answer is that it overlaps with all of them and fully replaces none. Here is how they compare on the things buyers tend to care about.
| AI legal assistant | General chatbot | Human paralegal | Legal research database | |
|---|---|---|---|---|
| Main strength | Fast drafts, summaries, and sourced answers in plain language | Broad writing help on any subject | Judgment, client context, follow-through | Authoritative, complete primary law with citators |
| Where answers come from | Retrieval over legal sources and your documents | Training data, sometimes the open web | Experience, firm files, research tools | Curated, editor-maintained collections |
| Citations | Linked to sources; must still be checked | Often missing or invented | Checked by a person | Exact, with treatment history |
| Speed | Seconds to minutes | Seconds | Hours to days | Fast to search, slow to synthesize |
| Confidentiality | Depends on vendor terms; varies widely | Consumer terms may allow training on inputs | Bound by duty and supervision | Generally strong; you are mostly searching, not uploading |
| Typical cost | Roughly $20 to several hundred dollars per user per month | Free to about $20 per month | Salary or hourly rate | Often hundreds per user per month or more |
| Can it give legal advice? | No, it supports the person who does | No | No, works under a lawyer | No, it's a library |
The column that surprises people is the last one. A research database is the most reliable source in the table, but it does very little synthesis; you still read and write. An AI legal assistant does the synthesis, but it sits on top of sources it may or may not have complete access to. Many teams end up using both: the assistant for the first draft, the database (or the primary source itself) to confirm the authorities that matter. If you are comparing the database side of that pairing, our roundup of the main legal research databases covers the options.
Where AI legal assistants fail
Every product in this category can fail in the ways below. The good ones make failures easier to spot; none make them impossible.
Citations that don't exist, or don't say what's claimed
The most famous failure is the invented case. In Mata v. Avianca (S.D.N.Y. 2023), lawyers filed a brief citing decisions that a general chatbot had made up, and the court sanctioned them. Similar incidents have kept appearing in courts since, including filings by experienced firms.
The quieter failure is more dangerous: a real case, correctly cited, that does not actually support the proposition attached to it. A link that resolves gives a false sense of safety. The fix is the same in both cases: open the source and read the relevant passage before relying on it. A tool that makes this one click away is doing its job. A tool that lists citations without links is asking you to trust it.
Confidentiality and where your documents go
Pasting a client's contract into a consumer chatbot may mean sending it to a company whose terms allow inputs to be stored or used to train future models. For lawyers, this runs straight into professional duties of confidentiality. The American Bar Association's Formal Opinion 512, issued in 2024, tells lawyers to understand how a generative AI tool handles client information before using it, and in some cases to get informed client consent.
For buyers, this means reading the data terms, not the marketing page. Look for a clear statement that your uploads are not used for model training, information on how long files are kept, and whether you can delete them.
Unauthorized practice of law
An AI legal assistant is not a lawyer and cannot form an attorney-client relationship. When these tools are sold directly to consumers, regulators pay attention to how they are described. In the United States, the Federal Trade Commission took action against DoNotPay over its "robot lawyer" marketing, which ended in a settlement the agency finalized in 2025.
In practice, the line works like this. A tool that explains how a law works, finds sources, or helps you draft a document is offering legal information and assistance. Deciding what you specifically should do in your situation, and standing behind that decision, is legal advice, and that stays with a licensed professional. Reputable tools say so plainly; be wary of any product that implies otherwise.
Old law, wrong place
Law changes, and different places have different rules. An assistant can return an answer based on a superseded statute, or quietly apply California law to a question about Texas. Check the date and jurisdiction of every source that matters, and tell the tool your jurisdiction up front rather than hoping it guesses.
Four myths worth retiring
"It will replace lawyers." It replaces some hours, mostly the first-draft and first-read hours. The judgment, accountability, and client relationship stay with people. Teams that adopt these tools well tend to redeploy time, not headcount.
"If it has citations, it's accurate." Citations are a way to check the work, not proof the work is right. See the Stanford numbers above.
"It's just ChatGPT with a legal skin." Some products are. The ones worth paying for add retrieval over legal sources, document handling, and data terms suited to confidential work. The difference shows up immediately when you click a citation.
"Only big firms can afford it." Enterprise products do carry enterprise prices, but there are now credible options priced for solo practitioners, small businesses, and students.
A buyer's checklist for AI legal assistant software
If you're evaluating AI legal assistant software, run every candidate through the same set of questions. Use a real (but non-confidential or anonymized) matter from your own work rather than the vendor's demo script.
- Grounding. What sources does it search before answering? Case law, statutes, regulations, your uploaded files? For which countries and states?
- Citation quality. Does every factual legal claim carry a citation? Does clicking it open the source at the right passage? Try ten citations and count how many hold up.
- Jurisdiction handling. Can you set the jurisdiction, and does the tool tell you when it's unsure?
- Data terms. Are uploads used for training? Where are they stored, for how long, and can you delete them? Get answers in writing.
- Document work. Can it review a contract against your preferred positions, propose edits, and compare versions? Test it on a document you know well.
- Drafting. Does it draft from templates or from scratch, and can you control style and governing law?
- Where it lives. Is it a web app, a Word add-in, a mobile app, or an integration with software you already use? Pick the shape that fits how you work.
- Price and limits. Is pricing per user or flat? Are there caps on document reviews or research queries? Is there a trial, and what does it cost?
- Honesty about limits. Does the product tell you what it can't do? Vendors that claim zero hallucinations are telling you something about their marketing, not their software.
If you want a ranked view of specific products after running this checklist, we keep a separate comparison of AI legal assistants for lawyers. For research-heavy needs, the list of AI legal research tools goes deeper, and if contract review is the main job, see the roundup of AI contract review tools.
One example of the category: LegesGPT
To make the abstract concrete, here is how one product maps onto the pattern described above. LegesGPT is a web-based AI legal assistant aimed at lawyers, small businesses, and individuals. It answers legal questions with verified citations and links to the underlying sources, searches case law and statutes, drafts contracts and other documents, and reviews uploaded files to flag risks and suggest changes. It also includes e-signature, a library of more than 100 templates, and a Deep Research mode for longer, structured research requests like Dana's 3 p.m. memo.
It runs in the browser only, so there is no mobile app, Word add-in, or API, which matters if your team lives inside Word. Pricing is monthly: Basic at $19.99, Plus at $49.99 with 50 document reviews a month, and Premium at $99.99 with unlimited reviews and priority support. Annual billing saves 30%, and there is a 3-day trial for $1 rather than a free tier.
It's one implementation among many, and the checklist above applies to it the same way it applies to everything else.
Where the category is heading
The current direction is toward assistants that carry out multi-step tasks rather than single answers: research a question, draft a memo from it, then check the draft against the sources. That shift raises the stakes on everything in the risks section, because errors can compound across steps. If that area interests you, our overview of legal AI agents tracks the products moving that way.
The takeaway
An AI legal assistant is best understood as a fast, well-read junior that has never met your client and occasionally misremembers what it read. Grounded in good sources and checked by someone who knows the law, it can take hours out of a week. Trusted without checking, it can put a fabricated case in a court filing. The software is only part of what you're choosing. The other part is the review habit you build around it.
Frequently Asked Questions
What is an AI legal assistant in simple terms?
It is software that takes a legal request written in everyday language and returns work product such as a sourced answer, a summary, a draft, or a list of contract risks. The better ones look up legal sources before answering and link to them, so you can see where each claim came from.
How is a legal AI assistant different from ChatGPT?
A general chatbot usually answers from patterns it learned in training, which is why it can produce case names that sound real but do not exist. A legal AI assistant adds a search step over case law, statutes, and your own files, then ties its answer to what it found. That makes errors less frequent and much easier to check, though not impossible.
Can an AI legal assistant give legal advice?
No. It can explain how the law works, find authorities, and draft or review documents, but deciding what you should do in your specific situation is legal advice and belongs with a licensed lawyer. Tools that suggest otherwise deserve extra caution.
Are AI legal assistants accurate?
They are accurate enough to speed up research and drafting, not accurate enough to trust without review. Independent testing in 2024 found even retrieval-based legal tools got at least one in six benchmark questions wrong or poorly supported. Open and read every source you plan to rely on.
Is it safe to upload client documents to an AI legal assistant?
It depends entirely on the vendor's data terms. Before uploading anything confidential, confirm in writing that your files are not used to train models, how long they are stored, and whether you can delete them. Lawyers should also check their local ethics guidance on technology and client confidentiality.
Who uses AI legal assistants?
Lawyers and in-house counsel use them for first drafts, contract review, and research starting points. Paralegals use them to summarize and organize documents. Small business owners and individuals use consumer-facing versions to understand contracts and draft routine documents, ideally with a lawyer involved for anything high-stakes.
How much does AI legal assistant software cost?
Prices run from about $20 a month for tools aimed at individuals and small practices to several hundred dollars per user per month for enterprise products. LegesGPT, for example, starts at $19.99 a month with a 3-day trial for $1, while large-firm platforms are usually quote-based.
