If you have been practicing for more than a few years, you have watched at least one technology wave arrive with big promises and leave behind a modest feature. Legal artificial intelligence is different in one important way: the underlying models can now read, summarize, and draft ordinary legal prose at a level that is genuinely useful on real matters. That does not make the marketing around it honest, and it does not make every tool safe to use with client data. This guide explains what the technology actually does, where the time savings are real, where the risks are real, and how to evaluate tools without taking a vendor's word for anything.
What legal AI actually does in 2026
Strip away the branding and nearly every legal AI product on the market performs some combination of five jobs. Each one follows the same pattern: the software produces a first pass, and an attorney reviews, corrects, and approves it.
| Job | What the AI produces | What stays with the attorney |
|---|---|---|
| Document drafting | First drafts of demand letters, motions, discovery requests, and correspondence, built from the facts in the file | Strategy, tone, legal theory, every edit, and the signature |
| Legal research | Research memos that summarize authority and surface citations for the attorney to verify | Reading the cases, confirming they exist and say what the memo claims |
| Deadline tracking | Dates extracted from filings and orders, placed on a matter timeline with reminders | Confirming computed dates against the applicable rules |
| Client intake | Structured summaries of intake conversations and documents, plus draft status updates | The engagement decision, conflicts, and everything sent to the client |
| Contract review | Clause-by-clause flags: missing terms, unusual provisions, deviations from a playbook | The judgment call on every flag, and the negotiation |
The interesting product question in 2026 is not whether a tool can do these jobs. Most credible tools can, to varying standards. The question is where the tool sits. Research assistants like CoCounsel and Lexis+ AI live outside your matters and answer questions when asked. Practice management suites like Clio have added chat-style AI to systems built for record keeping. A newer category, the one we are building Caseagent in, puts the agent inside legal case management software itself, so the AI works from the actual file rather than from whatever you paste into a chat box.
How it works under the hood
You do not need a computer science degree to evaluate legal AI, but three concepts explain almost everything a vendor will show you.
Large language models
A large language model (LLM) is a statistical model trained on enormous amounts of text. Given a prompt, it predicts the most plausible continuation, one token at a time. That is why LLMs are remarkably good at legal prose, which is dense with patterns, and why they can also produce confident text that is simply wrong. The model is optimizing for plausibility, not truth. Everything else in a well-built legal AI product exists to compensate for that fact.
Retrieval, or grounding the model in real documents
Retrieval-augmented generation (RAG) means the software first searches a trusted source, your case file, a case law database, your firm's precedent bank, and then hands the retrieved passages to the model with instructions to answer from those passages only. Retrieval is the difference between "write me a demand letter" and "write me a demand letter from this police report, these medical records, and this repair estimate." When a vendor says their tool is grounded or cites its sources, they are describing retrieval. It reduces fabrication substantially; it does not eliminate the need to verify.
Agents, or models that take steps
An agent is an LLM wrapped in a loop: it can plan a task, use tools (search a database, read a document, add a calendar entry), look at the result, and decide the next step. A chat assistant answers one question. An agent can notice that a new filing arrived, read it, update the matter chronology, extract the response deadline, and queue a draft, then stop and wait for a human to review. The leash matters: in legal work, a well-designed agent proposes and a lawyer disposes.
When you sit through a vendor demo, translate the pitch into these three parts. "Trained on legal data" describes the model. "Cites its sources" or "works from your documents" describes retrieval. "Handles the workflow end to end" describes an agent, and should prompt your sharpest question: where exactly does a human review before anything leaves the system?
Where it saves time
The honest framing is this: industry research, including Clio's Legal Trends reports, has consistently found that lawyers bill only around a third of their working day, with the rest lost to administrative work. Legal AI attacks specific slices of that unbilled time. Where it helps most:
- First drafts of routine documents. A demand letter or a set of standard discovery requests that would take a working session to draft from scratch can start as an AI first pass in minutes. The attorney's time shifts from producing text to editing it, which is faster for routine work.
- Getting current on a file. Summarizing a stack of medical records, a deposition transcript, or months of correspondence is exactly the reading-heavy, judgment-light work models handle well. See how this works in practice on our legal research software page.
- Research triage. AI research tools are fast at the first hour of research: framing the issue, surfacing candidate authority, sketching the memo. The final hour, verifying and reading the controlling cases, remains human work.
- Deadline extraction and calendaring. Pulling dates out of orders and filings, computing response windows, and putting them on a watched timeline is mechanical work that software should have absorbed years ago.
- Client communication drafts. Status updates and intake follow-ups are high-volume, low-complexity writing: ideal first-draft territory.
Be skeptical of any vendor quoting precise hours-saved numbers. Savings depend on practice area, matter mix, and how much routine drafting your week actually contains. The realistic claim is directional: tasks that used to take a working session start as a reviewable draft in minutes, and reading-heavy catch-up work compresses sharply.
It is equally worth naming where the savings are small. Bespoke appellate argument, novel legal theories, negotiation, counseling a frightened client, and anything that turns on judgment rather than production sees little direct speedup. AI compresses the work around the judgment so that more of your week is spent exercising it. If a vendor claims their tool improves the judgment itself, close the tab.
Real limits and risks
Hallucinated citations
The best-known failure mode in legal AI is the fabricated citation: a case name, reporter cite, and persuasive quote that do not exist. Courts in multiple jurisdictions have sanctioned or disciplined lawyers who filed briefs containing AI-invented authority, and those episodes have become the cautionary tale of the entire category. The lesson is not "never use AI." It is that a general-purpose chatbot is the wrong tool for research, and that no citation goes into a filing until a human has pulled and read the authority. Purpose-built research tools reduce the risk through retrieval and explicit verification flags; none remove the duty to check.
Confidentiality
Pasting client facts into a consumer chatbot may expose confidential information to systems that retain prompts or use them for model training. Before any tool touches client data, you need clear answers, in writing, on where data is processed, how long it is retained, whether it is used for training (the answer must be no), and what happens on deletion. Bar associations that have issued guidance on generative AI consistently center the duty of confidentiality and the duty of technological competence.
Unauthorized practice of law
Legal AI tools are built for legal professionals. Software that drafts documents or summarizes authority is not licensed, cannot exercise judgment, and cannot form an attorney-client relationship. A firm that lets AI output reach clients or courts without attorney review is not automating practice; it is abandoning it. Every serious vendor in this category, ourselves included, states this plainly: the output is a drafting and research aid that a licensed attorney reviews. It is not legal advice.
Quieter limits worth knowing
- Models are weakest on novel legal questions, cross-jurisdictional nuance, and very recent developments.
- Long, messy files can exceed what a model attends to well; good products chunk and index rather than stuffing everything into one prompt.
- Output quality tracks input quality. A thin file produces a thin draft.
How to adopt it at your firm
Adoption fails when a firm buys a tool and announces it. It works when a firm picks one workflow, measures it, and expands from evidence. A practical sequence:
- Pick one high-volume, low-risk workflow. Demand letters in a PI practice, intake summaries, or discovery request drafting. Not appellate briefs.
- Run the vendor's own demo with skepticism. Bring a sanitized fact pattern from your practice. If a vendor will not let you try before you commit, keep moving.
- Clear the confidentiality questions first. Data processing location, retention, training use, deletion, and a DPA if you need one. In writing.
- Write a one-page AI use policy. Which tools are approved, what data may enter them, and the non-negotiable: attorney review before anything leaves the firm.
- Pilot with two or three people for a month. Track drafts produced, edits required, and whether reviewers trusted the output enough to keep using it.
- Decide on evidence, then expand. If the pilot saved real time, roll the workflow out and pick the next one. If it did not, you spent a month, not a budget.
For a broader look at the tool landscape and what each category costs, see our overview of AI for lawyers.