Legal AI has moved past the chatbot. Most lawyers haven’t.

A practical guide to where legal AI stands at the end of 2026, and how to judge what comes next

Ask a room of lawyers what legal AI is and most will describe a chat window. You type a question, it answers, you check the answer. Two years ago that was roughly right. Today it describes a small corner of what these systems do, and the distance between what the tools can handle and what lawyers ask of them has become one of the more expensive misunderstandings in the profession.

This guide is for practitioners who want a straight account of where things stand. It covers what has changed in the technology, why that matters to the economics of a firm of any size, how specialist legal AI differs from the general assistants many lawyers already use, and what to ask before committing to the next phase.

1. Where legal AI is now

Most lawyers now use AI in some form. In the 8am 2026 Legal Industry Report, 69% of legal professionals said they personally use generative AI for work, up from 31% a year earlier.

How they use it tells a different story. The most common tasks were drafting correspondence (58%), general research (58%), brainstorming (54%) and summarising documents (47%). Analysing firm data or cases came in at 14%. Each of the top tasks is a single question with a single answer, which is exactly what a chatbot does well.

The products moved on some time ago. Through 2026, the leading legal AI platforms rebuilt their products around agents. An agent can now review thousands of documents against a list of questions and return the results as a table, or set out its plan for the lawyer to approve before it starts. Hand it a counterparty’s redline at the end of the day and it can work through the changes overnight, with a draft response waiting in the morning.

Very few firms are using this. The Thomson Reuters Institute’s 2026 report found that 15% of professional services organisations have adopted any agentic AI tool, and only 18% track the return on their AI spend. Clients feel the gap from their side. More than half of corporate legal departments want outside firms to use AI on their work, yet fewer than a third know whether their firms actually do.

Training explains a good part of this. In the 8am survey, 54% of firms offer no responsible AI training and have no plans to start. Most lawyers learned AI through consumer chat apps, so the chat window became their picture of what AI is. In fairness, many agentic features only reached general release during 2026, often in staged rollouts. If you tried legal AI in 2024 and decided it wasn’t for you, you were judging a different product.

2. From answering to doing: agents and the capacity question

What “agentic” actually means

The simplest way to see the difference is to ask who drives each step.

With a chatbot, you do. You write a prompt, read the output, decide what to ask next, paste in another document and go round again. The tool answers; you run the work.

With an agent, you set the goal and the system works out the steps. Ask it to prepare a chronology from a case file and it will find the relevant documents, pull out dates and events, flag where accounts conflict and produce a draft, pausing at the points you have set for review. The Thomson Reuters Institute describes agentic systems as completing multi-step tasks on their own, with people stepping in at pre-determined points. It adds a line worth keeping in mind: “More autonomous does not mean fully autonomous.”

That has a direct bearing on responsibility. An agent behaves more like a capable junior than a search engine. It does the groundwork, and you still read the draft before it goes to the client, as you would with any associate.

A rough test: if the task can be phrased as a question, a chatbot will cope. If it reads like an instruction you would give a trainee on Monday morning, it is agent work.

Large firms have always run on agents

Take away the technology vocabulary and an agent is a unit of delegated work. Large firms have bought that kind of capacity for decades, in the form of people.

Their structure is built for it. Across the Am Law 100 there were about 4.4 lawyers for every equity partner in 2023, and the ratio was rising. Behind those lawyers sit paralegals, trainees, knowledge management teams, document specialists and IT support. When a large firm takes on a major transaction, the partner’s judgement rests on hundreds of hours of work done by other people. That is how it handles volume and complexity at speed.

AI is going into those firms on top of the people, not instead of them. Thomson Reuters reports that law firm technology spending grew 9.7% in 2025 and knowledge management spending 10.5%, while lawyer headcount still rose 2.9%. In November 2026 A&O Shearman announced a USD 1.1 billion, five-year AI commitment and said 96% of its partners were already using its AI tools. Linklaters has rolled out a legal AI platform to just under 3,000 lawyers across 30 offices.

What changes for smaller firms and solo practitioners

Work is already moving towards smaller firms. In Above the Law’s 2025 survey of more than 500 in-house lawyers, 35% had moved work from large firms to midsize or smaller firms in the previous year, and 48% use those firms for complex, high-value work. Cost was the main reason. Responsiveness (48%) and personal attention (45%) came next. Clients go to smaller firms because they get the partner, rather than an account manager fronting a team of twenty.

What smaller firms have lacked is the capacity behind that partner. A solo practitioner can only draft so many agreements and review so many disclosure bundles in a day. Clio’s research shows how much of the day goes on work that is never billed: solo lawyers bill around 27% of their working day and small firms around 32%, compared with roughly half at mid-sized firms.

Agentic AI changes that sum. Work a large firm would hand to a trainee or a paralegal can now be delegated to software. For a boutique, that means more cases, or bigger ones, without putting layers between the client and the lawyer they chose. Billing models help. Clio found 75% of solo lawyers already use flat fees for some work, and on a fixed fee, time saved turns into margin and capacity instead of lost revenue.

So far the opportunity is mostly going unused. Clio’s 2026 research found that fewer than a third of solo and small firms have grown revenue through AI, against nearly 60% of the largest firms. Small firms use AI almost as widely as large ones; they just use it less deeply. The firms that close that gap will be the ones that treat AI as delegated capacity rather than a quicker way to search.

3. Why legal AI isn’t general AI with a legal plugin

What you can build yourself

General assistants have become very capable, and pretending otherwise helps nobody. In 2026 the four biggest AI companies all moved into legal work. Anthropic released a legal plugin for its Claude Cowork app in February and launched Claude for Legal in May, with connectors to LexisNexis, Thomson Reuters and iManage among others. Google previewed Gemini Enterprise for Legal in August, OpenAI launched Astra for Law in September, and Microsoft released Copilot for Legal in October. On the day Anthropic’s legal plugin was announced, Thomson Reuters shares fell as much as 18% in intraday trading.

On raw accuracy, the general models hold their own. In Vals AI’s October 2025 legal research study, ChatGPT matched the specialist legal tools tested, at around 80%.

A technically confident lawyer can therefore assemble a great deal. You can connect an assistant to your email and document storage, paste extracts from your legal database into a project, install plugins, write your own instructions, and build a deadline dashboard over a weekend of “vibe coding”. Some lawyers do exactly this and get real value from it.

What that costs you

It can be done. What matters for most firms is who does it, and what it takes to keep it running.

The first cost is that you become the legal engineer. Connecting tools, setting permissions, writing and testing prompts, checking everything still works after a model update: that is a job in itself. The vendors acknowledge as much. Analysts reviewing the Cowork legal plugins pointed out that they need technical configuration and enterprise contracts, and OpenAI has placed its own engineers inside large firms to build their custom tools. Most lawyers did not train for this, and every hour spent on it is an hour away from clients.

Then there is cost, which is harder to predict than it looks. General assistants are usually sold in tiers from about USD 20 a month to several hundred, with usage limits. Long-document legal work is heavy. When Harvey published the first results of its open Legal Agent Benchmark (LAB) in May 2026, the best-performing general model took about 22 minutes and roughly USD 51 in compute per task. We ran Ketrone through the same benchmark. It passed 90% of tasks in full, averaging 5.4 minutes and USD 1.10 per task, or USD 1.22 for each task completed to the rubric. Harvey’s figure comes from a different task set, so read it as a reference point rather than a like-for-like ratio, but the public leaderboard shows the same pattern: the models that score well on their own tend to be the slow, expensive ones. On usage-based pricing, the more seriously you use the tool, the more you pay and the harder it is to budget. A platform built for legal work, sold at a fixed subscription, puts the job of balancing speed, quality and cost on the vendor instead of on you.

The legal content also has to come from somewhere. The measurable gains come from grounding the model in curated legal sources. In the Vals study, raw accuracy was level, but the specialist tools beat ChatGPT by about six points on citing authoritative sources. OpenAI’s own testing found its new legal configuration passed 54% of a legal research benchmark, against 38.7% for the same model using web search alone. That is a real gain from adding legal sources, and it still leaves close to half the answers short of the mark. Build it yourself and you still need the database subscriptions, plus someone to wire them in. Check coverage while you are at it: OpenAI’s new legal research index covers US law only, and none of the four launches mentioned Gulf or Arabic-language primary law.

And someone has to maintain it. Models change every few months. A prompt that worked in spring may behave differently by autumn, and the rules you practise under keep moving. On a do-it-yourself setup, keeping up is your problem.

None of this changes the need to check the output, whichever route you take. Stanford researchers found in 2024 that specialist legal research tools gave incorrect information in 17% to 33% of queries, against 43% for a general model. Those tools have improved since, but courts have little patience for errors. Damien Charlotin’s database listed 2,312 court decisions worldwide involving AI-fabricated material as of November 2026. In Qatar, the QFC Civil and Commercial Court held a Dubai-based lawyer in contempt for citing cases that turned out to come from Google’s AI search.

What a purpose-built platform carries for you

Specialist platforms increasingly run on the same underlying models as the general assistants. What sets them apart is what is built around the model, often called the harness. In practice, that means:

  • legal reasoning built into the workflow, with tested instructions and checks for legal tasks that you don’t have to write yourself
  • primary law already connected and maintained, and cited in each answer
  • one defined workspace where your cases, documents, drafts and research sit together under one set of rules, rather than spread between an assistant and an inbox
  • a library of prompts and agent workflows that the vendor updates as models change and as clients raise new needs
  • support and security included in the service, with data commitments you can put in a contract

The harness is also what shows up in testing. On LAB, each task gives the agent a partner-style instruction and a real case file, and the work only counts if every criterion written by legal professionals is met, with no partial credit. On the public leaderboard, the strongest general model working on its own reached 26.7% on that measure. Ketrone passed 90% of 200 tasks across 25 practice areas, and every task in 18 of those areas. The model results come from a separate setup, and no benchmark covers every jurisdiction or language, so treat any score, ours included, as a starting point for your own testing. Our full run is open on GitHub, with every task, rubric and grade, so you can check the work behind the number.

Firm knowledge, memory and context

A general model knows what the world knows. It does not know your precedents, your house style, the clause you always negotiate out or what happened the last time you acted for this client. An AI system works from what it was trained on and what is put in front of it when you ask. Anything outside that, it fills in.

Picture a working desk. The context window is the desk itself: everything the AI can see while it works. Retrieval is the clerk who fetches the right files from your archive. Permissions are that clerk’s security clearance. Memory is the case file that stays on the desk between sessions, so you aren’t briefing the system from scratch every morning.

Get these right and the output starts to read like your firm’s work rather than a generic first draft. That is why the large vendors spent 2026 connecting their platforms to firms’ document management systems. In Ketrone, for example, a lawyer describes the task and the relevant precedent and history come back, limited to what that lawyer is allowed to see.

Permissions deserve more attention than they usually get. Eversheds Sutherland initially kept ethically walled cases out of its AI platform and only went firm-wide once the platform could enforce those walls. If a system can’t respect your conflicts and confidentiality rules on every request, your most important work stays out of it, and that defeats the purpose.

Two cautions. Grounding reduces errors without removing them: retrieval only helps if it finds the right, current document, so the state of your own files matters. And reasoning across several sources is still the hardest thing these systems do. In Vals AI’s latest legal research benchmark, questions that required reconciling multiple authorities scored lowest of all. Firm knowledge makes the output far more useful, but someone still has to read it.

4. What legal professionals need to understand next

Questions to ask before you buy

  1. What will it do without me prompting each step? Ask to see a multi-step task run on documents like yours, not a prepared demo.
  2. Does it know my firm? How does it take in your precedents and past work, how long does setup take, and does it hold context across the life of a case?
  3. Does it respect my permissions and ethical walls on every request, and is there an audit log?
  4. Where does my data go? Where is it processed and stored, who can access it, and is it used to train any model? Get the answers in writing. In the UAE, ask regime by regime: onshore, DIFC and ADGM each have their own data rules, and DIFC Regulation 10 adds risk assessment and notice duties where autonomous systems process personal data. In the EU, there is GDPR compliance checks.
  5. Can I check every answer? Each output should cite its source, whether that is a public authority or one of your own documents.
  6. Does it work natively in my jurisdictions and languages? Most public legal AI benchmarks are built on US law, so ask how the system performs on the law you actually practise.
  7. What will it really cost once usage charges, support and the hours your own people spend configuring and maintaining it are counted?
  8. What happens when the models change? Who updates the workflows, and how will you find out?

The Law Society of England and Wales’ guide, “Generative AI: the essentials”, covers contract terms and exit planning in more depth and is worth reading alongside this list.

How to run a pilot that tells you something

  • Start in one practice group, with a couple of case types you handle often.
  • Agree what good looks like before you begin, for example hours per task and rounds of revision.
  • Run a fair comparison. We suggest giving similar cases to a traditional team and an AI-supported team, then comparing timekeeping data.
  • Use your own precedents in the pilot, or you won’t learn whether firm knowledge makes a difference.
  • Redesign the workflow instead of automating the old one step for step.
  • Protect learning time. Ropes & Gray lets first-year associates put up to 20% of their billable hours towards learning AI tools.

Where this leaves firms

Most lawyers already use AI, so adoption is settled. What is still open is whether they keep treating it as a faster way to ask questions or start treating it as capacity, and take on work that until recently only firms with large teams behind every partner could handle.

For large firms, that capacity adds to the headcount they already have. For boutiques and solo practitioners, it may be the first real chance to compete for the same work while keeping the direct relationship that brought their clients to them in the first place.


Sources


Author: Tamara Essayyad

Leave a Reply

*

This site uses Akismet to reduce spam. Learn how your comment data is processed.