How Lawyers Use AI Now (and What Surveys Do Not Show)

Many people talk about AI and the law. But the daily use of AI by lawyers is simple. Lawyers usually use AI for these tasks:

• To summarise documents.
• To find information.
• To write first drafts.
• To make letters and emails better.
• To prepare work that they know how to check.

These tasks are not dramatic. But they help to explain how lawyers use AI.

A survey can tell you that 62% or 94% of lawyers “use AI”. But the survey does not tell you which tasks they give to AI. It does not tell you how they check the result, or who makes the decision.

This lesson uses surveys from 2025 and 2026, and data from real use. The important question is not “Which task is first?” The important question is “Why are the same tasks first in each survey?”

The same tasks are first in each survey

The six columns in the chart come from different countries and different sponsors. Each survey used its own list of tasks. But the top of each list is almost the same: research, summaries, document review, drafts and letters. The agreement between the surveys is more important than each number.

Look at the last row. Most surveys do not ask about strategy. One survey in Italy asks about it. In that survey, about 4% of the lawyers who use AI say that their main use is strategy or case prediction Censis · Rapporto sull'Avvocatura, Apr 2025 (opens in a new tab). In Texas, 3% of respondents use AI for client intake.

AI can write a negotiation strategy as fast as a summary. Thus, usefulness does not explain this result. Lawyers use a different filter to select the tasks that they give to AI. This lesson is about that filter.

Lawyers give AI the tasks that they can check

One idea agrees with the data. Lawyers give a task to AI when they can check the result quickly with a source. For example:

• Is the summary correct for the document?
• Did the AI find the correct clause?
• Are the dates in the table correct?
• Does the case exist, and does it say that?

For each of these checks, you compare the result with a source: the document, the clause or the database. The comparison gives the answer.

Other questions are different:

• Which argument is the strongest?
• Is this commercial risk acceptable?
• How much do we offer to settle?

For these questions, there is no source for comparison. To check the answer, you must make the same decision yourself.

Thus, the difference is not between easy tasks and difficult tasks. A summary of a 300-page file is a large task. But you can find the page for each sentence of the summary. A decision about an indemnity can be fast. But no document tells you if the AI decided correctly.

A survey of French lawyers (avocats) lets us test this idea. The French national bar council (CNB) asked 4,457 avocats: “Do you trust generative AI for this task?” CNB · Enquête IA, Sep 2025 (opens in a new tab)

The results agree with the idea. 66% trust AI to find the most relevant source. 32% trust AI to write submissions. Only 9% trust AI to decide a case.

Select the second view. Frequent users trust AI more for each task: from 73% to 14%. But they put the tasks in the same order.

The surveys agree with this idea, but they do not prove it. Other causes are also possible: professional rules, liability, habit, and the fact that clients pay lawyers for their decisions. It is possible that all these causes have an effect together.

In the same survey, 92% of the avocats who use general AI tools try to check the answers. For frequent users, the value is 93%. Lawyers who use AI frequently also check it frequently.

Trust is not a general opinion about AI. Trust is for one task, with one type of check, and with one cost of error. A lawyer can trust AI to change the format of a table. The same lawyer can refuse to trust AI with a citation. These two positions agree.

Five types of task that you can give to AI

Survey labels include very different requests. For example, “drafting” can mean these tasks:

• To write an email again with better words.
• To write a first draft from detailed instructions.
• To change a precedent for a new matter.
• To write an argument.
• To recommend a legal position.

The first task is almost an edit. The last task is almost legal advice. A better method is to sort each request by the action that you ask AI to do.

From left to right, the type of check changes. On the left, you compare the result with a text that exists: your text, the contract or the database. On the right, you must use your own decision to check the result.

Most use of AI is on the left and in the middle. Look again at the first chart. The rows with the most use are “Change”, “Find” and “Compare”. “Write” is also frequent, but lawyers usually edit the drafts. “Decide” is almost empty.

Lawyers give AI the steps that they can examine. They keep the steps where they make the decision.

The five types are not separate. One request can include more than one type. But these types show more than a label such as “document review”.

“Contract review” is not one task

Most surveys use the label “contract review”. Two lawyers can select this label and do very different work.

• Lawyer A puts a contract into a chatbot and asks: “Is there anything unusual here?”
• Lawyer B uses the team's written playbook. The AI compares each clause with the playbook and sorts the differences by risk. Then it takes alternative text from an approved list. Lawyer B approves each change.

Lawyer A gives two steps to AI. For one of these steps, there is no standard for comparison, because nobody defined “unusual”. Then Lawyer A makes a decision from a list that nobody specified.

Lawyer B gives eight steps to AI. For each step, there is a source that the team wrote. Lawyer B keeps the last decision. Lawyer B gives more work to AI, and has more control.

Legal research is similar. “AI for legal research” can mean a question to a chatbot. It can also mean a search for authorities in a database. In the first case, AI can do all the steps. This includes the application of the law to the facts of the client. In the second case, AI finds and summarises the authorities, and the lawyer does the legal analysis.

Thomson Reuters found that 80% of legal professionals who use generative AI use it for research Thomson Reuters · AI in Professional Services, 2026 (opens in a new tab). This means that 80% use AI for one or more steps of research. The number does not tell you which steps.

“I use AI” can mean different things

The same problem occurs at the level of the user. The statement “I use AI” does not tell you how the person uses it.

The surveys show these differences. In France, 62% of avocats use generative AI. Of these, 30% use it only “sometimes”, and 32% use it “often” or “regularly”.

In the Thomson Reuters survey, 81% of the law-firm users use generative AI each week or more. But only 18% of law-firm respondents say that it is a central part of their work. Frequent use and structured use are not the same.

Data from real use shows how people use AI in one conversation.

Anthropic, the company that makes Claude, publishes anonymous data about the use of Claude. In conversations about legal tasks, only 35% give all of the task to AI. This is the lowest value of all the job groups in the chart Anthropic · Economic Index data, Jun 2026 (opens in a new tab). For computer tasks, the value is 64%.

In the other legal conversations, the person works on a draft with the AI, or asks the AI to explain a subject. The data sorts the conversations by task, not by person. Thus, some of these people are not lawyers.

The depth of use is as important as the number of users. A yes-or-no question cannot show the depth of use.

The most frequent use is not the most important use

Surveys count the number of people who use AI for a task. Some surveys also ask how frequently. Few surveys measure these items:

• The time that AI saves.
• The complexity of the task.
• The cost of an error.
• The quantity of work that the lawyer gives to AI.

A lawyer who makes 20 emails better each day is in each survey about correspondence. A lawyer who uses AI one time to review 400 documents is possibly not in the results. But the second use can save days of work. It also has more risk.

A list of tasks in order of “percentage of users” is a list of common tasks. “The most popular use” means “the most common use”. It does not mean “the most valuable use”. Lesson 1.5 shows how much time AI saves.

Why lawyers use general chatbots and legal tools

General tools are frequent in the data. In France, 64% of avocats in firms that use generative AI use ChatGPT. Only 30% use Doctrine, the most used legal tool. In Texas, 63% of AI users use ChatGPT State Bar of Texas · AI survey, 2026 (opens in a new tab).

It is easy to compare these tools as competitors. A better question is: “Which tasks does each type of tool do?” One US survey asked this question.

Some tasks have more use with general chatbots: new ideas, letters and marketing. These tasks do not need legal sources. The lawyer supplies the information, and the lawyer can examine the result.

Legal research is only on the legal side. Document drafts have more use with legal tools. Summaries have the same use with the two types of tool, because you supply the document.

Specialised legal tools are necessary when the task needs one of these items:

• Authoritative legal sources.
• A case-law database.
• The precedents of the firm.
• Confidentiality controls.
• Fixed steps.

Firms that bought legal tools gave this main reason: they get AI in legal software that they already trust (52%).

The type of tool changes the confidence of lawyers. It does not change their checks. French avocats have more confidence in the answers of legal tools (54%) than in the answers of general tools (34%). But they check the two types almost equally (90% and 92%).

Lawyers use the two types of tool for different parts of their work.

Lawyers start to use AI before their firms do

Personal use is higher than the use by the firm:

• In a US survey by 8am, a legal software company, 69% of legal professionals use general AI tools for work. But only 46% of firms have these tools.
• In the Thomson Reuters survey, 55% of law-firm respondents use public tools such as ChatGPT. But only 41% say that their firm uses generative AI.
• More than one third of law-firm professionals use AI tools that their firm did not approve Thomson Reuters · Future of Professionals, Jul 2026 (opens in a new tab).

The cause is simple. A lawyer can start to use a chatbot in a few minutes. Before a firm can use AI, it must make decisions about these subjects:

• Confidentiality and security.
• Access rights.
• Procurement.
• Data retention.
• Professional obligations and supervision.

Thus, the questions “Do lawyers use AI?” and “Does the firm use AI?” measure different things. The two answers can be true at the same time. This explains many differences between survey results. Lesson 3.2 tells you what happens to data in personal accounts.

Why survey results are different

Recent surveys of lawyers give adoption values from 37% to 94%. Most of these values are correct. But the surveys ask different questions:

• “Did you ever use generative AI?”
• “Do you use it one or more times each week?”
• “Does your organisation use AI or test it?”
• “Do you use an approved legal AI product?”

Each question measures a different thing. The chart below shows each main result in the row of its question.

Select one survey and follow its results from row to row. The result decreases when the question is more specific. LexisNexis goes from 94% (use AI) to 34% (use AI each day). The Irish survey goes from 55% to 5%.

The question alone can change a result by 60 points. In each row, the differences that remain have other causes: the respondents, the country and the definition of AI.

Before you use a survey result, check these five items:

• Who answered. Surveys by software companies often go to their own customers. In the 2025 Clio survey, 1,202 of the 1,702 respondents were Clio customers. Surveys by bars and regulators go to more types of lawyer. Their results are usually lower.
• What counts as use. Ever, sometimes, each week or each day.
• Whose use. A person or a firm. The 94% from ILTA counts firms that “use or explore” AI. Each firm gives one answer.
• Which AI. “Any AI tool” can include translation tools and spelling tools. “Generative AI” is more specific.
• Where and when. The country, the size of the firm and the date of the survey change the result.

Differences between groups: read them carefully

Some differences between groups are in almost all surveys. But their causes are not clear.

Firm size. Large firms have money and legal platforms. In the 2026 ILTA survey, 84% of firms with more than 700 lawyers use or explore Harvey. For firms with fewer than 50 lawyers, the value is 22% LawSites · ILTA survey, Sep 2026 (opens in a new tab). But large firms also have slow procurement. A sole practitioner does not need approval, but often cannot get legal tools.

In-house teams. Many in-house lawyers get AI through company licences. In the Thomson Reuters survey, 62% of corporate legal respondents use company tools such as Microsoft Copilot. For law-firm respondents, the value is 38%. It is possible that this difference shows IT budgets more than the choices of lawyers.

Age and experience. In Italy, younger lawyers use AI much more. But in Texas, the highest use is by lawyers with 7 to 10 years of practice (82%), not by the newest lawyers (67%).

Junior lawyers and partners do different work. The tasks that AI does well are first drafts, research notes and summaries. These tasks are not equal across the two roles. The statement “Younger lawyers like AI” is too simple.

The common pattern

The data does not show that lawyers give “legal work” to AI in general. Lawyers give only some parts of their work to AI. They give a task to AI when these conditions are true:

• The task occurs frequently.
• The AI has the necessary information.
• You can check the result with a source.
• You can find and correct an error before it causes a problem.
• The lawyer keeps the important decision.

This explains why research, summaries and drafts are at the top of the lists. For each of these tasks, there is a source for the check, or a lawyer edits the result before it leaves the office.

These tasks are popular because lawyers can check them, not only because they are useful.

What occurs when these uses become more structured is a different question. Other chapters discuss it.

Five questions for each use of AI

You can use the same method for each use of AI: your own use, or the use that a supplier shows you. Ask these five questions:

• What do you give to AI? Change, find, compare, write or decide.
• How easy is it to check the result? With a source, with a standard, or only with your own decision.
• What is the cost of an error? An internal draft, or the position of the client.
• How much structure is there? A free chat, or a defined method with sources and approval.
• When do you make your decision? Before, after, during, or almost never.

What we can say about the present

The important difference is not between lawyers who use AI and lawyers who do not use AI. These questions are more important:

• What do lawyers give to AI?
• How much do they give?
• How do they check the result?
• What happens when the result is wrong?
• Where does the lawyer make the decision?

One adoption value cannot answer these questions. Together, the answers show how lawyers use AI now. Many lawyers use AI. They use it mostly for work that they can check. They keep control of the decision.

Many lawyers say that AI saves them time. Lesson 1.5 examines that claim: what the studies measure, and what they do not measure.