If You Are a Lawyer and You Are Not Using AI, Your Days Are Numbered

I am going to say it bluntly and qualify it afterwards, which is the proper order.

If you are a lawyer and, in 2026, you are still working as though artificial intelligence did not exist, the days of your current competitive position are numbered.

That does not mean you will close your firm next month, or that a machine will replace you tomorrow. It means something more concrete: you no longer determine on your own the pace at which the work can be done, its cost of production or the standard against which your clients will compare you.

And now for the qualification, because much of what is written on this subject is advertising in disguise:

Using artificial intelligence will not save you either. If you integrate it badly, it may even multiply your mistakes.

Courts have already sanctioned lawyers for submitting fabricated citations produced with the assistance of AI systems. Professional bodies are beginning to draw a clear boundary: the tool may be artificial; professional responsibility remains human.

But before we reach that point, we need to understand what has changed.

The battle in 2025: the samurai versus the Marine

Imagine an outstanding lawyer. Twenty-five years in practice. An extraordinary memory. An instinct for anticipating the other side’s line of attack. The ability to open a hundred-page contract and identify almost intuitively the clause that matters.

This lawyer does not use artificial intelligence. He reads, thinks, checks and drafts as lawyers have worked for decades.

We might picture him as a veteran samurai. He has spent half a lifetime mastering his weapon. He understands distance, terrain, his opponent and his own limitations. His advantage lies not only in the sword, but in everything he has learned by wielding it.

Opposite him, let us place a reasonably good lawyer who was already using AI in 2025 to summarise documents, prepare drafts, translate texts or accelerate research.

That second lawyer resembles a Marine with a modern rifle.

The comparison is deliberately brutal. A modern weapon multiplies the capacity of the person who uses it, but it does not automatically make that person a better strategist. It does not teach him where the adversary is or when to fire.

That was how AI worked for many professionals in 2025: as an extraordinary multiplier of intellectual productivity. The machine accelerated execution, but the human still decided what had to be done and when each step should be taken.

The samurai retained something the rifle could not supply on its own: twenty-five years of judgement about which question to ask, which document to examine first and when an apparently innocuous clause was dangerous.

That is why he could still win many battles.

The battle in 2026: the samurai versus an airmobile unit

In 2026, it is not only the models’ prose that has improved. The shift from conversing with an artificial intelligence to delegating work to it has become far more visible and accessible.

Agents existed before, but their capabilities, endurance, access to tools and presence in products used in real work have changed the scale of what a professional can assign to them.

An agent can receive an objective, divide it into tasks, use tools, consult documents and sources, evaluate results and continue without waiting for a new human instruction after every operation. Multiple agents can work in parallel, share their results and be coordinated by another system. With memory, tools and a suitable harness, they can maintain context, execute procedures, record what they have done and stop when they reach a limit or a decision reserved for a person.

Let us return to the battle.

The Marine of 2026 does not simply have a better rifle. He can call an entire unit over the radio. He can request reconnaissance, assign work to several specialists and ask an airmobile unit to reach one location while he continues dealing with another problem.

That is the qualitative difference.

In a suitable architecture, the unit may take the form of several agents reviewing contracts against common criteria, performing the first sweep of a due diligence exercise, reconstructing a chronology, checking citations against original sources and subjecting a draft to a separate critical review before it reaches a human reviewer.

These capabilities are not infallible, nor do they amount to finished professional work. A poorly directed unit may produce more errors—and do so faster. But the lawyer is no longer merely writing more quickly: the lawyer can coordinate execution capacity.

OpenAI describes this transition using data from its own product. In a random sample of 0.1% of individual Codex users observed in May 2026, it estimated that 70.2% had made at least one request equivalent to more than one hour of human work. In June, its active internal users at the 99th percentile regularly generated more than 60 hours of agent turns per day across multiple parallel runs. The second figure comes from an internal environment especially favourable to adoption and is not representative of the general user. Both are directional provider estimates based on an AI evaluator, not independent measurements of productivity.

The samurai may still be a better individual combatant than any member of the unit.

That is precisely the problem.

Being the best executor matters less when execution capacity is no longer the scarce resource.

The real lesson of Nagashino

Here, the history is considerably better than the cliché.

The samurai is often portrayed as the traditional warrior who rejected firearms as dishonourable and was defeated by modernity. That is not what happened.

The conventional account places the arrival of Portuguese firearms at Tanegashima in 1543, although historians dispute aspects of that narrative. What matters for our comparison is less controversial: the Japanese soon began reproducing them, and different warlords incorporated them in growing numbers. Adoption was rapid.

Thirty-two years later came the Battle of Nagashino. The popular version says that Oda Nobunaga invented a perfectly ordered rotation of three ranks of arquebusiers there to defeat Takeda cavalry. Current historiography questions that overly tidy image, the number of marksmen and various tactical details. The more cautious interpretation emphasises that the weapon formed part of a military organisation and did not act as a miraculous technology on its own.

Nor were the Takeda a medieval army that had never seen an arquebus. They had adopted firearms before Nagashino.

The interesting contrast, therefore, is not arquebus versus sword. It is technology integrated into an organisation versus technology insufficiently integrated into it.

The weapon could be purchased. The difficult part was deciding where to position it, how to protect the person using it while he reloaded, how to keep it supplied and how to coordinate it with the rest of the army.

Nagashino helps dismantle an idea we still repeat today: that adopting a technology means buying it.

Important technologies eventually spread. For a time, the advantage belongs to those who learn first how to reorganise people, knowledge, procedures and decisions around them.

That is precisely what is beginning to happen in law firms.

The pressure does not come only from competitors

Adoption surveys must be read with caution. Having used ChatGPT once is not the same as having legal processes integrated with AI.

The 8am 2026 Legal Industry Report, based on a survey conducted between September and October 2025 among more than 1,300 US legal professionals, reports that 69% used general AI tools at work, up from 31% the previous year. 42% used legal-specific tools. A sponsored summary published by the American Bar Association added that 54% worked at firms with neither training on the responsible use of AI nor plans to provide it.

This is a vendor survey limited to the United States. It does not, by itself, describe the global profession. It does show that basic adoption is no longer marginal in a significant part of the market.

In Europe, the Norwegian Bar Association and Karnov announced in March 2026 that they would conduct the comparative Fremtidens jurist study in Norway, Sweden, Denmark, Spain, France and Portugal. Its results were expected in the autumn; the existence of the research should not be confused with conclusions that have already been published.

But the most strategically important finding is not about other law firms. It is about their clients.

In a 2025 study by the Association of Corporate Counsel and Everlaw, conducted among 657 in-house legal professionals in 30 countries, 64% expected generative AI to reduce their reliance on external counsel and 61% expected to press their firms to change how they delivered and priced their services. Among US respondents, reported use had risen from 23% to 52% in one year.

Here too, we must remember who sponsored the study and the nature of its sample. But the economic direction deserves attention.

If an in-house legal department can complete in two hours a task it previously outsourced for two days, the pressure will reach pricing, organisation and hourly billing. The client will begin to ask an uncomfortable question:

Why am I paying for time spent if technology has reduced the time required?

Artificial intelligence does not transform production alone. It is beginning to transform the purchasing of legal services.

Capacity also multiplies error

The comfortable conclusion would be: buy helicopters. Obtain the best AI, automate, and put agents to work.

That conclusion is incomplete.

The same technology that expands a lawyer’s capacity can expand the speed and reach of the lawyer’s mistakes.

Researcher Damien Charlotin maintains a public database of court cases involving AI hallucinations. As at 14 September 2026, it contained 2,041 identified cases; 812 appeared in the category relating to lawyers and 1,174 in the category for self-represented litigants. The categories overlap, and the database exceptionally includes cases in which AI use was alleged but not confirmed, rather than sanctions alone. Nor does it measure every instance of misuse: it records decisions in which the court substantively addresses the appearance of hallucinated material.

It is not an incidence statistic. It is nevertheless a warning that is difficult to ignore.

In Spain, the Criminal Chamber of the High Court of Justice of the Canary Islands reported on 13 February 2026 that it had fined a lawyer €420 for including as many as 48 false case-law citations generated with a general-purpose tool in an appeal. The Chamber also referred the matter to the Bar Association for consideration of possible disciplinary responsibility.

The economic reasoning behind the decision is striking: in setting the amount of the fine, the court used the approximate annual price of a subscription to a legal AI tool as a reference and took into account the professional’s acknowledgement and remorse.

But the important lesson is not which brand of tool might have prevented the problem. No provider removes the duty to verify.

The problem is not using artificial intelligence. The problem is using it without the diligence required by professional practice.

What cannot be delegated

Here we reach the boundary that gives order to everything else. It is a professional-ethics boundary before it is a technological one.

The guidance approved by the Norwegian Bar Association in September 2025 states that a lawyer retains ultimate responsibility for advice and services provided when using AI. The content must be reviewed, verified and approved before professional use.

The CCBE’s 2025 guide for European lawyers starts from the same premise: using AI does not suspend the duties of competence, independence, confidentiality or professional responsibility. Its 2026 technical guide also explains the differences among tools, models and architectures that professionals must understand when evaluating them.

When a system writes something in your name, the client has not retained the model. The client has retained you.

That is why asking the same system to “check carefully” what it has just written is not enough. In high-risk work, the control must introduce independence: comparison with the original source, deterministic rules where possible, separate checks and human review before the output leaves the organisation.

The machine can produce.

Another layer must check.

And someone must authorise and answer for the result.

That is where genuinely professional architecture begins.

From executor to director of the work system

The right move for an outstanding lawyer is not to compete with the machine on speed. Nor does it require the lawyer to become an artificial-intelligence engineer.

It means moving up a level: from executor to director of the work system.

The capabilities required are remarkably similar to those a good partner has exercised throughout a career.

First, the objective must be defined: what is to be achieved, what lies outside scope, which sources may be used, what level of reliability the result requires and when the process must stop. An ambiguous objective given to a system capable of acting does not produce only a poor answer; it may produce a great deal of useless work.

Resources must then be allocated: which task can be performed by a small model, which requires a more capable one, what a junior lawyer should review, what needs a partner’s experience and what information may be processed externally.

Verification must then be designed: which assertions require a primary source, what can be checked automatically, which documents must be compared and which decisions require mandatory human approval.

Finally, someone must take responsibility for the result: authorise it, sign it and answer for it.

Our capacity to produce analysis is increasing rapidly. Our capacity to determine that an analysis deserves trust is not increasing at the same pace.

What will remain scarce

Some things do not become abundant simply because computing capacity increases.

The first is trust. Generating a memorandum may require a fraction of the work it did three years ago. Persuading a business owner that it is safe to buy a company on the strength of your analysis remains difficult. As production becomes cheaper, the bottleneck shifts towards reputation, judgement and the willingness to assume responsibility.

The second is knowing where not to use the machine. Some actions may be unsuitable for automation even if they are technically possible: irreversible decisions, especially sensitive communications, conclusions whose sources cannot be verified or judgements that depend on human information that is difficult to formalise.

Knowing how to use AI will be a basic competence. Knowing when not to use it will be a higher-order competence.

The third is training the next generation.

We are automating the school of lawyering

Among the tasks most often proposed for automation are precisely those traditionally performed by junior lawyers:

• researching case law;
• summarising files;
• reviewing contracts;
• preparing first drafts;
• building chronologies;
• comparing documents.

But those tasks were not merely production capacity. They were training.

They are how lawyers learn to distinguish an important judgment from an irrelevant one, recognise the clauses that tend to cause problems and notice when a document is missing.

Someone with twenty-five years in the profession can use AI to multiply judgement already acquired. Someone starting today presents a different problem:

How do we build twenty years of judgement if we automate precisely some of the tasks through which it was previously acquired?

I do not have a complete answer. I distrust anyone who claims to have one.

But one conclusion appears reasonable: training cannot consist of removing the junior lawyer from the process. It must allow the lawyer to inspect sources, formulate hypotheses, identify errors and justify decisions even when a machine has performed the first pass. If the junior merely reviews results that already look finished, there is a risk that the judgement needed to recognise when they are false will never develop.

Automating junior work also requires us to design a new learning system.

What to do on Monday

“Learn to use AI” is not a plan.

1. Choose one real, repetitive and thankless process. A recurring contract review, extraction of information from deeds, document classification or the initial analysis of a matter.
2. Measure it before changing it. Time, people, cost, errors and current controls.
3. Automate first what you can verify. Begin with outputs that can be checked quickly against a source.
4. Make verification mandatory. No legal citation should reach a filing without being opened in the original source.
5. Resolve confidentiality before scaling. Determine what information may be processed, where, by which provider and subject to which obligations.
6. Train the team and define the rules. Permissions, limits, documentation, accountable persons and human-approval points.
7. Measure again. If the process is not faster, less expensive or more reliable, perhaps the problem was not the absence of AI, but its design.

One point for Norwegian firms: in August 2026, the Government announced a new consultation on the future Norwegian AI Act and stated its ambition to submit the bill to the Storting in spring 2027. In the meantime, existing rules on privacy, responsibility, professional secrecy and intellectual property continue to apply. The temporary absence of a specific Norwegian AI Act is no reason to postpone internal controls.

Nagashino, once more

Let us return to 1575.

The lesson is not that a new technology enabled modern warriors to annihilate outdated ones. The Japanese had been using firearms for decades, and the Takeda knew the arquebus.

What mattered was integration.

The weapon could be copied. The system was much harder to copy.

Before long, saying that a law firm uses artificial intelligence will be about as informative as saying that it uses email. Access to models will tend to become widespread. The difference will lie in the architecture:

• what you delegate and what you do not;
• what information each system receives;
• how you verify;
• how you protect data;
• how you introduce independence between production and control;
• how you train junior lawyers;
• how you turn professional experience into usable procedures;
• and who is prepared to assume responsibility for the result when it enters the real world.

That is why the important question is no longer whether the human or the machine is the better lawyer.

It is this:

When legal work begins to be produced through systems made up of lawyers, knowledge, procedures and machines, what place do you occupy within that system?

Technology will eventually spread, as the arquebus did.

The advantage will not belong for long to those who merely possess it, but to those who learn first how to reorganise themselves around it.

Scope note

The adoption figures come from surveys with different methodologies, samples and interests and are not directly comparable. The agentic-productivity metrics cited are estimates published by the provider. The article uses Nagashino as an organisational analogy and does not seek to settle historiographical debates about the battle.