If you had asked me a year ago which artificial intelligence platform I would recommend to a lawyer working in a Microsoft 365 environment, Microsoft Copilot would have been an easy answer.

The reasoning was compelling. Copilot was already inside the environment where attorneys worked. It could take advantage of Microsoft Graph and organizational data. Access was governed through Microsoft Entra ID. Existing Microsoft 365 permissions continued to matter. Purview provided an established framework for information protection, retention, eDiscovery, and compliance. Attorneys did not have to move their work into an entirely separate ecosystem simply to experiment with AI.

That was a reasonable answer. It is also a good example of the problem with training attorneys on AI products rather than teaching them how to work with AI.

The landscape has changed dramatically. Microsoft now describes Copilot Cowork as a multi-model environment for long-running, multi-step work. It uses Anthropic models and can also offer OpenAI models within Microsoft’s enterprise environment. Anthropic’s Claude Cowork similarly emphasizes handing AI actual assignments involving files and connected workplace systems. ChatGPT Work can gather context from files and apps, take action across workflows, produce reviewable deliverables, and perform recurring or event-driven tasks.¹²³

Legal-specific platforms are moving in the same direction. CoCounsel Legal combines agentic workflows with Westlaw and Practical Law. Lexis+ with Protégé combines legal research, drafting, organization knowledge, skills, agents, and Shepard’s validation. Harvey positions purpose-built agents as systems that execute complex legal work end to end.⁴⁵⁶

What should we teach lawyers today that will not be obsolete six months from now?

The answer is not another class on “50 prompts every lawyer should know.” It is time to stop training lawyers primarily on AI products and start training them on AI work.

The AI Training Shelf-Life Problem

The pace of AI development creates a problem that traditional technology training rarely encountered. Microsoft Word changes, but an attorney trained to use styles, track changes, and document comparison does not wake up six months later to discover that those concepts are irrelevant.

AI is different. Models improve rapidly. Product names change. Features move between products. Companies integrate competitors’ models. Interfaces disappear. Capabilities that once required specialized software become standard features in general-purpose platforms.

Choosing “the best AI” is therefore becoming less like choosing one legal research service over another and more like selecting the appropriate combination of intelligence, information, tools, security, and legal authority for a particular assignment.

The important skill is no longer knowing where a particular button is located. It is understanding what you are asking AI to do, what information it needs, what systems it should be permitted to access, what sources should control its answer, how much autonomy it should have, and how its work should be reviewed.

Attorneys Need to Understand the Five Levels of AI Work

One useful way to teach AI is to stop treating every interaction as “prompting.” There are increasingly five levels of work an attorney can give to an AI system.

At the first level, the attorney asks: “Explain this provision,” “Summarize this case,” or “Give me ten topics to consider before this deposition.”

At the second level, the attorney asks AI to analyze supplied materials: “Compare these agreements,” “Identify inconsistencies among these witness statements,” or “Analyze these documents for evidence supporting causation.”

At the third level, the attorney asks AI to create work product: a memorandum, chronology, deposition outline, contract provision, client communication, presentation, spreadsheet, or first draft of a legal document.

The fourth level is fundamentally different. The attorney begins to delegate. Instead of requesting one output, the lawyer gives AI an assignment with several steps: research the issue, review the relevant documents, identify the strongest authorities, compare the law with the facts, develop an argument structure, and produce a first draft.

Finally, the attorney can automate work. The system performs scheduled, monitored, or event-driven tasks: reviewing regulatory developments every week, preparing a morning matter update, triaging new intake, or drafting a recurring status report.

Many lawyers are still being trained almost exclusively at Level 1. The largest productivity gains are increasingly available at Levels 3 through 5, but those levels also require stronger supervision and governance.

Prompt Engineering Is Useful. Context Engineering Is More Important.

Prompt engineering received enormous attention during the first wave of generative AI. Attorneys learned formulas involving roles, instructions, tone, formatting, and step-by-step directions. Those techniques can improve results, but the more durable concept is context engineering.

An AI system cannot reliably perform sophisticated legal work simply because an attorney crafted an eloquent prompt. It needs the right information.

The difference is not clever wording. The difference is context.

Attorneys should ask a fundamental question before assigning anything substantial to AI: What would I give a capable associate if I wanted that associate to perform this assignment correctly?

That mindset transfers remarkably well. The attorney identifies the objective, supplies the relevant record, defines controlling sources, communicates client and firm requirements, states what the system must not do, and specifies the desired work product.

Lawyers Must Understand Grounding

There may be no AI concept more important for lawyers than grounding. Every attorney should develop the habit of asking: What information is this answer based on?

There is an enormous difference between asking a general-purpose model what the law says and asking a legally grounded system to analyze an issue using identified primary authority. There is another enormous difference between asking AI what happened in a matter and asking a system with authorized access to the matter’s actual documents.

A fluent answer is not necessarily an authoritative answer.

General-purpose AI can be excellent for brainstorming, explanation, organization, synthesis, drafting, and analysis. Deep-research systems can search multiple sources, reconcile competing information, and produce sourced reports. But legal research adds requirements: Is the authority controlling? Is it current? Has it been negatively treated? Did the system find contrary authority? Can the attorney inspect the primary source?

That is why legal-specific systems remain important as general AI becomes more capable. The lesson is not “always use Lexis,” “always use Westlaw,” or “never use ChatGPT for research.” The durable lesson is: Know when authoritative grounding is required and verify that the system actually used it.

Verification May Be More Important Than Prompting

Generative AI can produce extraordinarily persuasive wrong answers. Verification is therefore a professional skill, not an optional AI feature.

An attorney should apply different levels of review depending on the consequence of the work. Brainstorming potential deposition topics requires a different standard than filing a brief. A document summary may require comparison with the source. A chronology should be checked against the evidence. Contract analysis requires substantive attorney review. Legal research requires verification of authorities and treatment.

The governing ethical principles are more durable than today’s tools. ABA Formal Opinion 512 addresses competence, confidentiality, communication, supervision, candor, and reasonable fees in lawyers’ use of generative AI. It explains that lawyers need a reasonable understanding of the capabilities and limitations of the technology they use.⁷

Do not teach attorneys that AI is categorically trustworthy or untrustworthy. Teach them how to determine the appropriate level of reliance for a particular task—and what verification must accompany it.

Research, Deep Research, and Agents Are Not the Same Thing

A chat system may answer, “What are the elements of fraudulent inducement?” A deep-research system might examine multiple sources, identify competing interpretations, look for recent developments, and produce a sourced report.

An agent might receive a broader assignment: research the relevant law; review the complaint and opposing motion; analyze the record for evidence relevant to each element; identify factual weaknesses; prepare an argument matrix; develop a response structure; and create a draft memorandum.

The attorney has moved from asking a question to assigning work. Lawyers do not need to become AI engineers, but they do need to become better delegators. That means defining the objective, supplying relevant materials, establishing constraints, reviewing the proposed approach, setting checkpoints, and evaluating the final work product.

AI Is Moving From Generating Words to Performing Work

Lawyers do not get paid simply to generate words. They review documents, compare agreements, extract information, build chronologies, analyze evidence, prepare tables, revise Word files, review spreadsheets, create presentations, organize matter information, monitor developments, and communicate with clients.

Modern AI systems increasingly operate on those artifacts directly. Training exercises should therefore focus on deliverables rather than isolated prompts.

The same principle applies to contracts. “Summarize this agreement” is a chat request. “Compare this agreement against our approved playbook, identify deviations, rank them by risk, explain the business significance, and prepare proposed revisions” is a work assignment.

Attorneys Need to Understand Agents Even If They Never Build One

At a basic level, an agent combines intelligence with instructions, information, tools, and permission to take actions. A chatbot can tell you how to perform a task. An agent may actually perform portions of it.

That creates possibilities—and risks—that simple chat does not. What systems can the agent access? Which documents can it read? Can it create or modify files? Can it send communications or initiate another process? What identity does it use? Where are its actions recorded? What happens when it makes a mistake? Who approves consequential actions?

An attorney does not need to understand the complete technical architecture. An attorney does need to understand why giving an agent access to a document repository, mailbox, or filing workflow is fundamentally different from pasting a paragraph into a chatbot.

Scheduled AI Will Quietly Become One of the Biggest Changes

One of the most important developments may appear mundane at first: AI that works when the lawyer is not actively talking to it.

Compare “Tell me whether the Department of Labor issued anything new concerning this regulation” with “Monitor the Department of Labor and notify me when guidance affecting this issue changes.” Or compare “Summarize these messages” with “Each weekday morning, review matter-related messages from the previous 24 hours and prepare an action summary.”

The AI is no longer waiting for an attorney to initiate every interaction. Firms need lawyers who understand when recurring AI is appropriate, what sources it should monitor, what conditions should trigger it, and where human approval must remain.

Attorneys Also Need Basic AI Architecture Literacy

Lawyers do not need to become developers, but sophisticated users should understand four concepts that determine what an AI system can and cannot do:

Permissions. AI should not gain access to information simply because the system exists. Access should continue to reflect appropriate user, matter, ethical-wall, and organizational permissions.

Connectivity. A standalone model knows what it learned during training and what the user supplies. A connected system may access SharePoint, Outlook, a document management system, a legal research platform, CRM data, or other authorized sources.

Tools. Giving AI a tool can let it perform an action rather than merely discuss it.

Skills and workflows. Reusable instructions can encode how the organization wants a particular task performed, reviewed, and delivered.

AI intelligence + appropriate context + authorized tools + controlled permissions = useful AI work.

So Which Products Should Attorneys Learn?

They should learn several—but as examples of capabilities rather than as the curriculum itself.

There is no reason to assume today’s winner in a category will remain the winner. The better question is not, “Which model should I learn?” It is, “Which environment gives this assignment the right intelligence, information, tools, security, and legal authority?”

What I Would Teach Attorneys Today

A durable curriculum should be built around actual legal assignments rather than software demonstrations. The following eight modules create a progression from basic user to effective AI supervisor.

Expectations by Role

AI competence should not mean the same thing for every person in a firm. Each role needs a defined standard tied to the work it directs, performs, supports, or governs.

A firm should not declare its workforce “AI trained” because everyone attended the same introductory webinar. Readiness is demonstrated when each role can perform its responsibilities under a shared governance framework.

What I Would Stop Teaching

I would reduce the time devoted to libraries of “perfect prompts,” current model benchmark comparisons, and tours of one vendor’s interface. I would be careful about presenting any single platform as the firm’s permanent AI answer.

Those topics may be useful today, but they are perishable knowledge. Teach them where necessary, then teach the durable concept underneath them.

A button may move. A model may disappear. A product may be renamed. The need to provide good context will remain. The need to distinguish authoritative information from plausible language will remain. The need to protect confidential information, supervise delegated work, verify legal authority, and take responsibility for the final work product will remain.

The Next Phase of Legal AI

The next stage of adoption is unlikely to be defined by attorneys spending more time inside chat windows. AI will increasingly operate inside Word, Outlook, SharePoint, document management systems, research platforms, matter workspaces, browsers, and specialized legal applications.

Model selection will become less visible as systems route tasks among models based on the assignment, reasoning needed, available tools, speed, and cost. The distinction between general AI and legal AI will also become less rigid: legal platforms are incorporating frontier models and agent technology while general platforms gain access to specialized sources and enterprise systems.

As that happens, one of a firm’s most valuable AI assets may not be its preferred model. It may be its own knowledge: precedents, templates, briefs, agreements, matter histories, negotiation positions, research, playbooks, and institutional experience.

AI readiness is therefore not only a software-purchasing decision. It is also a knowledge-management, information-governance, permissions, process-design, and training challenge.

Stop Training Lawyers on ChatGPT

There is nothing wrong with teaching attorneys how to use ChatGPT, Claude, Copilot, Harvey, CoCounsel, or Lexis+ with Protégé. They should understand several of them.

But those products should be the laboratory in which attorneys learn durable AI skills, not the skills themselves.

Teach attorneys how to provide context. Teach them how to choose appropriate sources. Teach them to distinguish asking from analyzing, creating, delegating, and automating. Teach them how agents use tools and permissions. Teach them what work can safely be delegated, how to verify the result, and where human judgment must remain.

Most importantly, teach attorneys to ask the question:

What work am I doing today that AI could help me analyze, create, delegate, monitor, or automate—without compromising the judgment, confidentiality, and accountability my clients expect from me?

That is the AI capability worth developing. Products will continue to change. The ability to work effectively with AI will not.