An AI agent can help complete parts of an assignment that a junior lawyer might otherwise perform: find documents, organize facts, compare provisions, research authorities, and assemble a draft for review. The overlap is in the tasks, not in professional status. A system does not become a lawyer because its interface lets someone “delegate” work. A lawyer must decide whether the output is correct, complete, current, and appropriate for the client and jurisdiction.
Editorial update, September 27, 2026. This article was first published January 31, 2026, when “tireless junior lawyer” was a provocative way to describe emerging legal agents. The phrase remains a useful question, but it is not an accurate job title or a measure of competence. This update retains the original publication date, revisits the prediction, and separates vendors' capabilities from results a firm has actually tested.
The more useful question is specific: Which parts of this matter can an agent perform within defined permissions, what evidence will it show, and where must a person make the decision? Our Agents section follows that question through the technology, workflow, oversight, and evaluation.
What changed after the January 2026 prediction?
By January, multi-step legal products were already real. In April 2025, A&O Shearman and Harvey announced agents for antitrust filing analysis, cybersecurity, fund formation, and loan review. The announcement described systems that work over matter documents and curated data, combine intermediate steps, and provide transparency and oversight. That is evidence of a product initiative and intended uses; it is not an independent finding that every output is accurate, that every deployment saves time, or that the systems can practice law.
Thomson Reuters now describes CoCounsel Legal as planning, researching, analyzing, and drafting from a plain-language request. Its product documentation says a user can stop, redirect, or refine a task. This is a current example of an assistant that can run a multi-step process—one reason “assistant” and “agent” cannot be divided into “reactive” and “proactive” categories. Product descriptions explain available features; they are not a substitute for a firm's own matter-based testing.
General-purpose agents expanded as well. OpenClaw is an example of a tool-using personal agent, but a local gateway does not mean all data stays on the computer: a selected hosted model or connected service may process it. OpenClaw's session documentation explicitly says the configured model provider still processes sent messages; its security guidance discusses tool permissions, message access, and the trust boundary. A legal team should evaluate those details before considering client information or connected accounts. None of these examples establishes that a product is a “digital associate” with a human associate's duties, training, or judgment.
The change since January is clearer product capability and a wider set of possible integrations, not proof that a system has acquired professional judgment. A vendor announcement can show what a tool is designed to do; a product manual can show which controls a user can operate. Neither tells a particular firm how often the system misses an exception in its documents, whether its sources stay current for a jurisdiction, or how much lawyer review time is required. Those questions need local evidence.
What makes an agent different from a chat response?
A chat response produces an answer to a prompt. An agentic system may choose among permitted tools, inspect the results, and take further steps toward a goal. A fixed workflow can also chain several model calls, but its path is specified in advance. Anthropic's agent design guide distinguishes model-directed tool use from predefined workflows and recommends the simplest design that serves the task. An agent may begin with a lawyer's request and pause for approval; it need not wake up on its own or work unattended.
Consider a fictional contract review. A lawyer could authorize a system to read a defined set of agreements, extract specified change-of-control and assignment provisions, return the source page for each extraction, and draft a comparison. The agent might choose which document to open next. It should stop when a source is missing or ambiguous and present a draft—not silently email the client, change a contract, or certify that nothing was missed. The lawyer then checks the original documents, the exceptions, the governing law, and the advice. The legal workflows guide traces this kind of assignment from intake to review.
The same assignment illustrates what “junior lawyer work” hides. Locating a clause and producing a comparison are bounded tasks. Determining whether the clause creates a deal risk depends on client objectives, the transaction structure, the rest of the agreement, and legal interpretation. Deciding whether to raise the risk with a counterparty or client is a professional decision with consequences. A system may assist with the first part and prepare material for the second; the analogy does not transfer the lawyer's responsibility to the system.
A useful handoff would identify the documents and versions read, quote each relevant provision, mark uncertain or missing items, and say which steps were not performed. If the search covered only two of three amendments, “no other change-of-control clause found” is too broad. The narrower, defensible statement is that none was found in the two documents actually examined, subject to review of the missing amendment. The reviewer can then decide whether the record is sufficient to proceed.
Adding more autonomy also adds places for mistakes. A tool may return the wrong file, a model may miss a qualification, or an early extraction error may flow into the final memo. A chain of steps is not proof of a more accurate answer. It must be tested against the actual task.
Does the agent learn from every assignment?
The January version of this article suggested that agents remember interactions and “learn from feedback” like junior colleagues. That confuses three different mechanisms. Conversation context is the instructions, messages, retrieved documents, and tool results supplied for a particular model response. Stored memory is information an application saves and may retrieve for a later run, such as a preference or matter note. Training or fine-tuning is a separate process that changes model parameters. Reusing a note or keeping a conversation available does not itself retrain the model.
Anthropic's discussion of context engineering explains how agents select information for a finite context window. A saved preference can help a future response if the application retrieves it, but a saved factual error can also be repeated. An agent may lose relevant context, use stale information, or fail to retrieve a needed file. A firm should verify what is saved, who can see it, how long it is retained, and whether a provider uses submitted data for training under the applicable product terms. Do not infer those answers from a feature called “memory.”
Suppose a lawyer corrects the agent: an exhibit was superseded on June 1. The correction may shape the next answer in the same conversation because it is in context. A separate saved matter note may cause a later run to mention the superseding exhibit if that note is retrieved. A model-training process would be different: it would use selected data to update model behavior beyond this one matter. Ordinary feedback in a chat is not such a process. Even with a saved note, the agent should open the current exhibit rather than rely on a remembered date. OpenAI's conversation-state documentation explains how prior messages can be carried into later model calls while each request remains subject to context limits.
Which tasks are plausible candidates for delegation?
Document comparison can be a candidate when the input set and the requested fields are explicit. A system can extract a clause, show its page, and highlight a difference between versions. The reviewer must still check that all versions and exhibits were included and that the comparison did not miss an exception in a definition or schedule.
Research preparation can mean finding possible authorities, recording the search path, and drafting a list of questions for a lawyer. The agent can help narrow a search; the lawyer must establish which authorities are controlling, current, and applicable to the facts. A case citation is evidence of a source's existence only after it is opened and checked. It is not proof that the case supports the sentence beside it.
Draft assembly can mean organizing approved facts and source material into a first-pass memo or chronology. The system should label assumptions and identify unsupported passages. The responsible lawyer decides what the client needs to know, which risks to emphasize, and whether the final language is accurate and appropriate. Client counseling, a filing, or a commitment to a counterparty requires human professional judgment and authorization.
The work that still requires a lawyer
Research needs checking against current primary authority and the right jurisdiction. A citation can look plausible while leading to an irrelevant, overruled, or nonexistent source. Document review needs sampling or complete verification appropriate to the risk, especially for negative findings: “no assignment clause found” is harder to establish than a quoted clause. Drafting needs assessment of client objectives, factual assumptions, strategy, tone, and consequences. Deadlines must come from the authoritative docket, order, or rule, with a responsible person and a tested calendar process; an agent's reminder is not a guarantee.
The American Bar Association's Formal Opinion 512 discusses competence, confidentiality, communication, candor, supervision, and fees in lawyers' use of generative AI. It applies the ABA Model Rules; the controlling duties depend on the relevant jurisdiction, court orders, and matter. The practical point is that a polished draft does not discharge the lawyer's obligation to understand and review the work.
Supervision should have an actual handoff. A named lawyer or authorized team member reviews the source record and exceptions, resolves open questions, approves any external use, and retains a way to reconstruct what material the system relied upon. A task that arrives in an inbox marked “complete” without that record is harder to supervise than one that clearly says what was checked and what was not. The level of review should reflect the task's risk, not the speed or confidence of the generated prose.
Tool access brings additional risks. A retrieved webpage, email, or document may contain text that tries to give the agent new instructions. NIST's agent-hijacking analysis describes this indirect prompt-injection problem. The same agent can also act in the wrong account, disclose a document through a connector, or exceed its intended scope if permissions are broad. A written instruction to “be careful” is weaker than enforced read/write limits, source boundaries, logs, and approval gates. See the permissions and oversight guide for a practical framework.
How a firm should test the “junior lawyer” claim
Start with one bounded assignment and fictional or appropriately approved data. Write down a human baseline: what the work requires, who checks it, how long it takes, and what errors matter. Give the system only the files and tools needed for that assignment. Run a representative set of easy, typical, and difficult matters, including incomplete or contradictory materials. Record each tool call and human intervention.
For the fictional assignment above, the test set could include a clean agreement, one with a later amendment, one with an incorporated exhibit that is absent from the packet, and one whose clauses use unusual wording. Define the expected findings before viewing the agent's result. Mark a missing amendment as a serious failure if the system confidently says the packet is complete. Mark a correct abstention or escalation as a useful result when the record cannot support a conclusion. This prevents a polished but unsupported memo from scoring better than a cautious, reviewable one.
Score more than the final prose. Check whether every legal proposition is supported by the cited source, whether material issues were missed, whether the agent acted without authorization, and how much lawyer time was needed to verify and repair the result. Count model, tool, setup, and review costs. Repeat trials because one successful demonstration may not represent the next matter. Anthropic's guide to agent evaluations explains why multi-step runs need task-specific graders and traces. Our agent evaluation guide supplies a matter-based scorecard.
This process may show that an agent helps with a narrow task. It may also show that a fixed workflow, ordinary search, or a human working directly is faster or safer. Neither conclusion should be assumed from a vendor demo or a comparison with an exhausted associate. A strong result is an auditable improvement in the firm's actual work, with lawyer review and client duties accounted for.
The decision should be revisited when documents, legal sources, models, integrations, or team procedures change. A pilot on old, clean examples does not validate a new matter type or a connected action tool. If a system fails a material case or requires extensive repair, narrow the assignment or return it to a simpler process. The choosing and piloting guide develops the adoption decision; this article's point is that the “junior lawyer” label supplies no substitute for that evidence.
What “tireless” should mean now
Software can run outside office hours. That does not mean it has complete recall, detects every issue, never misses a deadline, or produces court-ready work without review. In January 2026, “tireless junior lawyer” captured excitement about delegating longer sequences of work. As of this September update, the more responsible description is a tool-using system for bounded assignments under human professional control. It can be useful; its performance and permissions must be demonstrated, not presumed.
For the broader framework, begin with what an AI agent is. Then review a legal workflow, set permissions and approval points, and evaluate the run before expanding it to a real practice.
