An AI assistant helps a person complete work through a conversational or task-based interface. It may draft text, summarize a document, search a source, or use a connected tool. “Assistant” describes its role in relation to the user; it does not tell you exactly how much it can do or whether a person must prompt every step. For legal work, define the task, the information the system may access, and who approves its result.
What can an AI assistant do?
An assistant can turn an outline into a first draft, explain unfamiliar language, extract terms from documents, organize a research plan, or answer questions about material provided to it. Depending on the product, settings, and connected tools, it may also search the web, inspect files, run calculations, or work across several steps. OpenAI's current ChatGPT capabilities overview describes an assistant with optional tools and multi-step research. These are product capabilities, not promises of accuracy or permission to use client information.
For example, a lawyer might ask for an issue list from a fictional contract. The assistant can identify clauses to inspect and propose questions for the client. The lawyer still must compare the list with the actual agreement, governing law, negotiated context, and instructions. A fluent summary can omit an exception or mistake a draft for a signed version. An assistant's output is a starting point for review.
There are several ways an assistant can produce that list. It might answer from a clause pasted into the chat, retrieve passages from an approved document collection, or use a connected research service. The visible interface may look the same in each case, yet the evidence differs. A response about an uploaded draft is not evidence about an amendment the system never saw. A response citing a retrieved file is not evidence that the file is the operative agreement. Ask what material the answer actually used.
Useful output can be more specific than a polished paragraph. For the fictional contract, request a table of issues with the clause text, page or section, why it may matter, and a column for unresolved facts. The assistant should be allowed to say that an exhibit is missing or a defined term is inconsistent. The lawyer can then review each row against the signed documents. This format makes omissions easier to detect; it does not prove there are no omissions.
Is an assistant different from an AI agent?
An assistant is defined by whom it helps; an agent is defined by how it chooses and carries out steps. A conversational assistant may use an agent behind the scenes. A fixed workflow may also sit behind an assistant: software could always search a named repository, summarize the retrieved pages, and place a draft in a queue. An agent has more discretion to choose the next tool or step after seeing a result. The same product can offer all three modes.
The distinction does not turn on whether software waits for a user to start a task. A person can initiate an agent, set limits, and require approval at checkpoints. Nor does an agent become a lawyer because it completes more steps. Anthropic's engineering guide separates predefined workflows from systems where the model directs tool use; it also recommends the simplest design that meets a task. Our AI agents guide explains the model–tool loop, and the Agents section follows it into legal workflows and oversight.
Consider one request: “Compare these two versions of a lease.” In a simple assistant exchange, a person provides both versions and receives a comparison. In a fixed workflow, the application always extracts specified clauses, computes changes, and generates the same report format. In an agentic flow, the model may decide that an incorporated exhibit must be opened before it can answer, then ask for that exhibit when it is unavailable. All three may appear inside one assistant product. The amount of tool discretion, rather than the branding or chat layout, distinguishes the underlying process.
The right mode depends on the assignment. Repetitive, well defined extraction may favor a fixed path whose checks can be rehearsed. An open-ended question across varied materials may benefit from additional searches, provided the user can see the sources and stop the process. Greater freedom is not automatically better; it can increase latency, cost, and the chance that a bad intermediate result shapes the final answer.
Does the assistant remember earlier conversations?
The answer depends on the application. A model can use conversation context that is supplied with the current request. An application may also save memory or retrieve a past chat, file, or preference and add it to later context. Neither operation, by itself, means the model has been retrained. Training or fine-tuning is a separate process that changes model parameters. Ask what is stored, how long it is kept, who can access it, whether it is used for model training, and how it can be deleted. OpenAI's memory documentation illustrates application-level memory; Anthropic's context explanation explains how instructions, history, and retrieved material are selected for a model call.
A saved matter note can be helpful and still be out of date. A long conversation can also lose important detail when context is limited or compressed. For legal work, return to the source document, verify its version, and check any cited authority instead of trusting a remembered summary. Our context-window guide explains this limit in more detail.
The three mechanisms lead to different outcomes after a correction. If a lawyer says “that date is from the draft, not the signed agreement,” the assistant can use that correction while it remains in the current conversation context. If the application saves the correction and retrieves it in a future session, the later answer may reflect it. If the application does neither, the correction need not carry forward. None of these steps automatically changes the model's parameters or guarantees that it will make the same distinction on another matter. OpenAI's conversation-state guide explains how applications can carry earlier messages across model calls within context limits.
Memory controls are product and workspace features, not a universal property of AI assistants. OpenAI's current ChatGPT memory documentation says available controls vary by plan and workspace and that memory does not retain every detail. It also separates personalization features from model-training data controls. A firm should evaluate the actual account and product configuration it will use, including connected apps and retention, before treating an assistant as a place to store matter knowledge.
Where can an assistant help in legal work?
Useful starting tasks are bounded and reviewable: creating a checklist from a matter plan, comparing two drafts for changed language, organizing a deposition outline, or suggesting research questions. Use fictional or approved data for initial tests. For a real matter, select tools whose terms and configuration meet confidentiality and security obligations. Document what the assistant saw and what the human checked. A tool should not receive an entire client file when a redacted excerpt would suffice.
A bounded request names the source, output, and limits. For example: “Using only the attached fictional lease and amendment, list each assignment-consent provision, quote its section, identify any missing exhibits, and do not infer terms from documents you cannot open.” A reviewer can check each quote and missing-document flag. The wording helps define the assignment, while the application's access controls still determine what files or external tools the assistant can actually use.
Some tasks call for greater caution even when the requested output seems routine. An assistant might summarize a deposition transcript, but the reviewer must check whether a negation, speaker attribution, or objection changed the meaning. It might draft a client update, but the responsible lawyer must determine what can be said and approve delivery. It might suggest a deadline, but the deadline must be verified through the authoritative order, docket, and governing rules and entered through the team's established calendar process. A chat answer is not a deadline control.
The assistance should be judged by the complete workflow, not the first draft alone. Count missing facts, incorrect quotations, false or inapplicable citations, and the time needed to repair the answer. A repeatable set of examples is more informative than a single polished demonstration. OpenAI's agent evaluation guide describes traces and test datasets for assessing multi-step systems; the same habit of measuring errors applies to simpler assistants.
What must a lawyer verify?
Check facts against the record, citations against primary authority, dates and jurisdiction against the task, and every external communication or filing before release. Protect client information when choosing a product and connector. Set rules for who can use a tool, what matters may be entered, and who signs off on final work. ABA Formal Opinion 512 discusses competence, confidentiality, supervision, candor, and fees in lawyers' use of generative AI. The ABA opinion addresses Model Rules; applicable jurisdictional rules, court orders, and client terms still govern.
Verification should match the claim. For a quoted clause, open the cited agreement and check the exact words and document version. For a statement of law, read the authority and check its treatment and jurisdiction. For a negative claim such as “no consent requirement appears,” inspect the search coverage, definitions, schedules, and exceptions because a citation to one paragraph cannot prove that no other provision exists. For an executive summary, compare the summary against the full set of source documents and the purpose of the advice. Record material corrections so the team can decide whether this use remains worthwhile.
If a bounded assistant repeatedly succeeds under review, a team may consider a tool-connected workflow or agent for tasks whose steps vary. Permission scope and human checkpoints should grow with the consequences of an error. The permissions and oversight guide sets out those controls; a legal workflow example shows where review belongs.
Sources and update date
This guide was reviewed September 27, 2026. Product features and legal rules may change. Technical sources include OpenAI's assistant capability overview, OpenAI's memory explanation, conversation-state documentation, Anthropic's agent design guide, and Anthropic's context-engineering article. The legal source is ABA Formal Opinion 512.
