Asking the right way: how lawyers work with AI tools

The question determines the answer

A language model such as ChatGPT or Claude is not a colleague who knows the file. It only knows what is in the request. Anyone who types in "Write me something about notice periods" gets a text that sounds respectable and fits no specific case. The providers themselves advise thinking of the tool as a clever new member of staff who does not yet know how things are done in the office. The more precise the instruction, the more usable the result.

For small law firms that is good news. Prompting properly, that is, formulating an AI request cleanly, costs no software and no IT department. It costs two minutes of thought before typing. A simple test helps: would a trainee lawyer know what to do with this instruction? If not, neither does the AI.

Role, context, task, format

The specialist literature for lawyers always names the same four building blocks. First the role: "You are an employment lawyer writing for a client with no legal background." The role steers tone and depth. Second the context: who is the text addressed to, what aim does it pursue, what is the starting position? Without context, every answer remains average language. Third the task, as specific as possible: "Summarise the following judgment in no more than 200 words" rather than "What does this judgment say?". Fourth the format: bullet points, table, structure in the style of a legal opinion, length.

An example of a request of the kind that might come up daily in a firm with three lawyers: "You are a lawyer. A client has received an ordinary dismissal and wants to know what time limits apply to a claim. Explain in five short paragraphs, for a layperson and without citing provisions, what steps are now due." The result is a draft that the lawyer reviews and adjusts. Not a finished letter.

It also helps to give the AI a good template from your own files, anonymised. From it the model picks up the structure, tone and level of detail that are customary in the firm. If you have no template, you can first ask the AI to formulate the best prompt for the task itself, and then use that.

Flag gaps rather than fill them

Language models answer confidently, even when they are wrong. The German Federal Bar (BRAK) describes it this way in its guidance on the use of AI from December 2024: the answers often sound plausible but are substantively wrong. The phenomenon is called hallucination. A 2024 study by Stanford University found invented or incorrect information in 17 to 33 per cent of cases even in specialised legal research tools from major providers. With a general model the rate was higher.

A clear instruction in the prompt helps against this: "If you are not certain about a statement, mark it as uncertain. Do not invent sources." A second technique is self-review: "Name the three most important risks in this contract. Then name three weaknesses in your own analysis." Asking for counter-arguments to your own legal position also often yields more than asking for confirmation.

None of these techniques replaces review. The Bar calls it a plausibility check and regards it as a duty. Every source, every date and every time limit from an AI answer is looked up before it goes into a pleading.

Adjust rather than start over

A prompt is rarely finished at the first attempt. The advantage of a language model is the dialogue. Instead of discarding the request, the result can be improved step by step: "Make it more formal." "Cut it by half." "Add an example for the client." Once you have found a request that works well, save it as a text block. Over time this builds a small collection of tried and tested prompts for recurring tasks: client letters, judgment summaries, overviews of time limits, outlines.

An error in the other direction is the question that is too short. Specialist authors recommend deliberately detailed instructions with examples and explanations for complex tasks. Otherwise the model simplifies too early and leaves out precisely the subtleties that matter legally.

Client data stays out

The best prompt is worthless if it breaches client confidentiality. The duty of confidentiality applies without restriction when using AI. The German Federal Bar recommends making only anonymised or abstracted inputs wherever possible. A prompt is therefore built so that it manages without names, addresses and identifiable case circumstances. "An employee with eight years' service receives a redundancy dismissal" is enough for most questions. The names of the client and the employer do not belong in it.

Anyone uploading documents removes all personal data and case names beforehand. Anyone who nevertheless wants to enter client data into an AI system needs a contract with the provider that secures confidentiality and rules out use of the data for training, plus a data processing agreement under the General Data Protection Regulation (GDPR). For most small law firms, the abstract request is the simpler and safer route.