Two hundred pages of documents and a hearing on Tuesday: AI can draw up the summary. Six mistakes keep coming up, exposing the file or making the summary wrong. Here is how to avoid them.
Friday, six in the evening. A colleague hands you a case you are taking over at short notice: two hundred pages of documents, letters and written submissions, and a hearing on Tuesday. The question everyone asks: can I ask an AI to summarise it for me, and am I allowed to give it these documents?
The answer is yes, on two conditions: the documents must stay within a framework you control, and you must know what the AI has actually read. The mistakes below are the ones most often seen in law firms and in-house legal teams. Each one has a simple fix.
A lawyer's professional secrecy covers the documents, the exchanges with the client and even the very existence of the case. In Switzerland, its breach is punishable under Article 321 of the Swiss Criminal Code, and most other countries have equivalent rules. In-house lawyers and notaries have their own duties of confidentiality, often reinforced by contractual commitments. Summarising a case file with AI therefore starts with deciding who gets to see the documents. The guide on AI for lawyers sets out the general method; this one focuses on the large file that needs summarising fast.
The temptation is strong when time is short. But a consumer AI tool is a third party: it often keeps the history, and its provider may be subject to foreign laws that require it to hand over what it holds. "It was only a summary" changes nothing: the documents have left your hands.
The fix: work in a "Legal" space. It sets a formal tone and keeps the folders you add on the confidential models, open models installed on servers in Switzerland. For a summary, they are all you need.
You black out the client's name, then send the rest. Yet a case file is full of details that identify someone as surely as a name: the case number, the address of a property, the amount of a transaction, the date of a dismissal, the employer's name. Put together, they point to one person.
The fix: do not rely on redaction done by hand. The guide on anonymisation and pseudonymisation explains why. The simplest approach is still to keep the file on a confidential model. If you choose an external model for a side question, the Confidentiality filter first detects sensitive data and offers to answer with a confidential model; a document that is too identifying is handled by a confidential model, or only the question is sent.
This is the most misleading mistake, because the summary you get looks complete. A request can only hold a limited amount of text: beyond that, an attached document is marked "truncated", and the end of the file has simply not been read. The summary seems consistent, but it may ignore the decisive document, often the most recent one.
The fix: work in batches. Attach the documents in coherent bundles (procedure, correspondence, the other side's documents) and each time ask for a list of what has been read.
Here are documents 1 to 25 from a commercial lease case. For each document, give in one line: what it is, its date, its author and what it establishes. End with a list of the documents you could not read in full, or where part seems to be missing.
You then combine the batches into a single summary sheet.
On a computer, with Chrome, Edge or Opera, you can add the case folder to a space exactly as it exists on your disk, with the "Confidential models" option: it is then never sent to an external model. With a signed-in account, the assistant automatically finds the passages relevant to your question. This works very well for a precise question ("what does the lease say about the indexation clause?"), much less well for an overall summary: the assistant answers from the passages it has found, not from a full reading.
The fix: use the added folder for targeted questions, and batches of attached documents for the overall summary. On another browser or on a phone, attaching documents remains the normal route: they are read in your browser.
A summary with no reader in mind is flat: it follows the order of the documents, puts everything on the same level and does not tell you what matters for Tuesday. Yet what you need is not a summary but a hearing brief, a note to the partner or an update for the client.
The fix: state the purpose, the reader and the format. The space's instructions can set your brief format once and for all; the request specifies the rest.
Using the batch summaries above, write a one-page hearing brief for a lawyer taking over the case: each party's claims, the three questions the judge will have to decide, and for each one the documents that support us and those that work against us. Cite each document by its number and state nothing that does not appear in a document.
AI can be wrong with confidence: a shifted date, a document from the other side attributed to your client, a recalculated sum, an argument nobody raised. Across two hundred pages, these slips easily go unnoticed, and they end up in a pleading.
The fix: require a document number for every assertion, as in the request above, then check the points your argument rests on. Anything not tied to a document is suspect by default. The summary saves you the first read; it does not replace yours, and responsibility for what you argue remains entirely yours.
In a Legal space, the folders you add stay on the confidential models: they are processed on servers in Switzerland and are never passed to an external model. The index used to find passages stays in the folder, on your computer. Attached documents are read in your browser, and no file is kept on our servers. If you choose an external model, only pseudonymised extracts can be sent, and always after the Confidentiality filter.
Your conversation history, summaries included, lives in your browser and nowhere else: we could not produce it if asked to. On a shared computer at the firm, however, it can be read by anyone using the same session: work in a personal session, and store the final brief in the case folder rather than relying on the history.
Before opening a real case file, reread the data you should never paste into an AI, then try the batch method on a fictitious case: open the chat.
