Less time reading, searching and drafting – the legal judgement stays with the lawyer
I build AI solutions for law firms that cut the time spent reading, searching and drafting: summaries of case files, search across the firm's own templates and memos, and material for time recording.
The sources of law are checked by a lawyer. A language model is not a legal database.
A law firm sells time and judgement. Yet a large share of the time goes to things other than judgement: reading through pleadings, contracts and email threads, hunting for the clause someone wrote two years ago, and remembering afterwards what Tuesday's hours were spent on.
That is where AI can save time already. But few industries have so little room for error, and few have such clear rules about what may leave the building. For members of the Swedish Bar Association (advokater), the Swedish Code of Judicial Procedure (chapter 8, section 4) requires them to observe good professional conduct and to keep secret what they learn in their practice when good professional conduct demands it. Law firms without advokater are not covered by that provision, but have confidentiality written into their engagement terms – and clients expect the same discretion.
That is why an assignment at a firm begins with the question of where client data may be processed. What the tool should do comes second.
What works
Five tasks where AI saves time at a firm
Summaries of documents
A case file of a few hundred pages can become a chronology with parties, dates, amounts and who claimed what – with page references to where it says so. The lawyer still reads the parts that decide the matter. The summary shows where they are.
First drafts
Letters to the opposing party, client emails, standard agreements and the structure of a memo, written from the firm's own templates and wording. The language model writes. Whoever signs is responsible for the content.
Search across the firm's own material
Templates, earlier memos, the firm's own contract clauses and internal notes become searchable with ordinary questions, and the answer shows which document it comes from. It is the same technique as on the page about AI knowledge bases. It searches what the firm has written itself, not statutes and judgments.
Contract review against a checklist
An incoming contract can be compared with the firm's own checklist or standard version, and clauses that deviate are flagged: limitation of liability, termination, dispute resolution. What the deviation means for the client is for the lawyer to decide.
Material for time recording
Calendar, sent emails and document history can become draft time entries with a description of what was done. The lawyer adjusts them before anything is invoiced. Time that would otherwise be forgotten gets recorded, and the client receives a clearer breakdown.
What doesn't work
Cases that don't exist
A language model writes the text that is most likely in the context. Ask it for case law and the result can be references with the right form – parties, year, court, case number – to judgments that were never handed down.
The most widely discussed example comes from a federal court in New York. In Mata v. Avianca, lawyers had filed a brief citing cases that ChatGPT had made up, and in June 2023 the court ordered two of them and their firm to pay USD 5,000. The mistake was not only that the tool guessed, but that nobody checked the citations before the brief was filed.
It does not disappear on its own with new model versions, because it is tied to how language models work. I therefore build no solutions where the model is meant to produce sources of law from memory. References to statutes, legislative history and judgments are checked in the primary source, and the legal assessment is made by a lawyer – it cannot be handed over to a system.
Secrecy and GDPR
Client data in an AI service
The duty of secrecy applies whether a document is read by a colleague or processed by a system. What matters is who gets access to the data and on what terms.
The contract with the supplier. A business service with a data processing agreement, clear information on where the data is processed and stored, confidentiality in the contract and a commitment that client material is not used to train models.
Permissions that follow the matters. Anyone who may not see a matter should not be able to get answers from it either. This applies especially where the firm has walled off matters because of conflicts of interest.
The most sensitive material. Criminal cases, family law and personal injury involve data about health and about offences, which has stronger protection under GDPR (Articles 9 and 10). For that, the language model can run on the firm's own server instead, and that trade-off is described on the page about local AI.
The Swedish Bar Association (Advokatsamfundet) has negotiated framework agreements with several suppliers of AI services for lawyers. That makes choosing a tool easier, but the firm is still responsible for what data is entered and how the tool is used.
The information here describes the situation in November 2026 and is not legal advice – you know that part better than I do. More on what every business needs to have in place is on the page about AI and GDPR.
The approach
From the first conversation to a tool in use
A conversation about the firm's matters. You tell me which tasks take the most hours – reading, drafting or time entries – and get a straight answer on whether a build is worth the money. That meeting costs nothing.
Scoping with the lawyers and assistants. Which types of matter to include, which material must never go in, and who is responsible for keeping the templates up to date.
Testing on closed matters. The tool is first tested on matters where the answer is already known, so the firm can see where it goes wrong before it touches a live matter. The price is fixed and agreed before the build starts.
Connection to the firm's systems. Case management, document archive and email are linked through system integration, so nobody has to move documents by hand.
Training. Everyone who will use the tool is shown what a wrong answer looks like and how it is checked, following the format for AI training for businesses.
Honestly
When the firm should hold off
When the tools you already pay for are enough. Legal databases and case management systems have their own AI features. If they do the job, those are what the firm should learn, not a separate build alongside them.
When the templates are scattered. If the firm's standard wording exists in five versions in five places, a search just gets faster at finding the wrong version. Tidy up first.
When nothing may leave the building and the volume is small. Local operation costs money to set up and maintain. For a handful of matters a month it is rarely worth it.
When nobody has time to check. A draft sent without anyone reading it is not time saved, but a risk moved on to the client.
Background
Who builds it
I am not a lawyer, and law is not what I sell. I build AI systems that are used in live operation, and I say where they don't hold up. The law, and responsibility for it, stays with you.
In my own business an AI team works with a role each. For a firm I suggest that AI prepares the material and that a lawyer reads everything before anything goes out. If AI is to be introduced in more parts of the business at once, it starts with an AI strategy.
The business is called Glänne & Söner and is based in Lindesberg. I take assignments across Sweden, remotely or on site.
That depends on the service and on how the firm uses it, not on the technology itself. No software is “GDPR compliant” on its own.
What needs to be in place is a data processing agreement, clear information on where the data is processed and stored, a commitment that client material is not used to train models, permissions that follow the matters, and a habit of entering only the information the task requires.
Does using AI at a law firm breach professional secrecy?
Not in itself. But putting a client's documents into an external service means handing them to a supplier, and that requires the supplier to be bound by confidentiality in the contract and the firm to know where the data ends up.
A private account in a free chat service is the wrong place for client data, however good the answers are.
Can AI find case law and precedents for us?
Not from memory. A language model can produce references to judgments that look genuine but do not exist.
What works is searching the firm's own material, searching the legal databases with their own tools, and checking every reference in the primary source before it is used.
Will AI replace associates?
No. AI takes away some of the reading and the first draft. The judgement, the client relationship and responsibility for what leaves the firm stay with the lawyers.
How much does it cost?
It depends on which task is to be taken off your hands, how sensitive the material is and which systems it has to work with. The firm receives a fixed quote after a first meeting, free of charge. How a price is usually set is explained in the article how much does an AI consultant cost.
Which hours at the firm go to things other than judgement?
Describe which tasks take time in your matters, and we will go through what can be taken off your hands and what is best left alone. The meeting is free of charge, and any quote has a fixed price.
