AI for law firms

Why AI in a law firm starts with the case files, not with an "AI lawyer"

Law firms are offered AI from the spectacular side: a chatbot answering clients, "AI that writes pleadings", a tool that "predicts the judgment". The common problem with those projects is that their output is hard to check — and in a lawyer's work an output that cannot be verified in a few minutes does not save time, it moves the time to later corrections. We have implemented AI in law firms and seen where the time actually goes.

It goes into reading. In a case with several volumes of files, someone has to establish who said what and when, and where the documents contradict each other; a three-hour meeting recording has to be listened to in order to pull three sentences out of it; in the court portal every case has to be clicked into every day to see whether a new document or a date has appeared. These are text- and audio-based processes on material the firm already lawfully holds, with rules that can be written down and a result the lawyer verifies by clicking through to the source.

That is why a good first AI project in a law firm looks unspectacular: it gives no advice, writes nothing to the court and talks to no client. It prepares the material the lawyer starts from — and does so in a way that can be checked.

What we actually implement in law firms

• Document and evidence analysis. Reading, classifying and cross-referencing large sets of pleadings, contracts and case files: extracting facts, dates, amounts and parties, pointing to the places where documents contradict each other, and summaries in which every sentence refers to its source page. The model proposes, the lawyer checks and decides — and checking means clicking through to the page of the file, not re-reading the volume.
• Transcription and analysis of recordings. Transcription of hearings, client meetings and recordings from the case file, with the speakers separated and labelled. With video material from the file the system works on the labels a lawyer assigns — who is who in the first minutes — and groups the segments of that recording by those labels, preparing time-coded excerpts for the lawyer to check and for pleadings. This is assistance in reviewing material the firm already lawfully holds, on its own legal basis: the system establishes nobody's identity, matches nobody against external sources, creates no reusable biometric templates and is not a face-recognition service.
• Information maps for litigation. A structured map of the facts, people, documents and dates in a case, with the relations between them — so the case can be navigated instead of being re-read before every pleading. The map is built from the file, every node points to its source, and the whole exports to the system in which the firm keeps its documents.
• An MCP server for the Polish courts' information portal. Our own solution, delivered only as part of an AI implementation in the firm — never as a separate product. It connects the firm's assistants and automations to the Portal Informacyjny Sądów Powszechnych through the firm's own account: case status, new documents, hearing and session dates reach the assistant, the calendar and the document workflow without clicking into every case every day. The integration only reads: it sees exactly what the account's user sees, does not download data in bulk, does not receive served documents and does not file pleadings, and the firm remains the holder of the account and the access rights. What the portal exposes, and what cannot be automated, is described in our article on automating case monitoring in the courts' information portal.
• Drafts citing the firm's own precedents. Draft pleadings that cite earlier pleadings, opinions and rulings from the firm's archive — showing where every argument comes from — instead of an assistant that "knows everything" and cannot say from where.
• Intake, deadlines and anonymisation. Preliminary ordering of client enquiries before a lawyer's call; automatic calendar entries for the dates in pleadings and the portal, with reminders; and anonymisation of documents before anything leaves the firm — the model receives only what a given step needs.

The common denominator: the lawyer approves the result, the system prepares it. Which processes qualify for this kind of automation at all — the five conditions a good candidate must meet — is described on the AI automation page; in a law firm those conditions are checked exactly the same way, only with a stricter requirement of verifiability.

Data, GDPR and the AI Act

Case files, recordings and video are the personal data of parties, witnesses and third parties — often of special categories. The firm processes them on its own legal basis: it determines the Article 6 GDPR basis and, where the material contains special categories of data, the Article 9(2) condition — in litigation usually the establishment, exercise or defence of legal claims (point f). We do not assess that basis; we build to the determination of the firm and its data protection officer. The data stays in the firm's infrastructure or with a provider with an EU processing region that the firm itself chooses, with training switched off and a data-processing agreement in place; the model receives the minimum a given step needs, anonymised wherever possible.

Which AI Act obligations apply to a given tool depends on the firm's role, the system's intended purpose and its risk category: telling people they are dealing with a system (Article 50, where the system interacts with people), human oversight (Article 14 for high-risk systems), no prohibited practices (Article 5). We treat them as the firm's decisions that we build to: a register of systems, logs, an approval mode. A lawyer approving every result is our delivery rule regardless of whether a provision requires it. Classifying the system and assessing its impact are determinations for the lawyer or the data protection officer, not for us — we describe the boundary on the AI Act technical compliance sprint page.

What we do not promise

• Legal advice. Neither from us nor from the system we implement. AI prepares the material — sourced findings, summaries, drafts — while the legal assessment, the strategy and the signature belong to the lawyer. A tool that "advises" is a tool nobody is accountable for when it errs.
• Automatic filing or receipt of served documents. The court-portal integration reads and flags. Receiving a document in the portal has procedural effects, and filing a pleading is an act of the representative — both remain human acts, performed deliberately.
• "AI wins cases". We do not build tools that predict judgments or assess the credibility of witnesses. An offer that promises this is promising a statistic in place of responsibility.
• Biometric surveillance products. We do not build remote biometric identification, face recognition as a service, or systems that infer people's emotions or characteristics. The recording analysis we implement assists the review of material the firm already holds, on labels assigned by a lawyer — and ends there.

How we work

The same way as in every implementation: one process, a number measured before the start — for example hours from receiving a file to a ready summary with references — the narrowest possible pilot, and parallel running before the switch. At first the system prepares and a lawyer approves every result. The full course — four phases and what you get at the end of each — is described by the delivery methodology, and the full scope of the service, from the process audit to maintenance, by the AI implementation page.

A separate topic that comes back in every law-firm conversation is data: professional secrecy, evidence, recordings. Where data goes in a model integration, and what to ask any provider, is set out on the security and GDPR page; the full list of projects we do not take on, on the what we don't do page.

What it costs

Three things drive the cost: the number of systems to connect, the state and volume of the material, and the level of certainty required. The number of lawyers matters less — what counts is whether the files are digital and orderly or go through the scanner afresh for every case. We set the scope and the price after a first conversation; all the components, maintenance included, are broken down in our article on what an AI implementation costs.

Frequently asked questions

Where should a small law firm start with AI?

With one kind of material you already have and currently read in full: the files of one category of cases, recordings of client meetings, or the daily check of the court portal. A good first project has a number measured before the start — hours to prepare a summary of a case file, the number of cases checked by hand in the portal — and a result a lawyer can verify in minutes, because every sentence points to its source page. We do not start with an assistant that "answers everything"; we start with a process in which a mistake is visible at once.

What happens to client data and evidence?

It stays where the firm decides: in its own infrastructure, or with a provider whose processing region is in the European Union, with training on your data switched off and a data-processing agreement in place. Personal data in case files, recordings and video is processed by the firm on its own legal basis — in litigation that is usually the establishment, exercise or defence of legal claims, and for special categories of data the condition in Article 9(2)(f) GDPR — and it is the firm, not us, that determines that basis. Before the integration, not after it, we settle together which data leaves the firm at all, to what extent and after what anonymisation. The rules, and the questions to put to any provider, are set out on our security and GDPR page.

Is the court-portal integration available without an implementation?

No. We deliver the MCP server for the Portal Informacyjny Sądów Powszechnych (the Polish courts’ information portal) only as part of an AI implementation in the firm — we do not sell it separately, nor as software to install on your own. The reason is practical: the integration alone gives the firm a list of changes in its cases, and the value appears only when that list reaches the assistant, the calendar and the document workflow that we configure and test together. The integration runs on the firm’s own account and sees exactly what a user logged into that account sees.

What does automating hearing and meeting recordings require?

A lawful recording. Whether the recording could be made at all — the consent of the meeting’s participants, the court’s permission to record, the electronic record from the portal — is the responsibility of the firm and its lawyers; we work on material the firm already lawfully holds. Technically it takes audio of reasonable quality and, for video, a few minutes in which a lawyer labels who is who. From that point the system separates the voices, groups statements and segments by the labels the lawyer has assigned, and prepares time-coded excerpts for the lawyer to check. It establishes nobody’s identity, matches nobody against external sources and creates no reusable biometric templates.

How is the AI’s output verified?

The way a trainee’s work is, only faster: every finding, date, quotation and summary points to the page of the file or the minute of the recording it comes from, so verification means clicking through to the source rather than re-reading everything. In the pilot we run the work in parallel — the team does what it did before, the system prepares its own version — and we count how many findings were correct, what errors occurred and what they cost. Only those numbers decide which tasks may go on without every result being approved. In litigation the answer to that question is most often: none.

How much does an AI implementation in a law firm cost?

Three things drive the cost: the number of systems to connect — the practice-management software, email, the calendar, the document repository, the court portal — the state and volume of the material, and the level of certainty required. The number of lawyers matters less than whether the files are digital and orderly or go through the scanner afresh for every case. Maintenance is a separate line: usage-dependent fees and periodic tests after model version changes. We set the scope and the price after a first conversation and a short look at the material; the cost components are broken down in our article on what an AI implementation costs.