A short summary with page links
A partner needs a short case summary before a meeting, and so do a new lawyer on the team and a client who wants to understand where they stand. AI prepares it from the firm’s template: parties, subject of the dispute, claims, positions, key documents, deadlines, open questions.
The main rule: every statement links to a document and a page. A summary without links cannot be checked, and so it cannot be used in work.
The firm sets the summary template, and it is the same for all lawyers: the partner gets summaries in one format, whatever the case.
Summary formats for the task:
1. One page for the partner: the essence, the risks, and what to decide this week.
2. A detailed summary by volume for the lawyer who runs the case.
3. A summary of testimony: what each witness said and where it differs from the documents.
4. A letter to the client in plain language, which the lawyer edits and sends personally.
When new documents come into the case, AI updates the summary and shows which conclusions changed. The partner sees not a whole new version but the difference from the last one.
From such a summary, trial preparation follows: a timeline, a map of evidence, and questions for witnesses.
Finding facts and contradictions
The second mode of work is a question to the materials. The lawyer asks when the buyer first claimed defects, or who signed the acceptance reports for the contractor, and AI answers with quotes that link to pages.
In correspondence of hundreds of emails, AI collects threads by topic and finds the email where a party first admitted a delay or agreed to change a term. By hand, such an email takes the longest to find.
AI looks for contradictions by comparing documents:
1. Dates: the acceptance report has one acceptance date, the email another.
2. Amounts: the invoice and the payment order for the same delivery show different amounts.
3. People: a document was signed by a person whose power of attorney gave no such authority.
4. Positions: in its response, a party admits a fact that it denied in its claim letter.
The lawyer checks each contradiction against the source documents. The model finds candidates, and a person decides what it means for the case: even legal services in the Stanford RegLab and HAI test made errors in one answer in six or more often.
This is what AI for document work looks like when it is built into the legal process: the machine reads and compares, and the lawyer checks and decides.
Scans and handwriting
Case materials rarely arrive as text. Scans, photos of documents, and handwritten notes first go through recognition, and AI marks the pages where quality is low.
What matters for a lawyer:
1. The recognized text is stored next to the original, and the link leads to the page of the scan, not to a retelling.
2. AI marks handwritten notes, stamps, and signatures separately and does not present them as the document’s text.
3. Pages with poor recognition go onto a list for manual checking.
A quote from a scan in a court document is checked against the original. A recognition error in one digit changes an amount or a date, and the text does not always show it.
For tables and statements, AI keeps the structure: rows, columns, totals. If the total of the rows does not match the total in the document, it goes onto the list of contradictions.
AI reads materials in several languages together, such as a contract in English and correspondence in Russian, and quotes in the original language, so the lawyer checks the text and not a translation.
Where to store the materials
Case materials are protected by privilege. Under Article 8 of Russia’s Federal Law on Advocacy, attorney-client privilege covers any information related to legal help for the client, and only the client can lift it. A case volume in a public AI service means handing this information to a third party.
So the materials stay inside the firm’s systems. In our deployments, personal data is processed on servers in Russia under 152-FZ, Russia’s personal data law; the text does not go abroad or to services without a data processing agreement, and the system, data, and access stay with the client.
For any outside service, the rule is simple: if there is no data processing agreement with the vendor, case materials do not go there.
AI rights are limited: a separate account sees only the cases it was admitted to, and its actions are logged. The model does not follow commands from uploaded documents: if an email asks to forward the materials, that stays the text of the email.
If the firm wants to keep the model in-house, it is installed on the firm’s server or in a cloud in Russia: the options are covered on the page about local AI for lawyers. Summary templates and past cases are easy to keep in the knowledge base of a law firm.