MMOOZZGG.com

AI for lawyers

Local AI for lawyers: a private model on the firm’s own server

5 min read4 sources

Local AI for lawyers runs on the firm’s own server or in a cloud in Russia, so clients’ documents go neither to the vendor of a public service nor abroad. For a lawyer this answers three questions: passing information to a third party, cross-border transfer of personal data, and data localization under Part 5 of Article 18 of 152-FZ, Russia’s personal data law. Below: how a closed model works, where to place it, which tasks it can do, what it costs and who maintains it.

On this page · 5 sections

How a closed model works

A closed neural network for lawyers is an open language model installed inside the firm’s network, with a document base next to it. The model answers from the firm’s documents, and requests and answers never leave the server.

In legal terms, what matters is not the name of the model but where three things are kept:

1. Documents and the knowledge base: in the firm’s systems or in a cloud in Russia.

2. Model requests and answers: in a log on the same server, not with an outside vendor.

3. Access rights: each lawyer sees their own cases, and the AI has a separate account with rights only to its own process.

For an advocate this is a matter of Article 8 of Russia’s Law on Advocacy, which protects attorney-client privilege: the client’s information stays with the advocate and the firm, and there is no outside recipient in the chain. It is still worth getting the client’s consent to AI, but its subject is different: processing inside the firm, not transfer to an outside service.

What such a solution consists of and how to update the model without internet access is covered in the “Sovereign AI” section, on the page about AI in a closed network.

A server in the office or a cloud in Russia

A server in the office gives the firm the most control: the hardware is its own, access follows its rules, and there is no intermediary in the chain.

Renting capacity in a Russian cloud for the pilot helps test the process before you buy a server. The same rules on data processing on instruction apply to it as to any cloud.

A cloud in Russia is easier to launch, but the cloud provider becomes a party that processes data on the firm’s instruction. You need a contract with the terms of Part 3 of Article 6 of 152-FZ: the list of data, the actions, the purposes, confidentiality and protection requirements. Also check where the servers physically stand.

In both options, data of Russian citizens is stored in databases on Russian territory, as Part 5 of Article 18 of 152-FZ requires. Under Part 8 of Article 13.11 of Russia’s Code of Administrative Offenses, breaking this rule costs a legal entity up to $72,000.

Since 1 September 2026, 243-FZ has been in force: data of sovereign and national models is processed in Russia, and the Government decides when such models are mandatory. More in the review of 243-FZ for business.

Which tasks an open model can handle

An open model with Russian on the firm’s server is enough for work with documents the lawyer already has:

1. A case summary and timeline from the materials: the claim, the response, correspondence, transcripts.

2. First-pass contract review against the firm’s checklist and a search for deviations from the standard position.

3. Drafts of letters, claims and standard documents from the firm’s templates.

4. Search over the firm’s knowledge base: past opinions, positions, samples.

5. Anonymizing documents before they leave the firm.

A closed network has no internet access. The model takes court practice only from databases loaded inside, and a lawyer checks every reference. AI hallucinations in courts show why this cannot be skipped.

From the model you expect a draft and the found fragment of a document, not a finished opinion. The position on the case, the decision and the signature stay with the lawyer.

Price and timeline

The cost is made up of hardware and rollout. Hardware prices from the “Sovereign AI” section as of 29.09.2026:

1. A MacBook Pro laptop in the top memory configuration: $7,960.

2. Mac Studio M3 Ultra in the top memory configuration: $37.4K.

3. An RTX PRO 6000 graphics card for a server: $12.8K.

4. An H100 accelerator for a server: $27.1K.

5. Renting an H100 in a Russian cloud: from $3 an hour.

Rollout: a pilot on one process from $1,499 in 14 days, not counting hardware. An AI department of 3–5 AI agents from $4,999 in 4–6 weeks.

A pilot usually starts with one process, such as case summaries, and the hardware is chosen for the volume of documents in that process. If the calculation shows a cloud in Russia is cheaper than your own server, you see it before you buy hardware.

The full monthly cost with depreciation, electricity and maintenance is worked out by the local AI cost calculator. A comparison with other options is on the page about the cost of AI for lawyers.

Maintenance

A model on your own server needs support: model updates, checks of answer quality after each update, control of access rights and the log.

A new model version is first tested on a set of the firm’s typical requests and only then made the main one. The old version stays for rollback.

Support costs $480 a month. If the firm drops support, the system keeps working: the system, the data and the access stay with the firm, and the model can be replaced without rewriting the rules.

For a lawyer, one thing matters most in maintenance: every settings change is recorded, and in an audit you can show who had access to a case and when.

The technical data protection rules are in the block below, and what breaks in a server and how to fix it is on the page about maintenance and failures.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. Brace Law Firm (brace-lf.com)
  2. ppt.ru, 26.07.2026 (ppt.ru)
  3. ConsultantPlus, Federal Law No. 152-FZ of 27 July 2006 “On Personal Data”, Article 6, 27 July 2006 (consultant.ru)
  4. MMOOZZGG.com prices and terms

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min
Contact us