Local, self-hosted or hosted

Should a Law Firm Run AI on Its Own Computers or Use a Hosted Private AI?

Updated November 13, 2026

Short answer: Run AI locally if nothing may leave the office and each lawyer can work alone on a capable computer. Self-host if you want one shared tool under your control and have someone to run it. Choose a hosted tool that clears client content, such as Faradex or Proton Lumo for Business, if you would rather not run hardware or servers yourself. Hosted is less private than fully local, because prompts leave your office, but it is far less work. One caution applies to every local app: the popular ones now also connect to cloud models, and the privacy claim holds only on the local path.

The three ways to run private AI

Where the model runs Who keeps the data Who does the work
Local app On each lawyer's own computer Each computer Each user installs and updates their own copy
Self-hosted On a server your firm runs, or a cloud model the server calls Your firm's server Someone at the firm, or your IT provider
Hosted private AI On the vendor's infrastructure Depends on the vendor; some clear content after use The vendor

Local apps: the most private, the most limited

When a lawyer uses a local model, nothing has to leave the computer. LM Studio says "Nothing you enter into LM Studio when chatting with LLMs leaves your device" (LM Studio). AnythingLLM Desktop is "A private AI assistant that runs entirely on your computer. No accounts, no API keys, no token limits" (AnythingLLM), and its documentation says to choose Desktop if "Everything needs to stay only on your device" (AnythingLLM docs). Jan describes itself as "an open source alternative to ChatGPT that runs 100% offline on your computer" (Jan on GitHub).

The trade-offs:

  • Quality depends on the computer. Larger models need more capable hardware, and only open models run on the local path.
  • No shared workspace. Each person usually runs their own copy. AnythingLLM Desktop's own comparison lists "Multi-user support ❌", and describes Desktop as "a 'single-player' application".
  • Each copy is its own security question. A local app keeps what it keeps on that computer, so the firm's device security, backups and disk encryption now cover AI chats too.

The catch: "local" apps also connect to the cloud

"Local" is no longer a clean category. Each of the popular local apps now also connects to, or sells, cloud models:

  • LM Studio sells cloud plans alongside the free local app. Its pricing page says "Bionic cloud inference is US-hosted, with Zero Data Retention (ZDR)", and its home page says "Our cloud services are Zero Data Retention (ZDR) across the board" (LM Studio pricing, LM Studio).
  • Jan runs offline, but also offers "Cloud Integration: Connect to GPT models via OpenAI, Claude models via Anthropic, Mistral, Groq, MiniMax, and others" (Jan on GitHub), and its home page invites you to "Choose from open models or plug in your favorite online models" (Jan).
  • Ollama titles its site "Run AI models locally and in the cloud", and sells cloud plans. Its home page says "Access the latest open models with complete privacy locally or in the cloud" and "Your prompts are never stored or trained on" (Ollama).
  • AnythingLLM also sells AnythingLLM Cloud, a hosted "Private instance" where you "bring an LLM API Key" (AnythingLLM Cloud). The AI provider whose key you bring applies its own terms; we did not check AnythingLLM Cloud's own retention.

So "we use a local AI app" is not a complete answer. If a lawyer picks a cloud model inside the app, the prompt leaves the computer, and the cloud provider's terms apply. A firm policy should say which models may be used, and someone should check.

Self-hosted: control, with an IT job attached

Self-hosting means running one AI tool for the whole firm on a server you control. AnythingLLM Docker "is both a single-user and multi-user application you can install on any webserver using docker", with "admin and rule-based access for workspaces and documents" (AnythingLLM docs). Open WebUI is another such tool; it keeps "Your chats, users and settings" on your server, so your firm sets the retention (Open WebUI). Its own enterprise page says "Compliance responsibility sits with the deploying organization" (Open WebUI).

The trade-offs:

  • Someone has to run it. Installing the software, applying updates, keeping backups and managing who has access are ongoing jobs, for a person at the firm or your IT provider.
  • Check where the model itself runs. If the firm's self-hosted chat app is wired to a cloud model from OpenAI or Anthropic, every prompt goes to that company under its terms. Keeping the model in-house as well means buying a server powerful enough to run it.
  • Your server becomes the store. Chats and documents live on your server, so they are part of your firm's records, backups and anything you may be asked to preserve.

We do not give hardware or staff-time costs here: they depend on the models you choose and on who does the work, and we have no vendor figure to quote.

Hosted private AI: no hardware, less control

Hosted tools run the AI for you. The ones built for privacy differ from general chat apps in what they keep.

Faradex. Faradex is a private AI workspace for small and mid-sized firms that handle confidential client work. Each firm gets its own Workspace on dedicated infrastructure in the United States. Files and conversations are held in memory and are gone when the Workspace is reset. AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them. Pricing: from $120 a month, no seat minimum; see faradex.ai/pricing.

The honest trade-off: a hosted Workspace is less private than a fully local app, because prompts leave your office and the model is reached through AWS Bedrock. In return, the firm runs no hardware or server of its own, and gets a shared Workspace using Opus via AWS Bedrock.

Proton Lumo for Business. "$11.99 /user per month" billed yearly, self-serve, with a 14-day free trial. Proton says "After Lumo processes your query and generates a response, the data is erased", and that its models "run exclusively on servers Proton controls" (Proton). The catch: open-source models only, and a company "Built and based in Europe".

General business plans such as ChatGPT Business, Claude Team and Microsoft 365 Copilot are hosted too, but they store your chats for a time. The ChatGPT Enterprise page covers what each keeps.

How to decide

If your firm needs... Consider
Nothing to leave the office, and lawyers who work alone A local app with a local model, and a policy on cloud models inside it
One shared tool you control, and you have IT help Self-hosted, with care about where the model runs and how the server is backed up
No hardware or server to run, and a shared Workspace A hosted tool that clears client content
Familiar tools with admin controls, and stored history is acceptable A business plan from OpenAI, Anthropic or Microsoft

Two questions cut through most of the choice. Who will look after it? And where does a client's information sit after the lawyer closes the window? The vendor questions help with the second.

The ethics rules ask lawyers to understand the technology they use; our companion site covers what ABA Formal Opinion 512 asks.

This guide is general information, not legal advice. Check each product's current terms on its own pages before you rely on them.

About Faradex. Faradex is a private AI workspace for small and mid-sized firms that handle confidential client work. Each firm gets its own Workspace on dedicated infrastructure in the United States. Files and conversations are held in memory and are gone when the Workspace is reset. AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them. Pricing: from $120 a month, no seat minimum; see faradex.ai/pricing.