Open-weight AI models now run on a workstation, a server in your equipment room, or GPU servers in a data center. For work the firm prefers not to send to an outside model provider, we select, install, secure, and train your firm on a local setup — and review everything else in its data path.
The question is no longer "should we use AI?" — in Clio's 2025 Legal Trends Report, 79% of legal professionals surveyed said they already do. The question is where the data goes when they do. Scroll the timeline; the diagram follows.
Review the account settingsTraining use, retention, access and account ownership depend on the product and settings.
Verify the plan's controlsConfirm the agreement, training use, retention, access, audit features and offboarding.
Review the complete setupCheck the model license, telemetry, connectors, backups and remote support.
Personal account. In this example, the provider processes prompts under the account's terms. Training and retention choices vary. Check who owns and can administer the account, and which uses the firm's policy permits.
Business plan, configured. The provider processes requests under the applicable agreement. Available controls vary by product, plan and feature; verify their settings. Search, connectors and agents may involve additional services.
Local, open-weight model. With local inference configured, the model processes prompts on firm-managed hardware. Review any other services that receive data. Cloud, local and hybrid setups can each suit different tasks.
A personal account may offer useful capabilities for permitted work. Review the current product terms and settings: whether inputs may be used for training, how long data is retained, who can access it, and who owns the account. Also confirm whether the firm can administer access and offboard departing staff. The account label alone does not establish which tasks or data are appropriate.
Team, Business, and Enterprise plans of Claude, ChatGPT, Gemini, and Copilot generally exclude business data from training by default, and their agreements — often with a data processing addendum built in — set out how data is handled. Retention settings, single sign-on, audit logs, and in some cases zero-data-retention or a HIPAA business associate agreement can be added. These are different protections: no-training is not zero retention, and features such as web search, connectors, file stores, and agents can change where data goes. We set this up under AI & LLM Adoption and document what we verified.
Open-weight models from Qwen, Google (Gemma), OpenAI (gpt-oss), Meta (Llama), Mistral, and DeepSeek can now handle much of the high-volume work firms need — summarizing, extracting, searching, first drafts — on hardware a firm can buy. Configured local-only, prompts are processed on that hardware rather than sent to a model provider. That reduces outside processing; it doesn't remove every other party. Runtime telemetry, update channels, remote support, connectors, backups, and software licenses all still need review. It is a third option beside a well-configured business plan, not a replacement for one.
ABA Formal Opinion 512 asks lawyers to understand where client information goes when they use generative AI and to protect it. In California, SB 574 — effective January 1, 2027 — adds a specific condition: confidential, personal identifying and other nonpublic information may go into a generative AI system only if access to those inputs is restricted to the attorney and authorized people bound to confidentiality. A local model can make that question easier to answer for the model itself. The rest of the setup still has to be checked.
This does not replace the research platforms. Lexis+, Westlaw, and Midpage bring the authoritative databases and citators that an open model doesn't have.
Local AI is not one product. Below are three typical starting points — illustrations, not a rule that maps firm size to hardware. The real selection is measured: the model you need plus its context and KV cache, simultaneous users, acceptable latency, output quality on your own tasks, retrieval and tool use, storage, backup, recovery, operating effort, and how you would exit. Prices are September 2026 street ranges; we quote at order time.
One machine on one desk, running a capable model against the firm's own documents. No server room, no rack, no specialist on staff.
We configure and then test that the runtime stays local-only — several popular tools added optional cloud routing in 2026 — and we check telemetry and update settings while we're there.
A GPU server in a lockable equipment room, serving the whole firm through single sign-on, with a document index that only shows each person the matters they are allowed to see.
The most common mistake in firm-wide document search is a retrieval index that ignores ethical walls. Matter-level permissions are not optional here; we build them in from day one.
For firms with data-center space, in-house IT, and workloads or client requirements that justify it.
We design, specify, and govern; installation runs alongside your data-center partner and an authorized NVIDIA reseller. Most large firms start with a few nodes, not a full rack. Product names are used to describe compatibility; Firm Upgrade is an independent consultancy, not an NVIDIA or Apple partner or reseller unless expressly stated.
"Open-weight" means the model file is published and can be downloaded and run. Licenses differ and matter for a law firm: we check each model's license against your intended use and record it in the roadmap.
Retrieval-augmented generation (RAG) is what turns a chat window into "ask our files." Done carelessly, it also becomes the fastest way to breach an ethical wall inside the firm. Ours is built around three rules:
For a lot of work, a business plan can be the right home — if the agreement and the admin settings match what you need. Four steps we run on every deployment, for every vendor.
Identify the exact product, plan, and agreement — including any DPA, which is often incorporated into the terms rather than signed separately — and what it says about training, retention, subprocessors, and support access. Request zero data retention or a HIPAA BAA only where it is offered for your features and the practice needs it.
Verify the tenant's training setting, then review features that change the data path — web search, connectors, file storage, agents — and disable what nobody approved or what falls outside the reviewed terms.
Conversation retention aligned to your records policy, single sign-on so offboarding actually offboards, and logs you can produce if a client or regulator asks.
A dated evidence record — the agreement and any DPA that apply, screenshots of the settings, the features in use, the written AI policy — that makes a client security questionnaire far quicker to answer accurately.
What each vendor publishes about training, contracts, and retention — summarized September 2026, with the API and feature notes rechecked October 2, 2026. These are different protections: no training is not zero retention, a DPA is not a BAA, and none of them is the same as your duty of confidentiality. Terms change often — we re-verify every one at deployment time.
| Platform | Trains on your data? (business tier) | Contract (DPA) | Retention controls | Where it runs | Our take |
|---|---|---|---|---|---|
| ClaudeAnthropic · Team / Enterprise | No, by defaultConsumer plans may train unless the user opts out | Included in commercial terms | Admin-set retention; API ZDR by arrangementStateful API features (e.g. Files, Batch, code execution, Managed Agents) are not ZDR-eligible | Anthropic cloudCertifications and BAA availability confirmed for the plan at deployment | Our go-to for research and drafting; legal connectors including Midpage. |
| ChatGPTOpenAI · Business / Enterprise | No, by defaultConsumer plans train unless "Improve the model" is off | Executable self-serve | Workspace retention settings; API ZDR by approvalAPI stateful endpoints (Assistants, Threads, Vector Stores, Conversations, Agents, Batch, fine-tuning) are not ZDR-eligible | OpenAI cloudCertifications and BAA availability confirmed for the plan at deployment | The one your staff already knows — which is exactly why it needs governance. |
| GeminiGoogle · Workspace / Gemini Enterprise | No (qualifying Workspace editions)Consumer Gemini and third-party apps used with Gemini are under different terms | Built into the Workspace / Cloud DPA | By contractual amendmentEligible enterprise customers | Google CloudScope differs between Workspace Gemini and Gemini Enterprise — we confirm which you're buying | The natural fit for firms already living in Google Workspace. |
| Microsoft 365 CopilotMicrosoft · commercial tenants | No — Enterprise Data Protection | Automatic in Microsoft's DPA | Your tenant's retention policies apply | Your Microsoft 365 tenantGenerated web-search queries go to Bing; Microsoft states its DPA does not apply to them. Web search can be turned off by policy | For firms in Word, Outlook, and Teams all day. Tenant hardening comes first. |
| Clio Manage AI & VincentClio · vLex | Verified in your Clio agreementWe confirm the training and retention terms before enablement | Clio subscription terms | Per Clio policy | Clio cloud | AI inside the practice-management system your matters already live in. |
| Lexis+ AI with ProtégéLexisNexis | No — "never uses customer data to train" | Vendor agreement | Customer-held encryption keysAdded May 2026 | LexisNexis cloud | Authoritative research with Shepard's Verify; the citator an open model can't give you. |
| CoCounsel LegalThomson Reuters · Westlaw | No — prompts not used to train | Vendor agreement | Per TR policy | Thomson Reuters cloud | Agentic research and drafting grounded in Westlaw and Practical Law; next generation went GA in August 2026. |
| MyCase IQMyCase · built on OpenAI | No — contractual with OpenAI | Vendor agreement | Per MyCase policy | MyCase cloud | Inherits MyCase's own permissions — a good fit for small firms already on it. |
| MidpageAI-native case law | Confirmed in terms at deploymentSOC 2 Type II | Vendor agreement | Per Midpage policy | Midpage cloud | Research and cite-checking inside Claude through its MCP connector. |
| Local & open-weightQwen, Gemma, gpt-oss, Llama, Mistral, DeepSeek | Not by a model providerCheck runtime telemetry and any optional cloud features | Model license + other vendorsOS, runtime, support, and backup vendors still have terms | Set by the firm | Hardware you control | A third option for particular matters — after the rest of the data path is reviewed. |
Summarized from each vendor's published privacy, enterprise, and trust documentation as of September 2026; the Anthropic and OpenAI API retention notes, Google Workspace privacy hub, and Microsoft Copilot web-search page were rechecked October 2, 2026. Public documentation describes what a vendor offers, not the settings or agreement of your account. Not legal advice; not a substitute for reading the agreement you sign. Product names belong to their owners.
Six modules, taught separately to attorneys, staff, administrators, and the people who maintain your automations — synthetic examples first, then your approved workflows. Each module maps to a duty from ABA Formal Opinion 512, state guidance, and, for California, SB 574. Not MCLE-accredited unless a course says so in writing.
What a language model is and isn't, why it hallucinates, where it is reliably strong, and how to tell the difference before it reaches a client.
Duty: competenceYour firm's own map: which approved tool may receive which kinds of information, based on the reviewed access and settings of each — including the personal identifying information SB 574 lists. A one-page guide every person keeps at their desk.
Duty: confidentialityHands-on with the admin console: confirming the agreement, training and sharing settings, retention, SSO, audit logs, feature controls, offboarding, and keeping a dated record of what was verified.
Duty: confidentiality & supervisionCite-check discipline for AI-surfaced authority using Shepard's, KeyCite, or Midpage's citator — and the attorney's own personal verification of every citation in a court filing, recorded against the version filed.
Duty: candor to the tribunalCourt disclosure for AI-assisted filings, the separate decision about public content, when client consent may be needed, engagement-letter language your attorneys approve, and how AI-assisted work is billed.
Duty: communication, candor & feesYour written AI policy — what's allowed, what's prohibited, who reviews what — plus the quarterly refresh that keeps it current as the tools change.
Duty: supervisionDelivered live, recorded for your library, and refreshed as your tools change — the same formats as our Training & Enablement service.
Part of the fixed-fee assessment: which work might run locally, how much of it there is, and whether local AI is justified at all — with a short pilot on representative, non-confidential tasks.
Hardware from the measured workload, model choices and licenses, RAG design with matter-level permissions, the full data path, physical and network security, backup, recovery, and exit — priced at order time.
Build, runtime tested local-only, telemetry and updates configured, SSO, logging, a tested restore, and a permission test against the retrieval index before anyone relies on it.
The AI Fluency Program for every role, runbooks and an inventory, then optional support with written scope for patches, model updates, and the questions that come up in week three.
The fixed-fee assessment maps where your data goes today, measures the workload, and tells you honestly whether you need local hardware at all.