On-premise AI for law firms means the models, documents and logs run on servers the firm controls, so client files are not sent to a vendor's cloud.
It is one way to address confidentiality duties, alongside private-cloud and vetted SaaS options. The right choice depends on client terms, matter sensitivity and IT capacity.
This guide is for law-firm CIOs, knowledge-management leaders and general counsel's offices. It compares deployment options, maps them to published ethics guidance, and lists cost drivers and due-diligence questions.
Important: This article is general information, not legal advice. Following any checklist or workflow here does not guarantee compliance with ethics rules or court requirements. Obtain advice for your jurisdiction.
Why Deployment Model Matters for Lawyers
ABA Model Rule 1.6(c) requires a lawyer to "make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client" (ABA Model Rule 1.6).
ABA Formal Opinion 512 (July 2024) says competence, confidentiality, client communication and reasonable fees principally apply to generative AI use (ABA announcement). State rules and opinions may differ.
Deployment model does not satisfy these duties by itself. It changes who can access data and how the firm proves control.
Deployment Options Compared
| Option | Where data and models run | Firm controls | Main trade-off |
|---|---|---|---|
| Public SaaS | Vendor's shared cloud | Settings and contract terms | Fastest; relies on vendor controls |
| Private SaaS tenant | Vendor-hosted, isolated tenant | Contract, tenant settings | Faster than self-hosting; still vendor-hosted |
| Firm's cloud (VPC) | Firm's own cloud account | Network, keys, logs | Needs cloud operations skills |
| On-premises | Firm's data center | Hardware, network, logs | Highest custody; highest operating load |
| Air-gapped | Isolated network, no internet | Everything | Slowest updates; for the most sensitive work |
Decision Checklist
Work through these with the general counsel, IT and risk teams:
• Client requirements: do outside counsel guidelines or engagement terms restrict cloud processing or AI use?
• Matter sensitivity: are there matters (government, investigations, trade secrets) needing isolation?
• Informed consent: does the firm's analysis under Opinion 512 call for client consent for certain tools?
• Ethical walls: can the system enforce matter-level access and conflicts restrictions?
• Logging and retention: can prompts and outputs be retained or deleted under the firm's records policy?
• IT capacity: does the firm have staff to run and patch an on-premise stack?
• Supervision: who reviews AI output before it reaches a client or court?
For supervision design, see how law firms should supervise AI legal review.
This checklist supports a firm's analysis; it does not establish compliance.
Architecture of an On-Premise Legal AI Stack
• Document connectors: DMS and matter-management integration with permission sync
○ ➤ Access must mirror ethical walls, not just folder permissions
• Model serving: open-weight models on firm GPUs
• Retrieval: a search index over approved sources, filtered by matter
• Review workflow: outputs routed to the responsible attorney
• Audit logs: who asked what, which documents were used, who approved
Cost Drivers
• Hardware and hosting for models and indexes
• Integration with DMS, matter management and identity systems
• Security work: access mapping, testing, monitoring
• Staff to operate, patch and support the stack
• Licenses for legal AI software deployed in the firm's environment
Worked Example: Comparing Annual Cost
Illustrative example. Hypothetical figures for a planning exercise, not a quote or client result.
A firm compares a vendor-hosted deployment with on-premises for one AI workflow. It assumes an annual software fee of $90,000 for the hosted option. A vendor states on-premises adds 40% to the software fee plus a one-time setup.
• Hosted: $90,000 per year
• On-premises software: $90,000 × 1.4 = $126,000 per year
• Plus setup (year one): assume $75,000
• Plus firm-side operations: assume 0.5 FTE at $160,000 fully loaded = $80,000
• Year-one on-premises total: $126,000 + $75,000 + $80,000 = $281,000
The firm then weighs the difference against client requirements and matter risk, not cost alone.
Limitations
• On-premise is not automatically compliant: confidentiality also depends on access controls, supervision and staff training
• Models may be less capable: some frontier models are not available for self-hosting
• Updates take effort: patches and model upgrades become the firm's job
• Ethics guidance is evolving: ABA and state bars may issue new opinions; recheck before relying on this analysis
Our own product (disclosure: LuMay sells Lexintis). LuMay's pricing page lists Lexintis legal agents as enterprise-led, from $7,500 per month, and on-premises deployment aimed at law firms and government, from $90K per year plus $75K setup; on-prem adds about 40% per agent.
Lexintis includes Intake Flow, Virtual Paralegal, BillingGuard, BillingOps, OCG Compliance and Metrics Insights agents. Prices read October 30, 2026.
Next Step
If client terms or matter sensitivity push you toward private deployment, discuss your requirements with our team. Discuss a private Lexintis deployment.
Sources
• ABA Model Rule 1.6: Confidentiality of Information
• ABA: First ethics guidance on AI tools (Formal Opinion 512)
