Does your firm need on-premise AI deployment?

Quick answer: HHS guidance addresses cloud use for electronic protected health information, or ePHI, by HIPAA covered entities and business associates. It permits that use under specified conditions. This guidance does not settle what every US law requires of an AI deployment.

On-premise AI deployment means the AI model, your records, and the logs all run on hardware you own, with no calls to an outside AI service. Few firms need it. Omni Studio builds private AI setups for firms that have those terms, starting in the firm's own cloud account.

What HHS says about HIPAA and cloud services

A HIPAA covered entity or business associate may use a cloud service to store or process ePHI when it signs a HIPAA compliant business associate agreement with the provider and otherwise complies with HIPAA. Risk analysis and risk management still apply. This is guidance about HIPAA and cloud services for ePHI, not a conclusion about all US laws or every AI use. HHS cloud computing guidance

What does force a private AI deployment?

Client contracts. The FDIC's outside counsel deskbook (§1.17, April 2026 edition) bars sensitive information in any generative AI tool that runs in an "open environment." Its compliant "closed environment" keeps models inside the organization's own infrastructure, "ensuring user data never leaves established security firewalls." The Texas Attorney General lets outside counsel use internal AI only where the environment meets TX-RAMP or FedRAMP (§2.2, July 31, 2025).

Some clauses ban AI on client work, and no setup cures them. New York State DOT prohibits "any use of Artificial Intelligence utilizing Department information" (consultant instructions, item 5, October 6, 2025).

These terms are still rare: 20 of 1,054 published outside counsel guidelines carry AI terms, and one bans AI (Fulkerson Advisors, September 2026). Read your own client agreements before you price hardware.

Vendor defaults are the other driver. Zero retention is never the default. Anthropic's Covered Models, which include Claude Fable 5.1, "require 30-day data retention" (Anthropic BAA page, October 21, 2026). Microsoft says its personnel review prompts and completions for harmful content (Microsoft Learn, May 19, 2026). A June 2026 Commerce directive took two frontier models offline for 19 days (Mozilla, September 15, 2026).

How do the four setups compare?

Costs are public market prices, shown for context only.

Setup	Who it fits	Market cost context	What it protects	What it leaves open
Public cloud with a BAA	Practices with patient data and no client clause naming the environment	The BAA is mostly self-serve at no added charge (the five major vendors' pages, October 21, 2026). Claude API input costs $0.10 to $10 per million tokens (claude.com/pricing)	No training on your data, by contract, and a BAA	Up to 30 days of retention at OpenAI without zero retention; reviewer access at Microsoft; data leaves your account
Private cloud	Firms whose clients or licenses want data in an account the firm controls	Per-token model fees plus your cloud bill. An H100 GPU rents for $2.85 to $3.25 an hour (Silicon Data, October 21, 2026)	Records, files, backups, and logs stay in your AWS or Azure account. Bedrock content "is not shared with any model providers" (AWS) AWS data protection	Model calls run on the vendor's service under its terms, and the vendor controls availability
Hybrid	Firms with one class of data a BAA or consent cannot cover	Private cloud plus local hardware. A 256GB Mac Studio M5 Ultra is about $11,299 (kingy.ai, October 8, 2026)	Records you flag go only to a model on your hardware	A routing rule that misses a record sends it to the cloud
Fully offline	Defense suppliers with CUI or ITAR data, and clients that forbid outside processing	About $11,299 for a 10 to 25 seat desk machine (our estimate) up to $275,000 for an eight-GPU H100 server (Viperatech, October 21, 2026)	Nothing leaves your network, and no vendor can switch the model off	Lower model quality; you own patching; no vendor BAA, SOC 2 report, or on-call team

Cloud models score higher: on GDPval-AA office tasks, the top closed model scores 1866 and the best open model 1746 (Artificial Analysis, October 21, 2026).

Which setup fits your industry?

Healthcare and dental

HIPAA requires a BAA and a documented risk analysis, and every major cloud vendor signs a BAA. The case for more control is data a BAA cannot cover. Substance use records under 42 CFR Part 2 need a qualified service organization agreement, with compliance due February 16, 2026 (Federal Register, February 16, 2024). Fit: private cloud, with hybrid for Part 2 records.

Legal

ABA Formal Opinion 512 asks lawyers to weigh disclosure risk, get informed consent before client information goes into a self-learning tool, and disclose AI use where a client's guidelines require it (July 29, 2024). Courts have split on privilege. U.S. v. Heppner (S.D.N.Y., February 17, 2026) denied it to consumer AI documents partly because the platform's policy allowed training and third-party disclosure (Gibson Dunn, February 20, 2026). Warner v. Gilbarco (E.D. Mich., February 10, 2026) held that ChatGPT use did not itself waive work product (AI Lex Intelligence, July 30, 2026). Fit: private cloud, or fully offline for a client that demands a closed environment.

Finance and tax

FINRA's 2026 oversight report counts chat messages as retained communications and asks firms to store prompt and output logs (December 9, 2025). Tax return data is the hard case: Treasury rules require written consent naming the specific recipient before disclosure (26 CFR §301.7216-3). IRS Publication 4557 tells firms to encrypt files and use cloud storage (Rev. 6-2024). Fit: private cloud with exportable transcripts, and hybrid for return data where the consent cannot name the model provider.

Insurance

The NAIC model bulletin (December 4, 2023), adopted in 25 jurisdictions by April 1, 2026, requires a written AI program, a model inventory, and vendor contracts with audit rights. A private instance shrinks the third-party scope an examiner reviews. Fit: private cloud.

Real estate

Your MLS license is the binding term. Bright MLS bars uploading its data into any AI tool that lacks a zero-training guarantee (rulebook update, July 7, 2026). Metro MLS data may not be used for AI training without written permission (NAR, July 15, 2026). A cloud model under no-training terms meets the Bright test. Fit: private cloud, or a vendor tool with a contractual zero-training promise.

Government contractors

DFARS 252.204-7012 requires cloud services that hold covered defense information to meet FedRAMP Moderate equivalency and requires 72-hour incident reports (acquisition.gov, last changed May 7, 2026). Under ITAR, releasing technical data to a foreign person is an export (22 CFR 120.54). DCSA says DoD "does not have an official policy for AI processing of" CUI and advises "caution and or avoidance" (DCSA CUI FAQ, May 27, 2025). Fit: fully offline inside your assessed boundary, built per client.

What does on-premise AI deployment cost to run?

Upkeep is the larger line. Across five security flaws in Ollama, a popular tool for running models locally, only 0.43% to 2.90% of exposed servers were upgraded in place (ACM IMC 2026, September 7, 2026). We do not sell a private build as a saving.

How does Omni Studio build a private AI deployment?

We lead with private cloud: Omni Studio and its AI employees installed in your own AWS or Azure account, using Amazon Bedrock or Azure OpenAI (Amazon's and Microsoft's hosted model services) under no-training or BAA terms you sign. Hybrid and fully offline setups are built per client and scoped in writing first. Pricing is scoped with you, from the Custom Build plan, or Enterprise by quote. Our Security page lists private cloud, VPC, and isolated database options set up for you during onboarding.

Each engagement covers one workflow, one metric, and 60 to 90 days, and we propose the kill criterion before work starts. In MIT NANDA's study, 95% of generative AI pilots showed no measurable value (Fortune, August 18, 2025).

In a build, AI employee Olivia drafts and files documents, and AI employee Sam watches licenses and deadlines and flags risk to a person, both on the model in your account. Anything that sends, books, signs, spends, or changes records waits for your approval, and each approval records who said yes. Our guide to AI approval gates shows how those gates are set.

What we are not

• We hold no DoD or agency authorization, and we claim none for CUI.
• We have no SOC 2, ISO 27001, or HITRUST report to share. SOC 2 and ISO 27001 are in progress, and we make no HIPAA compliance claim (Security page, October 21, 2026).
• We accept no clinical or legal liability for drafts your staff approve.
• We quote no savings or return we have not measured with you.
• Nothing we build finalizes a note, filing, or decision without a person signing off.
• We do not sell a preloaded server.

Questions about on-premise AI deployment

Is AI prohibited by HIPAA?

A HIPAA covered entity or business associate may use a cloud service to store or process ePHI when it signs a HIPAA compliant business associate agreement with the provider and otherwise complies with HIPAA. Risk analysis and risk management still apply. This is guidance about HIPAA and cloud services for ePHI, not a conclusion about all US laws or every AI use. HHS cloud computing guidance The problem is consumer chat tools with no BAA. OpenAI excludes its Free, Plus, and Team plans from HIPAA eligibility (OpenAI, October 21, 2026).

Can defense contractors use AI with CUI?

DoD has no official policy for AI processing of CUI and advises "caution and or avoidance" (DCSA, May 27, 2025). Cloud that holds covered defense information must meet FedRAMP Moderate equivalency under DFARS 252.204-7012. An offline system inside your assessed boundary avoids the cloud question.

Does ABA Formal Opinion 512 require on-premise AI?

No. Opinion 512 (July 29, 2024) asks lawyers to check a vendor's confidentiality, breach notice, and retention terms, and to get informed consent for self-learning tools. A private cloud build lets you answer those questions from your own contract.

Will Omni Studio sign a BAA?

Not today. If you need us to sign a business associate agreement before work starts, we are not a fit yet. In a private cloud build, the model runs under the BAA you sign with AWS or Microsoft, and the scope sets what our team may access.

This is general information, not legal advice.

See the private AI setups and talk to a person

---

Sources (all read or checked October 21, 2026 unless dated otherwise)

1. HHS OCR, Guidance on HIPAA and Cloud Computing (October 2016): https://www.hhs.gov/hipaa/for-professionals/special-topics/health-information-technology/cloud-computing/index.html
2. ABA Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024): https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
3. eCFR, 16 CFR Part 314, Safeguards Rule: 16 CFR Part 314
4. NAIC, Model Bulletin on Use of AI Systems by Insurers (December 4, 2023): NAIC model bulletin
5. Federal Register, Executive Order 14365 (published December 16, 2025): https://www.federalregister.gov/documents/2025-12-16/2025-23092/ensuring-a-national-policy-framework-for-artificial-intelligence
6. Colorado General Assembly, SB 26-189 (signed May 14, 2026): https://leg.colorado.gov/bills/sb26-189
7. FDIC, Outside Counsel Deskbook §1.17 (revised April 2026): https://www.fdic.gov/doing-business-fdic/outside-counsel-deskbook.pdf
8. Texas Attorney General, Outside Counsel AI Policy §2.2 (July 31, 2025): https://www.texasattorneygeneral.gov/sites/default/files/files/divisions/general-counsel/OC-Acceptable-AI-Use-Policy.pdf
9. New York State DOT, consultant instructions item 5 (October 6, 2025): https://www.dot.ny.gov/portal/pls/portal/MEXIS_APP.BC_CONSULTING_NONAE_ADMIN.VIEWFILE?p_file_id=49663
10. Fulkerson Advisors, 1,054 outside counsel guidelines searched (September 15, 2026): https://www.fulkersonadvisors.com/research/ai-in-outside-counsel-guidelines
11. Anthropic, Business Associate Agreements for Commercial Customers: https://privacy.claude.com/en/articles/8114513-business-associate-agreements-baa-for-commercial-customers
12. Microsoft Learn, Data, privacy, and security for models in Microsoft Foundry (May 19, 2026): https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy
13. Mozilla, The State of Open Source AI v1.1 (September 15, 2026): https://stateofopensource.ai/
14. OpenAI, HIPAA eligible products and functionality: https://help.openai.com/en/articles/20001069-hipaa-eligible-products-and-functionality
15. AWS, HIPAA Compliance: AWS HIPAA compliance; Bedrock data protection: Amazon Bedrock data protection
16. Claude API pricing (October 21, 2026): https://claude.com/pricing
17. Silicon Data, GPU rental price indices (October 21, 2026): https://www.silicondata.com/products/silicon-index
18. kingy.ai, M5 Mac Studio for local AI (October 8, 2026, third party): https://kingy.ai/blog/m5-mac-studio-local-ai-m5-max-vs-m5-ultra/ ; Viperatech, Dell XE9680 with 8 H100 GPUs (October 21, 2026): https://viperatech.com/product/dell-poweredge-xe9680-rack-server-210-bfxh-bundle-with-8-x-h100-gpus
19. Artificial Analysis, GDPval-AA v2.1 leaderboard: https://artificialanalysis.ai/evaluations/gdpval-aa
20. Federal Register, 42 CFR Part 2 final rule (February 16, 2024): https://www.federalregister.gov/documents/2024-02-16/2024-02544/confidentiality-of-substance-use-disorder-sud-patient-records
21. Gibson Dunn on U.S. v. Heppner (February 20, 2026): https://www.gibsondunn.com/ai-privilege-waivers-sdny-rules-against-privilege-protection-for-consumer-ai-outputs/
22. AI Lex Intelligence case tracker, Warner v. Gilbarco (July 30, 2026; single source, order not retrieved): https://ailexintelligence.com/cases/
23. FINRA, 2026 Annual Regulatory Oversight Report (December 9, 2025): https://www.finra.org/sites/default/files/2025-12/2026-annual-regulatory-oversight-report.pdf
24. eCFR, 26 CFR §301.7216-3: https://www.ecfr.gov/current/title-26/section-301.7216-3
25. IRS Publication 4557 (Rev. 6-2024): https://www.irs.gov/pub/irs-pdf/p4557.pdf
26. Bright MLS rulebook article (July 7, 2026): https://www.brightmls.com/article/fewer-rules-more-protection-the-new-bright-rulebook
27. NAR, Make It a Policy to Protect MLS Data From AI Misuse (July 15, 2026): https://www.nar.realtor/news/aexperience/make-it-a-policy-to-protect-mls-data-from-ai-misuse
28. eCFR, 48 CFR 252.204-7012: https://www.ecfr.gov/current/title-48/chapter-2/subchapter-D/part-252/subpart-252.2/section-252.204-7012
29. eCFR, 22 CFR 120.54: https://www.ecfr.gov/current/title-22/chapter-I/subchapter-M/part-120/section-120.54
30. DCSA, CUI FAQ v3 (May 2025): https://www.dcsa.mil/Portals/128/Documents/CTP/CUI/DCSA%20CUI%20Frequently%20Asked%20Questions%20(May%202025).pdf
31. Ollama in the Wild, arXiv / ACM IMC 2026 (September 7, 2026): https://arxiv.org/abs/2609.07115
32. Fortune, MIT NANDA report (August 18, 2025): https://fortune.com/2025-08-18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
33. FluidLabs, citing S&P Global (April 14, 2026): https://fluidlabs.com/resources/why-42-percent-enterprise-ai-abandoned-2025
34. omni.studio/pages/security and /pages/pricing, read October 21, 2026