Private AI for Massachusetts Businesses: What It Means

TL;DR: Private AI means your prompts and files are not used to train someone else's model, are kept only as long as you agree, are seen only by people you approve, and run where you can say where. A business can get it three ways: a business AI account, a hosted private tenant, or self-hosting.

Private AI is what most owners mean when they ask whether AI is safe to use on client data. The worry is fair. The good news is that "private" is not a mood or a marketing word. It is four plain questions, and you can check every one of them this week.

This note defines private AI for a business, walks through where your data actually goes, compares the three ways to get it, and ends with what Massachusetts rules already expect of you. It is written for a 10 to 50 person firm, not a data center.

What does private AI actually mean for a business?

Strip away the labels and private AI is a set of promises about your data that you can read in writing. If a tool cannot answer the four questions below, it is not private, whatever the sales page says.

The four questions that define private AI

1. Training. Is anything you type or upload used to train or improve a model that other customers use? The answer you want is no, in the contract, with no opt-in buried in a setting. This is the model training opt-out, and it should be the default, not a setting you have to find.
2. Retention. How long are prompts, files and answers kept, and can you change that? Data retention is the question that is easiest to skip. Some business plans offer zero data retention, which means the provider processes the request and does not store it afterward.
3. Access. Who can see your data: the provider's staff for abuse review, subcontractors, other people in your own company? A private AI chatbot that lets every employee search every folder is private from the outside and wide open on the inside.
4. Location. Where does the model run and where is the data stored? On the provider's shared service, on servers dedicated to you, or on hardware in your own office.

Those four answers are the whole definition of private AI. Everything else, including which model is "smartest" this month, is a separate decision.

Where does my data go when I use AI?

When someone on your team types into an AI tool, the text leaves your building, travels to the provider, gets processed by a model, and the answer comes back. Along the way it can be logged, stored, reviewed or kept for training, depending on the account and the settings.

The federal government's own risk guide for generative AI lists data privacy as one of its core risks: "leakage and unauthorized use, disclosure, or de-anonymization of biometric, health, location, or other personally identifiable information or sensitive data." It also warns that "models may leak, generate, or correctly infer sensitive information about individuals." (NIST AI 600-1, July 2024)

The Federal Trade Commission has been just as direct about why providers want your data. Its technology staff wrote that model-as-a-service companies "have a continuous appetite for data to develop new or customer specific models," and that breaking a promise not to use customer data "to train or update their models" can put a company on the wrong side of the law. Their summary: "There is no AI exemption from the laws on the books." (FTC Office of Technology, January 9, 2024)

Here is the example that makes retention real. In May 2025, a federal court in the New York Times copyright case ordered OpenAI to keep ChatGPT logs it would normally delete. OpenAI usually erased deleted chats after 30 days, but under the order it kept them. Companies using the API under a zero data retention agreement were not affected, and the Enterprise and Edu editions of ChatGPT were exempt too. (SiliconANGLE, June 6, 2025; National Law Review, July 25, 2025) The requirement to keep new logs ended on September 26, 2025, though logs already saved stayed available. (Engadget, October 11, 2025)

Put those two facts side by side. Same provider, same months, and the account type decided whether a "deleted" chat was really gone. That is why private AI starts with the account, not the brand.

Our note on what AI security means for a small business covers the other two sides of the problem: AI you connect to your systems, and AI aimed at you.

What are the three ways to get private AI?

For a small or mid-sized firm there are three realistic routes to private AI. They are not ranked. Each fits a different kind of work.

1. A business AI account

The fastest route is the business or enterprise tier of a tool your team already uses. A business AI account is the lightest form of private AI. It usually comes with written terms that the provider will not train on your data, admin controls, and sometimes retention settings.

The large providers say this plainly in their own documentation. Microsoft states that for Copilot in Microsoft 365, "prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs." (Microsoft Learn, updated November 2026) Google states for Workspace that "your content is not human reviewed or otherwise used for Generative AI model training outside your domain without permission." (Google Workspace Privacy Hub, updated October 2026)

Good fit: drafting, summarizing and searching data you are allowed to share with that provider. Watch for: the provider writes the terms, and they can change. The FTC has warned that it "may be unfair or deceptive" for a company to start using data for AI training and announce it only through a "surreptitious, retroactive amendment" to its terms. (FTC Office of Technology, February 13, 2024) Reread the terms once a year. My longer note on whether staff can put customer data into AI tools (opens in a new tab) goes tier by tier, and which AI business plan to pay for (opens in a new tab) covers the plans themselves.

2. A hosted private AI tenant

The middle route is hosted private AI, running in an environment set up for your business alone. Someone else operates the servers, but your data, your rules and your logs sit in a space no other customer shares. The model can be one of the big commercial ones reached through a no-training, low-retention agreement, or an open model running on the same dedicated machines.

Good fit: firms that want AI working on client files, email and records, with one place that holds the rules and a record of what the AI did. Watch for: "private" has to cover access inside the tenant too. Our one-week least privilege playbook for AI agents is the checklist for that part, and our note on zero trust architecture explains the principle behind it.

3. Self hosted AI on your own hardware

The third route is private AI you run yourself, on a server or workstation you control. This became practical because many capable models now publish their weights. The National Telecommunications and Information Administration noted that these open models let users work with a model "without sharing data with third parties," which it described as "increasing confidentiality and data protection." (NTIA, July 30, 2024)

Good fit: very sensitive data, and firms with someone who can patch, monitor and back up the box. Watch for: you now own the security of that machine. The joint guidance from the NSA, CISA, the FBI and partner agencies on deploying AI systems exists because a self-run model is another system to harden: it covers protecting, detecting and responding to "malicious activity against AI systems and related data and services." (CISA, April 15, 2024) I wrote about why we run our own Ollama instead of calling an API (opens in a new tab) for some of our own work.

One more thing about private AI models in general: the model and the privacy are separate choices. A commercial model can be used privately under the right agreement, and an open model can be run carelessly on a laptop with no backups and no logins. Judge the setup, not the model's name.

Is self-hosting AI cheaper?

Self-hosting private AI is sometimes cheaper, and less often than it looks. I am not going to print a price range here, because I could not find a neutral, reputable source for one that would hold for a 10 to 50 person firm. What I can give you is what drives the number:

• Hardware. Larger models need graphics cards with a lot of memory. A model small enough for a modest machine is cheaper, and also less capable.
• Usage. A rented account is priced per seat or per use. A server you own costs the same whether five people use it or fifty. Heavy, steady use favors owning; light or spiky use favors renting.
• Operations. Someone has to patch, monitor, back up and update the model. That labor is the cost people leave out of the spreadsheet.
• Replacement. Models improve fast. Plan for the hardware to be swapped or the model to be changed within a few years.

My partner Amelia built a full total cost model for self-hosting (opens in a new tab) if you want to run your own numbers, and her piece on self-hosted versus cloud tradeoffs (opens in a new tab) covers what the spreadsheet leaves out. Kief Studio prices custom work around the use case, the size of the business, what it needs and the stack it already runs, so the honest answer for your firm comes from a short look at your setup.

What do Massachusetts rules expect from private AI?

Massachusetts did not wait for an AI law to cover this. The rules you already follow for personal information apply to private AI the same way they apply to any other system.

The data security law. Under chapter 93H, "personal information" means a resident's name combined with a Social Security number, a driver's license or state ID number, or a financial account or card number. (M.G.L. c. 93H, section 1) If that kind of data goes into an AI tool, the tool is part of your security program, private AI included.

The vendor rule. 201 CMR 17.00 requires a business that holds that information to oversee service providers by "taking reasonable steps to select and retain third-party service providers that are capable of maintaining appropriate security measures" and by "requiring such third-party service providers by contract" to keep those measures in place. (201 CMR 17.03, via Cornell LII) A private AI provider handling client data is a service provider like any other. Ask for the contract language. Amelia's guide to Massachusetts WISP requirements under 201 CMR 17.00 (opens in a new tab) walks through the written program itself.

The Attorney General's view. In April 2024, Attorney General Andrea Campbell issued an advisory saying the state's consumer protection, anti-discrimination and data security laws apply to AI. WBUR reported that the advisory says "data security laws still apply to emerging technologies, including AI systems," and that developers, suppliers and users "must take steps to safeguard personal data used by AI systems." (WBUR, April 17, 2024) Note the word users. That is you.

The privacy bill. The Massachusetts Data Privacy Act is still not law. As of November 9, 2026, the official bill history shows the Senate and House passed different versions, both chambers appointed conference committee negotiators in June 2026, and no compromise or signature appears after the June 17, 2026 entry. (Massachusetts Legislature, bill S.2619) Our note on the Massachusetts Data Privacy Act for small businesses explains what the two versions would ask of you. The Massachusetts compliance sources on MyRHC (opens in a new tab) keep the primary texts in one place.

How do you choose private AI for your business in one afternoon?

You do not need a committee. You need an afternoon and a short list.

1. List the work. Write down the five tasks your team most wants AI for. Next to each, write the most sensitive thing it would touch: client names, health details, account numbers, contracts.
2. Sort by sensitivity. Anything with 93H data or confidential client material goes in one column. Public or internal-only material goes in the other.
3. Ask the four questions. For every tool in use today, write the answers on training, retention, access and location. If you cannot find an answer in writing, treat it as a no.
4. Match the route. The low-sensitivity column can usually live on a business AI account. The high-sensitivity column needs a hosted private tenant or self hosted AI, with access rules and logs.
5. Close the free accounts. Personal and free accounts are the ones you cannot see or control. Give people a sanctioned tool so they stop needing the unsanctioned one. Amelia's case for governing shadow AI by enabling it (opens in a new tab) is the right frame.
6. Write it down. One page of private AI rules: which tool is approved for which data, who owns the account, and when you will check the terms again. Put the date on your calendar for six months out.

Picture a 20-person accounting firm in Waltham. The partners want AI to draft engagement letters and summarize client documents. Letters are low risk and fit a business account. Tax documents carry Social Security numbers, so they belong in a private setup with per-client access and a log. Two columns, two kinds of private AI, one page. If the bigger goal is answering staff questions from your own files, our note on an AI knowledge base for staff questions covers how to require sources on every answer.

This article is general information, not legal advice. Kief Studio is not a law firm, and reading it does not create an attorney-client relationship. Laws and regulations change, so check the current text and talk with a licensed Massachusetts attorney about your situation.