Which Business AI Plan Keeps Your Data Private? ChatGPT, Copilot, Claude, and Gemini

Every major business plan now promises not to train on your data. The questions that separate them are what the AI can find, who can read it, and how long it stays.

Most of the companies I work with did not choose their AI the way the comparison articles suggest. They chose it the way they chose their email. Microsoft 365 shops lean toward Copilot. Google Workspace shops lean toward Gemini. That instinct is more sensible than it sounds, and it points at the real privacy decision.

On a business plan, ChatGPT, Microsoft Copilot, Claude, and Gemini all state that they do not train their models on your data by default. So the choice comes down to three other questions: what the assistant can find across your files, who at the vendor can read what your people type, and who controls how long it is kept. Copilot and Gemini answer those inside the Microsoft or Google environment you already run. ChatGPT and Claude are a separate place your data goes, under their own terms.

One note before the details. These policies are evolving, and several changed in the months before this was written. Everything below reflects each vendor's published terms as of October 2026. The purpose here is to lay out the considerations your company needs to settle, so you can check them against whatever the terms say on the day you buy.

Which business AI plans train on your data?

None of the four business plans trains on your data by default, according to each vendor's own privacy documentation.¹³⁶⁹ OpenAI states it for ChatGPT Business and Enterprise. Microsoft states that prompts, responses, and the data Copilot reaches through Microsoft 365 are not used to train foundation models. Anthropic excludes its commercial products unless a customer opts in. Google states that Workspace does not use customer data for training without permission.

The consumer versions are a different matter. Free and paid individual plans can use conversations for training unless the user turns that off, which is why a personal account on a work task is where much of the exposure begins. I walked through that difference, and what "no training" leaves uncovered, in does AI train on your data.

Since training no longer separates the business plans, the useful comparison lives in three places: indexing, human review, and retention.

What is indexing, and how is it different from training?

Indexing is building a searchable map of your content ahead of time so an AI assistant can find the right email or document the moment someone asks. Training changes the model itself. Indexing leaves the model alone and changes what it can find about your business.

Think of a library catalog. It leaves every book as written and makes each one findable, including the ones that were misshelved.

Microsoft documents this in the most detail of the four. For companies with paid Copilot licenses, it builds a semantic index from SharePoint, OneDrive, and each user's mailbox, stores it inside your Microsoft 365 tenant, and does not use it to train models.⁷ Microsoft also says the index is switched on automatically and "can't be disabled." Copilot shows each person only what that person already has permission to open.

Google describes Gemini as retrieving content across Gmail, Drive, Calendar, and Chat, limited to what the asking user can access.⁹ Google has not published whether that runs on its existing Workspace search index or something built for Gemini. For a buyer, the practical answer is the same either way. In both suites, the only thing standing between a sensitive file and an employee's question is that employee's permissions.

How can indexing share data nobody meant to share?

Indexing can expose information without anyone deciding to share it, because it makes old permission mistakes findable. Two paths matter most for a smaller company: your file shares and your employees' phones.

The file share path. Picture a salary spreadsheet sitting on a SharePoint site that was opened to the whole company years ago so a team could share a template. For years nobody looked for it, so nobody found it. Microsoft's index can surface it to anyone who asks Copilot the right question, because on paper they already have access. The AI did exactly what it was built to do. It found what the permissions allowed.

The phone path. In September, Apple released Siri AI with iOS 27. Apple says it works from "the Spotlight index and App Toolbox, which work entirely on device."¹¹ By default that index covers apps such as Mail, Messages, and Notes.¹² For heavier requests, Apple routes the work to its Private Cloud Compute servers, where it says personal data "is not stored nor made accessible to Apple or anyone else."¹¹ Apple also states it does not train its foundation models on users' private personal data.¹⁵ That is a careful design.

Now follow one ordinary chain. An employee reads work email in Apple Mail on a personal iPhone, so that email is in the phone's index. They have switched on Siri's ChatGPT extension, which ships turned off, and they ask Siri for help with a reply. Siri offers to hand the request to ChatGPT. Apple asks before sending anything, and the employee says yes. If they are signed out of ChatGPT, OpenAI is not permitted to store the request or train on it. If they are signed in to a personal ChatGPT account, Apple's own privacy page says OpenAI "may log your request, attachments, and session history" and use that data to "train or improve their models."¹³

Each step was reasonable. The sum is company email under consumer terms. That is the pattern shadow AI follows, and a phone makes it faster.

Two things in this chain I could not confirm from Apple's documentation: whether email in a company-managed mail app is excluded from the Siri index, and how the additional AI providers reported for iOS 27 will handle data. Test both on your own devices before you write policy around them. Apple does let companies that manage devices restrict Siri AI and the ChatGPT extension under iOS 27.¹⁴

How do the four business plans compare on privacy?

The business plans differ most on human review and on who controls retention, and the two suite-based assistants keep both closer to home.

Swipe sideways to see all four plans.

Privacy question	ChatGPT Business	Claude Team	Microsoft 365 Copilot Business	Gemini in Google Workspace
List price, as listed on October 22, 2026	$20 per user per month, annual	$20 per seat per month, annual	$21 per user per month, annual (promoted at $18), added to a Microsoft 365 plan	Included in Workspace plans; Business Standard $14 per user per month, annual
Trains on your data by default	No	No	No	No
Will the vendor sign a HIPAA agreement for this plan?	No. Enterprise, Edu, or the API only	No. Enterprise or the API only	Yes, by default for customers subject to HIPAA; Copilot's web searches excluded	Yes, accepted in the Admin console (confirm your edition); Gemini Notebook and Gemini in Chrome excluded
Vendor staff may review content for abuse	Yes, authorized staff and contractors	Flagged chats retained up to 2 years	No; Copilot has opted out of abuse monitoring that includes human review	Not without permission
Who sets retention	Workspace admins, per OpenAI's Business FAQ	Custom retention on Enterprise only	Your admins, through Microsoft Purview	Your admins, from 90 days to indefinite
Zero data retention for the chat product	No	No	Not applicable; stays in your tenant	Not applicable; stays in your tenant
What it can search	Connected apps on request, plus memory of past chats	Connected apps on request	A semantic index of your tenant	Gmail, Drive, Calendar, and Chat, by user permission

Sources: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 16, 17, 18, 19, 20

Two cautions about this table. OpenAI's own privacy page lists retention control as an Enterprise feature in its summary and as a Business feature in its FAQ, so ask your rep which applies to your workspace.¹ And Claude's Enterprise plan, which carries custom retention, now starts at a 20-seat minimum plus usage, which puts it within reach of more midsize companies than its name suggests.⁵

Should you just use the AI that comes with your business suite?

For many small and midsize companies, matching the AI to the business suite you already run is a sound default, provided your file permissions are in order. Copilot and Gemini keep retention, review, and audit inside the environment your IT team already governs. That is a real advantage.

The condition is the permissions. Among the Copilot customers I work with, a common practice is to set "Allow this site to appear in search results" to off as the standard, then turn search on site by site in a controlled way as each site's access is reviewed. Microsoft notes the trade-off: that setting removes a site from ordinary Microsoft Search as well as the Copilot index, so people lose search on those sites until they are opened back up.⁷ It is a cost worth paying for the control it buys.

ChatGPT and Claude are strong tools, and their business plans make the same training promise. Companies on those plans need to know one thing clearly: the business chat products offer no zero data retention option.¹⁴ Anthropic states that its Team and Enterprise chat interfaces are not eligible. OpenAI offers zero retention only on qualifying API endpoints. So the use cases should be chosen with that in mind. General drafting, research, and analysis fit well. Confidential or regulated work belongs on an arrangement where you control retention, such as the API with zero retention, a model running in your own cloud tenant, or hardware you own.

What should a small or midsize company do first?

Settle permissions and personal accounts before you settle the vendor. In practice that means four steps.

First, review who can see what in your file shares, and treat any site open to the whole company as a site the AI can surface. Second, approve one business plan and make it easy to use, so nobody has a reason to reach for a personal account. Third, set a plain rule for phones: no personal AI accounts on devices that carry company email. Fourth, if you manage company phones, decide your Siri AI and extension settings on purpose instead of by default.

This is the governed rollout of the right AI plan, the starting point I described in what AI consulting costs a small business. It is usually the least expensive step in an AI program, and the one that protects every step after it.

Proxigee Services is an AI advisory firm in the Pittsburgh area that helps small and midsize companies choose, govern, and get value from AI, including deciding which capabilities to keep under their own control.

The AI you choose matters less than what it can find. Clean up what it can find, and the choice of plan gets much easier.