Is it safe to give AI access to my business data?

What the major AI providers say about your business data

The fear I hear most is "will my data end up training the AI, and then show up in someone else's answer?" For business plans from the big providers, their published position is no, by default.

As of late 2026, here's what they say:

• OpenAI says on its business data page that it doesn't train its models on your organization's data by default, covering its business ChatGPT plans and its API.
• Anthropic says in its privacy center that by default it won't use inputs or outputs from its commercial products, such as Claude for Work and the API, to train models, unless you opt in or send feedback.
• Google says in its Workspace generative AI privacy hub that Workspace doesn't use customer data to train models without permission, and that your content isn't used for other customers.

Personal and free plans are a different story. They often have a training setting you need to switch off yourself; OpenAI, for example, explains how personal users opt out. If your team is pasting work into personal accounts, that's the bigger risk.

Two cautions. First, "not used for training" doesn't mean "not stored." Providers keep data for some period for things like abuse monitoring, and retention rules vary by plan. Second, these policies change. Read the current terms for the exact plan you're paying for, not a summary like this one.

Give AI the least access that gets the job done

Security people call this "least privilege." In plain English: an AI tool should be able to reach exactly what its job needs, and nothing else.

Some examples of what that looks like:

• Calendar: read-only to check availability, or permission to add events but not delete them.
• Files: one shared folder for the task, not your whole drive.
• Email: create drafts for a person to send, at least until you trust the output.
• CRM: a user role that can add notes and update stages, but can't export the full contact list or delete records.
• Accounting: read-only for reporting, or draft-only for invoices. Never payment rights.

Start narrow and widen access only when there's a clear reason. It's much easier to grant a permission later than to explain why an agent deleted something it never needed to touch.

Separate logins, not shared passwords

Small businesses share passwords constantly. One login for the CRM, written on a sticky note, used by four people and now an AI tool. That's the setup I'd fix first, AI or not.

Every person and every AI agent should have its own login. That gives you three things:

1. A record. You can see which user did what, including the agent.
2. A clean off switch. Disable one user, and only that access stops. Nobody else is locked out.
3. Two-factor login. Separate accounts can each have their own two-factor authentication.

Use a password manager for the logins you do need to share. Business password managers let you share access to a login without anyone seeing or pasting the actual password. And never paste a password into an AI chat window. If an agent needs to sign in somewhere, it should use a stored credential or a proper connection, not a password typed into a conversation.

What access do AI agents actually need?How I handle access, what agents never touch, and the five-minute off switch.
Read the security page

What you should never hand to AI

Some things don't belong anywhere near an AI tool, no matter how good the plan's privacy terms are:

• Bank logins and payment rights. No agent should be able to move money, pay a bill or change bank details. If it's tricked by a fake vendor email or misreads a number, the mistake costs real dollars.
• Payment card numbers. Your payment processor handles those. Keep them there.
• Government ID numbers and scans. Social Security numbers, driver's licenses, passports, tax IDs for individuals.
• Master admin accounts. Agents get their own limited user, never the owner's account.
• Sensitive personnel files. Medical notes, disciplinary records, background checks.

If a task seems to require one of these, the task needs redesigning. Usually the agent can prepare everything and a person does the final step that touches the sensitive piece.

Regulated data needs extra care

If you handle health, legal or financial information, the general rules above are the floor, not the ceiling.

Health information. If you're a covered entity under HIPAA, federal guidance says you need a business associate agreement with a cloud provider that handles protected health information. That generally applies to AI vendors too. Some business AI plans offer one; many don't. No agreement, no patient data.

Legal information. Lawyers have confidentiality duties, and state bar guidance on AI keeps developing. Client information should only go into tools the firm has reviewed and approved.

Financial information. Accountants, insurance agencies, lenders and advisors often have their own privacy and data security rules. Check with whoever handles your compliance before connecting client data.

This is general guidance, not legal advice. If you're in a regulated field, talk to your attorney or compliance advisor before setting anything up.

Keep a record and an off switch

You should always be able to answer two questions: what did the AI do, and how do I stop it?

Good setups keep a log of every action an agent takes, use separate logins so actions are traceable, and can be shut off in minutes by disabling a user or revoking a connection. Review access every few months and remove anything no longer in use. More on how I handle this is on my security page.

A practical checklist before you connect any AI tool to business data:

• [ ] We're on a business plan, and I've read its data and training terms.
• [ ] The AI has its own login, not someone's personal account.
• [ ] Its access is limited to what the task needs, read-only where possible.
• [ ] Two-factor authentication is on.
• [ ] Shared logins live in a password manager, not in chats or sticky notes.
• [ ] It has no bank access, payment rights or government ID data.
• [ ] Regulated data has the right agreements in place, or stays out.
• [ ] Actions are logged and I know how to switch it off.
• [ ] Someone reviews access every few months.

What this looks like in a real small business

Example: picture a nine-person property management company. Rental applications include Social Security numbers, pay stubs and ID scans. The owner wants AI to handle maintenance requests and tenant follow-ups.

First, the team moves from personal AI accounts to a business plan and turns on two-factor login for everyone. The agent gets its own user in the property management software, with access to maintenance tickets and tenant contact details only. Applications, screening reports and bank details stay out of its reach.

The agent reads incoming maintenance requests, sorts them by urgency, drafts replies to tenants and schedules vendors for routine work. Anything involving money (a deposit question, a rent dispute) goes to a person.

Shared logins move into a password manager. Every agent action is logged, and the office manager knows that disabling one user shuts it off. Once a quarter, someone reviews what the agent can access and trims anything it no longer needs.

The honest part

No setup is zero risk. AI providers can have outages or breaches, same as any software company. Agents can be fooled by cleverly written emails or documents. The goal isn't perfection; it's limiting what can go wrong and how fast you can stop it.

In most small businesses, the more likely problem isn't the AI provider. It's an employee pasting a customer list into a free personal account because nobody told them not to. A short written AI policy does more for safety than any setting.

If your business runs on one shared admin password and no two-factor login, fix that before adding AI. And if you're hiring someone to set this up, ask how they handle access. Here are questions to ask an AI consultant before you hand over any keys.