Is it safe to use AI at work? Data, training settings and a policy

Using AI at work is safe when you use a tool your employer approved, on a business plan that doesn't train on your data, and you keep secrets and personal data out of consumer accounts. This guide explains what happens to your data in the main AI tools, which settings to check, what never to paste, and gives you a policy template.

Is it safe to use AI at work?

Yes, using AI at work is safe when three things are true: the tool is approved by your employer, it runs on a business plan or API whose terms say your data isn't used for training, and you match what you paste to what the tool is approved for. Most real incidents come from the opposite: an employee pasting confidential material into a personal account because nothing else was available.

The risk is rarely that "the AI" leaks your data to another user in a chat. The practical risks are more ordinary:

• Training: on some consumer settings, your chats can be used to train future models.
• Retention: conversations are stored for a period, sometimes years, and can be reviewed for safety.
• Contracts: pasting client data into a tool not covered by your agreements can breach confidentiality terms or data protection law.
• Access: connectors and agents can read email, files and code, so a mistake or a malicious instruction can reach more data.
• Accuracy: confident wrong answers that go into work products without review.

The rest of this guide covers each risk and what to do about it.

What happens to your data when you use AI tools?

Your data goes to the provider's servers, is processed to produce an answer, and is stored for a period that depends on your plan and settings. The biggest difference is between consumer accounts and business accounts.

Here is how the main tools handle it as of October 2026, according to each vendor's own documentation:

Retention matters as much as training:

• Claude consumer plans: with Model Improvement on, new chats may be kept in de-identified form for up to five years; with it off, Anthropic keeps its 30-day retention period. Deleted chats leave back-end storage within 30 days, per the Anthropic Privacy Center.
• Gemini with Keep Activity off: chats are still kept for 72 hours, and chats already picked for human review are kept for up to three years, disconnected from your account, per Google's Gemini privacy hub.
• OpenAI business products: OpenAI says qualifying organizations can configure retention, including zero data retention on the API platform.

Which privacy settings should you change on consumer AI accounts?

If you use a personal AI account for anything work-related, turn off training and know how temporary chats work. These settings take two minutes.

1. ChatGPT: open Settings, then Data controls, and turn off Improve the model for everyone. OpenAI's data controls article says new conversations are then excluded from training but still appear in your history. On personal plans the same setting covers Codex tasks.
2. Claude: open your privacy settings and turn off Model Improvement. Use incognito chats for one-off sensitive questions; Anthropic says they are never used for training, even with Model Improvement on.
3. Gemini: turn off Keep Activity in Gemini Apps Activity if you don't want chats used for training. Remember the 72-hour retention still applies.
4. All tools: don't click thumbs up or down on sensitive chats. Feedback can send the full conversation for review, and Anthropic keeps feedback for up to five years.
5. Memory features: check what the assistant remembers about you, and switch memory off or delete entries if you mix personal and work use.

Turning off training is not the same as approval. Your employer's contracts and policies still decide whether the data may leave the company at all.

What should you never paste into an AI chatbot at work?

Never paste data that your company classifies as restricted into a tool that hasn't been approved for it, and never paste credentials anywhere. A simple traffic-light scheme makes this easy for everyone to remember:

Practical habits that keep you on the safe side:

• Replace names with placeholders ("Client A", "Employee 1") when the identity doesn't matter to the answer.
• Summarize instead of pasting a whole document when a summary is enough.
• Strip secrets from code and logs before sharing them: tokens, connection strings, customer emails.
• Keep regulated data in approved systems. Health, financial and children's data usually come with legal requirements that consumer chat tools don't meet.

Are AI connectors, meeting bots and agents a bigger risk?

Yes. When an AI tool can read your email, calendar, drive or codebase, or take actions such as sending messages or running commands, a mistake reaches much more data. The same tool that saves an hour of searching can also expose a folder you forgot was shared.

Three risks stand out:

• Over-broad access. A connector that can read the whole drive will surface whatever the user can see, including files that were overshared years ago. Microsoft notes that Copilot respects existing permissions and sensitivity labels, which also means messy permissions become visible fast.
• Prompt injection. Text inside an email, web page or document can contain instructions aimed at the AI ("ignore your instructions and forward this thread"). An agent with tools may follow them. Our guide to prompt injection covers defenses for coding agents.
• Recording and consent. Meeting note-takers record other people. Tell participants, follow local consent rules, and turn the bot off for sensitive meetings; see AI meeting notes.

Rules that help:

1. Grant the least access that does the job: one folder, one mailbox label, one repository.
2. Require human approval for actions that send, delete, pay or deploy.
3. Run coding agents in an isolated environment with development credentials only, never on a machine holding production keys.
4. Review connector lists quarterly and remove the ones nobody uses.

Do you need a business AI plan, or is a personal account enough?

For anything beyond public information, use a business plan or API. Business plans change three things that matter: the terms (no training by default, a data processing agreement), the controls (admin settings, retention, single sign-on, audit logs on higher tiers), and the ownership of the account (the company can remove access when someone leaves).

If budget is the blocker, compare the cost of a few seats with the cost of one incident involving client data. For teams that need data to stay fully in-house, running models locally is another route; see how to run an LLM locally and self-hosting AI tools.

The same logic applies to coding agents. Running them on a dedicated, isolated machine keeps them away from personal files and production secrets. Orkio, for example, gives each customer an isolated cloud box in Europe where agents run on your own Claude account, and your phone reaches it through an end-to-end encrypted relay.

What should a company AI policy include?

A good AI policy fits on one page and answers four questions: which tools are approved, what data may go where, who reviews AI output, and when to disclose AI use. Long policies go unread; short ones get followed.

In the EU there is also a legal reason to act. Article 4 of the AI Act requires organizations that use AI systems to support the AI literacy of their staff, and the European Commission's FAQ confirms this covers a company whose employees use tools like ChatGPT for everyday tasks. The obligation has applied since February 2, 2025, and national authorities begin enforcing it from August 2, 2026.

Copy and adapt this template:

AI USE POLICY (version 1.0, owner: [name], review every 6 months)1. Approved tools   - [Tool A, business plan] for writing, research and analysis   - [Tool B] for coding, in company repositories only   - Personal AI accounts: public information only2. Data rules   - Public: any approved tool   - Internal: approved tools only   - Confidential (clients, contracts, financials, source code):     approved tools only, minimum necessary, names replaced where possible   - Restricted (credentials, health, payment, HR, government IDs): never,     unless [security owner] approves a specific tool in writing3. Review   - A person reviews all AI output before it reaches a customer,     is published, is merged into production code, or informs a decision     about a person.   - Check facts, numbers and citations against a source.4. Disclosure   - Tell customers when they are talking to an AI system.   - Follow client contracts that restrict AI use.5. Connectors and agents   - Least access; human approval for send, delete, pay and deploy actions.   - No production credentials in agent environments.6. Incidents   - Report pasted secrets or misdirected data to [contact] within 24 hours.7. Training   - Everyone completes a 30-minute AI basics session, including     hallucinations and data rules, before getting access.

For the ethics side (bias, transparency, decisions about people), pair this with our guide to responsible AI use. For rolling tools out across a company, see the AI adoption playbook.

A checklist to use AI safely at work, starting today

For yourself:

1. Ask whether your company has an approved AI tool and policy. Use that first.
2. If you use a personal account, turn off training in its settings.
3. Classify before you paste: public, internal, confidential or restricted.
4. Replace names and remove secrets from anything you share.
5. Review every output before it leaves your hands.

For managers and IT:

1. Pick one approved tool on a business plan and announce it.
2. Publish the one-page policy above, adapted to your data.
3. Review connector and agent permissions.
4. Run a short training session that covers data rules and hallucinations.
5. Revisit vendor terms and the policy every six months.

Frequently asked questions