AI for Business

How to Protect Company Data When Using AI

How to Protect Company Data When Using AI

AI assistants are now part of everyday work: drafting emails, summarising reports, cleaning up spreadsheets. That is good news for productivity. It also means that, every day, people in your company may be pasting customer details, contracts and internal figures into tools you never approved.

You don’t need to ban AI to protect your business. You need clear habits. This guide, based on the free course How to Be Smarter Than AI, explains what the real risks are, what should never go into a chat, and how to keep almost all of AI’s usefulness while sharing far less.

A chat window is not a safe

It is easy to treat an AI chat like a private notebook. It isn’t. Depending on the provider and your settings, conversations may be stored, reviewed by people for quality or safety, or used to improve future models. Any online service can also suffer a security flaw or a leak. And when staff use personal accounts on tools the company hasn’t vetted (often called “Shadow AI”), you lose visibility over where your data goes.

None of this means AI tools are reckless. It means you should assume that anything you type could, one day, be seen by someone other than you, and decide what to share on that basis.

What should never go into an AI chat

The course gives a clear list. As a business, treat these as off-limits unless your company has an approved setup designed for that kind of data:

  • National ID numbers, passport details and other official identifiers
  • Bank account details, card numbers and payment data
  • Health information about employees or customers
  • Scanned official documents (contracts, IDs, licences) with personal details visible
  • Company source code, trade secrets and unreleased plans
  • Passwords, API keys and access tokens

Also watch the small details. A first name here, a job title there and a city in another chat can add up to a profile, especially in tools that keep a memory between conversations.

Key tip: Use the course’s simple rule: if you wouldn’t say it in public, don’t say it to AI.

Strip the data, keep the meaning

The good news is that most tasks don’t need real data to work. The AI needs the structure of the problem, not the identity of your client. The course calls this technique “strip the data, keep the meaning.”

Use placeholders

Replace real details with labels: [client name], [contract value], [employee A], [city]. Ask the AI to keep the placeholders in its answer. Then re-insert the real details yourself, locally, in your own document.

Summarise instead of pasting

Instead of pasting an entire contract, describe the clause you need help with in your own words. Instead of uploading a full customer list, share three anonymised sample rows to show the format.

Clean images before uploading

Screenshots often reveal more than you think: email addresses, tabs, account numbers, names in a sidebar. Crop or blur anything that isn’t needed for the task.

Settings help, but they are not full privacy

Many AI tools let you switch off the use of your chats for training. It is worth doing. But be clear about what it does and doesn’t do: turning off training does not stop your data from being stored, reviewed under the provider’s policies, or exposed in a breach. Menu locations and options also change often, so check your tool’s current privacy settings (the course walks through this in its privacy station).

For truly confidential files, consider a model that runs entirely on your own computer. The course explains that real local privacy only holds when the whole data path stays on your device, which means turning off web search, connectors, cloud models and telemetry. Local isn’t free either: it costs hardware, electricity and setup time.

A practical data-protection checklist for your team

  1. Decide which AI tools are approved for company work, and on which accounts.
  2. Write a one-page list of data that must never be pasted into AI tools.
  3. Train staff on the placeholder technique: [client name], [amount], [employee].
  4. Turn off training on chats where the tool allows it, and recheck settings periodically.
  5. Keep memory for general preferences (role, writing style), never secrets.
  6. Crop or blur screenshots before uploading them.
  7. Give AI tools connected to email or drives the minimum permissions they need.
  8. Keep API keys in a secure secret manager, never in chats, code or shared documents.
  9. For highly sensitive work, evaluate a fully local setup or an enterprise agreement.

Common mistakes

  • “It’s just a quick question.” Small pastes of real data add up. The habit matters more than any single message.
  • Assuming “training off” means private. It reduces one risk; storage, human review and leaks are separate risks.
  • Banning AI without offering an alternative. A blanket ban often pushes usage onto personal accounts you can’t see. Approved tools plus clear rules work better.
  • Putting secrets in memory or project files. These features make things easier to recall, not safer to store.
  • Uploading third-party files without reading them. Documents, emails and shared “skills” can contain hidden instructions. Treat their content as data, not commands.

Frequently asked questions

Is it safe to upload a contract to get a summary?

Only if your company has approved that tool for confidential documents. Otherwise, remove names, amounts and identifying details first, or summarise the relevant clause in your own words and ask about that.

Does turning off chat history or training make my data private?

It helps, but it isn’t complete privacy. Data may still be stored for a period and handled under the provider’s policies. Check your tool’s current terms, as they change often.

Should small businesses worry about this, or only large companies?

Small businesses often hold the same sensitive data (client details, payroll, contracts) with fewer security staff. Simple habits like placeholders and an approved-tools list are low-cost and protect you.

Is running AI locally the answer?

It can be for specific confidential tasks, as long as nothing leaves your device. It requires suitable hardware and some setup, and local models may be less capable for some tasks. Many businesses combine both: cloud tools for general work, local tools for sensitive files.

Learn the full approach, free

Protecting data is less about fear and more about habits your whole team can follow. The free course How to Be Smarter Than AI covers this step by step in the stations on your privacy and AI, memory and projects and running a model on your own computer. Share it with your team and build the habits together. Start the course here.