How-to6 min read · Updated

How to use AI at work safely: accounts, data and checks that keep you out of trouble

AI can save you hours at work, and most of the risk comes from three avoidable habits. This guide shows the account, data and checking rules that let you use it with confidence.

The short answer

Using AI at work safely means using your company's business account, keeping confidential and personal data out of personal accounts, and checking every output before it leaves your hands. ChatGPT Business, Claude Team and Enterprise, Google Workspace with Gemini and Microsoft 365 Copilot with a work account do not use your content to train models by default. Personal plans, including paid ones, do unless you switch it off.

What you need before you start

  • Your company's AI policy, or the name of the person who decides. If there is no policy, ask before you paste any company data into a tool. If you are that person, start with how to write an AI policy for your company.
  • To know which account you are signed in with. A personal ChatGPT, Claude or Gemini account and a work account on the same tool follow very different data rules.
  • A rough sense of your data. Public, internal, confidential and personal data about customers or colleagues each need different handling.

Using AI at work safely is mostly about three things: the right account, the right data and a human check before anything leaves your hands. None of it needs technical knowledge.

Step by step: how to use AI at work safely

  1. Use the work account your company provides. On business plans the defaults protect you. ChatGPT Business (for 2 to 200 people) and Enterprise content is not used for training by default. Claude Team (2 to 150 users) and Enterprise content is not used for training by default. With a Google Workspace account, Gemini content is not human reviewed or used for training outside your domain without permission. Microsoft 365 Copilot with a work account runs under enterprise data protection, and prompts, responses and Microsoft Graph data are not used to train foundation models.
  2. If you must use a personal account, lock it down first. In ChatGPT, open Settings, then Data controls, and turn off Improve the model for everyone. In Claude, open Settings, then Privacy, and switch off Help Improve our AI models. In Gemini, turn off Keep Activity in Gemini Apps Activity. Then keep confidential data out anyway: a personal account has no admin, no contract and no audit trail.
  3. Sort the data before you paste. Public material, such as your website copy, is fine anywhere. Internal material, such as meeting notes, belongs in a work account. Confidential and personal data, such as contracts, salaries, health details and customer records, belongs only in a tool your company has approved for it, and often not even there. Replace names with roles when the task does not need them.
  4. Connect apps one at a time. Plugins and connectors let an assistant read your email, calendar and files. ChatGPT has Gmail, Google Calendar and Outlook connections; Claude has Gmail, Google Calendar and Google Drive connectors, and on Team and Enterprise the owner has to enable them first. Connect only what the task needs and check what it can see.
  5. Keep agents on a short lead. Tools that act for you, such as Claude Cowork tasks, ChatGPT Work and ChatGPT dots, can send, edit and delete. Claude asks before sending, replying to or forwarding email; with a dot you set what it may do alone. Start with "draft and wait for my approval" for anything that leaves the company.
  6. Check every output that matters. AI still states wrong facts confidently. Check names, numbers, dates, quotes and legal or medical statements against the source, and ask the tool to show its sources. See how to check if AI is telling the truth.
  7. Be open about AI use where it matters. Your manager and clients should not be surprised. The European Commission says people using chatbots should be made aware they are talking to a machine, with the AI Act's transparency rules taking effect in August 2026.

Worked example: a client proposal in under an hour

You are preparing a proposal for a prospective client. Your company has ChatGPT Business.

  1. Sign in with the work account, not your personal one, and check the workspace name in the corner.
  2. Paste your company's two best past proposals with client names replaced by "Client A" and "Client B", and the prospect's public brief from their website.
  3. Ask for a proposal outline that follows your house structure, with a list of questions you still need to answer. Do not paste internal pricing sheets; add the price yourself.
  4. Draft each section, then ask the assistant to check the draft against the brief and list anything promised that you cannot deliver.
  5. Read the whole thing as if you were the client, fix the tone, add the price and have a colleague review it before it is sent.

The AI did the structure and the first draft. You kept the price, the promises and the final decision, which is exactly the split that keeps you safe.

What to keep human, whatever the tool

  • Decisions about people: hiring, performance, pay and dismissal.
  • Anything you sign or send in your company's name without reading it.
  • Legal, medical and financial advice to customers.
  • Final numbers in reports, invoices and filings.
  • Apologies and bad news to clients. People can tell.

Common mistakes and how to fix them

  • Using a personal paid plan for work. Plus, Pro, Max and Ultra are personal plans, and training is on by default. Fix: ask for a work seat, or at least switch training off and keep confidential data out.
  • Pasting a whole spreadsheet of customers. Fix: remove names and contact details first, or ask the AI to write the formula and run it yourself.
  • Trusting the opt-out to undo the past. Anthropic says switching off training stops future use of new chats, not training already done. Fix: set it before you start.
  • Letting an agent send email unsupervised. Fix: require approval for every external message until you trust the workflow.
  • Copying AI text into a client deliverable without checking. Fix: read it, check every fact, and make it sound like you.
  • Using a free translator for confidential files. DeepL Free trains on submitted text, including corrections, with no opt-out. Fix: use a paid plan, where texts are deleted after translation.

Which tool to use at work

Use the one your company has a business plan for, because that is where the protection comes from. If you are choosing: Microsoft 365 Copilot suits companies that live in Outlook, Teams and Excel; Gemini with Google Workspace suits companies on Gmail and Docs; ChatGPT Business is the broadest all-rounder; Claude Team is our pick for careful work on long, sensitive documents. Compare them in ChatGPT vs Copilot, ChatGPT vs Claude and Copilot vs Gemini, see the best private AI for sensitive data, and pick up role-based prompt sets in Yuzoki kits.

Prompts to copy

Draft from your house style, not your secrets

Here are two past proposals with client names removed and the public brief from a prospect. Write an outline for a new proposal in the same structure and tone. Leave [PRICE] and [TIMELINE] as placeholders, and list the questions I still need to answer. [paste proposals and brief]

You get: A structured first draft with the sensitive parts left for you

Find the promises before the client does

Check this draft against the client brief. List every promise, deadline or number in the draft, say whether the brief asks for it, and flag anything that sounds like a commitment we may not be able to keep. [paste draft and brief]

You get: A risk list to fix before sending

Is this safe to paste?

Before I share anything, tell me which of these items count as personal or confidential data and how to rewrite each so the task still works: [describe the data you want to use, without pasting it].

You get: A quick data check that never needs the data itself

Show your working

Answer this question and, for every factual claim, give the source link and its date. Then list the claims you are least sure about, so I know what to check first: [question].

You get: An answer that tells you where to look

Ask before acting

Draft replies to the three most urgent emails in my inbox from today. Do not send anything. Show me each draft with the original email, and wait for me to approve, edit or discard each one.

You get: Inbox help that keeps you in control; needs a connected mail account

Keep going

Start the ChatGPT lessons Start the Claude lessons