Is your company data safe in ChatGPT, Claude, or Copilot?

Our plain-English answer on AI and company data. What a paid plan protects, where the real risk sits, and the rules we set before AI connects to anything.

On a paid team plan with training on your data switched off, your company data is about as safe in ChatGPT, Claude, or Copilot as it is in your email inbox. The bigger risk in the companies we meet is staff pasting files into free personal accounts, with no policy telling them what's allowed. Fix those two things first, then decide what AI is allowed to connect to and change.

What a paid plan gets you

Julian compares it to what happened when fax machines went away. You weren't supposed to email credit card details. Email then got better security and two-factor login, and companies now send almost everything through it.

Paid team plans for Claude and ChatGPT are at the same point. They have two-factor login and the same security certifications as the other software your company runs, and training on your data can be switched off. Copilot comes inside the Microsoft suite many companies already run, which is strong on security and compliance.

Could one of those platforms be breached? Yes, and so could your email provider. If data leaked from a paid plan with training off, every customer's data would be exposed along with yours. You already accept that risk with the rest of your software.

We'd still keep credit card numbers out. Most other work can go in.

Where the risk is

The risk sits in accounts the company doesn't control. A free ChatGPT account can use what a person types to train future models unless they turn that setting off, and few people open the settings at all. On a screen-share with Taylor, one business owner went through hers for the first time. She found the training switch turned on and said she hadn't known it was a feature.

One firm told us its staff were pasting contracts and confidential information into free ChatGPT. Julian's answer was that this is common, so the firm wasn't far behind, and that it still had to be dealt with. People use free tools when nobody has given them an approved one or told them the rules.

The fix is a paid plan for everyone, with the data settings checked on the first day. Checking takes two minutes per tool.

The policy fits on one page

An AI policy for a company this size says what's allowed, what isn't, and what staff should expect. We hand it to every person, and one of our engineers walks the team through it so they know what to put in and what to keep out. Before writing it, we survey the staff on what they use today and what worries them.

At a flagship hotel we trained, one of the 40 sales reps had a paid AI seat when we arrived. Staff were using AI anyway, and the information moving through it was guest and business information. By the end of the week all 40 reps were on the hotel's team plan, under a policy the hotel wrote and published itself.

Team plan or enterprise plan

Buyers ask us whether a team plan is enough or whether they need enterprise before anyone touches AI. Both keep your data out of training by default. Enterprise adds more admin and security controls on top.

We think a team plan with the right settings is enough for most of the companies we work with. The decision belongs to your IT team, and we make it with them on a call before any training or build begins.

Locking everything down has a cost

Taylor describes a pendulum. Startups use whatever gets them furthest and think little about security. Large enterprises think only about security and their people fall behind. Mid-sized companies sit somewhere between, often without having discussed where they want to be.

Some companies allow Copilot and nothing else. That's a safe place to start, and we train teams inside those limits. The cost shows up in your staff. People at restrictive companies have told us they resent being kept off the best tools, and some have started looking for employers where they can build AI skills. They understand the security concern. They can also see how far their approved tool lags the ones leading the market.

A workable policy sets a clear limit and gives staff a way to ask for a new tool and have it tested.

When AI connects to your systems

A chat window is one level of risk. An AI that reads your database or inbox is another, and analysts are right to be nervous about it. These are the rules we follow on every build.

  1. IT signs off before anything connects. Our engineers meet your IT team first.
  2. AI gets read-only access wherever it has no need to change anything, through a restricted user account.
  3. AI drafts and a person approves. It writes the email, and you read it and press send.
  4. Email comes in through lists of approved senders and recipients. We never take in a whole inbox, and we start with active projects only.
  5. We collect only the personal information the job needs, and route data through Canadian data centres when that's required.
  6. Code does the math, so the model can't guess at a number.

In a demo for a real estate firm, built on the firm's own sales data, we asked the system for six comparable sales. It found four and said so. For a market with too few sales it declined to write the valuation memo. That behaviour came from checks built into each step of the process.

Keep your data outside the model

Buyers often ask whether building around Claude locks them in. It doesn't if your documents, meeting notes, and system data live in a layer you own, outside any AI model. We run North Group this way. Scheduled automations pull our email, meetings, and files into one store, and we could plug that store into a different model tomorrow.

Owners also ask whether they'll need a model of their own one day. Julian's view is that every company may eventually have one built on its own history, and that going straight there costs far too much today. The order we recommend is trained people first, then the data layer, then cheaper open-source models for the work that suits them.

What to do today

Open the settings of every AI tool your company uses and read what happens to your data. Then count how many staff are on personal accounts. If you find more than a handful, the first job is a single paid plan and a one-page policy.

If you'd like help with either, book a free audit and we'll go through your settings with you.

More guides.

Find out what we'd build for you.

The free audit is a 30-minute call. We ask how your team works today and tell you where we'd start.

Book a free audit ↗︎