How to Use AI at Work Without Leaking Your Data

Last updated June 2026. First published with current OpenAI and Microsoft data-handling references, a what-not-to-paste list, safe patterns, and the consumer vs enterprise vs managed comparison.

AI tools are genuinely useful at work. They draft emails, summarize long documents, explain error messages, and turn rough notes into something presentable. The problem is not whether to use them. It is that the easiest way to use them, pasting real work into a free personal account, is also the way most likely to put your business data somewhere you did not intend.

This is a practical guide, not a warning to stop. Used with a few simple rules, AI is safe for most day-to-day work. Here is what actually happens to your data, what to keep out of public tools, and the patterns that let your team move fast without creating a privacy problem.

What Actually Happens to Data You Paste In

When you type into a public AI chatbot, your text leaves your device and is processed on the provider's servers. Two things matter for privacy: whether the provider uses your conversations to train future models, and how long they keep your data.

On consumer tiers, training is often on by default. OpenAI states that content from personal ChatGPT accounts (Free, Plus, and Pro) may be used to help improve its models unless you turn that setting off, while business products are treated differently. In its own words, "we don't train our models on inputs and outputs from" business offerings such as ChatGPT Team, Enterprise, and the API by default. See OpenAI's guidance on how your data is used to improve model performance for the current settings and opt-out steps.

The risk is not theoretical. A 2023 Cyberhaven analysis of real workplace usage across roughly 1.6 million workers found that a meaningful share of what people paste into ChatGPT is confidential, reporting that 11% of data employees paste into ChatGPT is sensitive. That is the everyday reality a policy has to account for: well-meaning staff pasting client lists, contracts, or code into a tool because it is the fastest way to get help.

Two practical takeaways. First, the account tier you use changes the rules, so it matters whether your team is on a personal login or a business one. Second, "it deletes after a while" is not the same as "it was never used," so the safest habit is to control what goes in rather than rely on what happens after.

What You Should Never Paste Into a Public AI Tool

A simple rule covers most cases: if you would not post it publicly or email it to a stranger, do not paste it into a personal AI account. Concretely, keep these out of consumer chatbots unless you are on a business tier with a data protection agreement in place:

• Customer and client personal information: names tied to contact details, account numbers, health or financial data, anything covered by privacy law.
• Credentials and secrets: passwords, API keys, access tokens, connection strings, or anything that grants access to a system.
• Source code and proprietary logic: internal code, algorithms, or configuration that represents your competitive advantage.
• Unreleased or confidential business material: contracts, financials, M&A details, pricing models, or strategy documents.
• Regulated data: personal health information, payment card data, or anything you are contractually or legally required to safeguard.

For Canadian businesses, that last point connects directly to your obligations. Pasting a client's personal information into a tool that may retain or learn from it can undercut the safeguards you are expected to maintain. If that applies to you, our PIPEDA compliance checklist is a good companion read.

Safe Patterns You Can Use Today

You do not have to choose between "use AI freely" and "ban it." These patterns keep the value while removing most of the risk.

• Redact before you paste. Replace real names, numbers, and identifiers with placeholders (Client A, $X, account 0000). The AI can still help with structure, tone, and logic without ever seeing the real data.
• Work on the shape, not the secret. Ask for a template, a checklist, or a rewrite of a generic version, then fill in the sensitive details yourself afterward in your own systems.
• Use a business or enterprise tier for real work. If staff genuinely need to process company information with AI, give them an account that contractually excludes their input from training, rather than letting them improvise on personal logins.
• Prefer tools that connect to your own data securely. AI that works inside your existing environment, under your access controls, keeps data within a boundary you control instead of copying it into a public chat window.
• Turn off training where you can. On consumer tiers that allow it, switch off the "improve the model for everyone" setting so new conversations are not used for training.
• Write the rule down. A one-page acceptable-use note that lists what is fine and what is off-limits prevents most accidental leaks, because people generally follow clear guidance when it exists.

The redaction habit alone removes a large share of the risk for everyday tasks like drafting and summarizing. It costs a few seconds and means the tool never holds anything you would mind it remembering.

Consumer vs Enterprise vs Managed AI

The same chatbot can be safe or risky depending on the tier. Here is the honest difference.

The major vendors draw the same line for their business products. Microsoft, for example, states that for Microsoft 365 Copilot, "your prompts, responses, and data accessed through Microsoft Graph aren't used to train the foundation models," as described in its enterprise data protection documentation. The pattern is consistent: consumer tiers may learn from your inputs, business tiers contractually do not.

Managed AI goes one step further than simply buying the business tier. Instead of each person deciding what is safe to paste, the AI is built into the systems your business already runs on, kept inside your access controls, and monitored so the safe path is the default path. That is the model behind ClayGen's Managed AI service: built in, monitored, and secured, so staff get the help without each one having to be a privacy expert.

When AI Is Not the Right Tool

Being honest about the limits is part of using AI well. There are tasks where a public chatbot is the wrong choice no matter how you redact:

• When accuracy must be guaranteed. AI can be confidently wrong. For legal, medical, financial, or safety-critical answers, treat it as a drafting aid that a qualified human verifies, not the source of truth.
• When the data simply cannot leave your environment. Some regulated or contractual data should never go to a third-party tool at all. In that case the answer is AI that runs inside your own boundary, not a consumer chatbot with careful wording.
• When the task is trivial. If a saved template or a two-line script already does the job reliably, adding AI just adds a data-handling decision you did not need to make.

Knowing when not to reach for AI is what keeps the tool trustworthy. A team that uses it for the right things and verifies the rest gets the upside without the nasty surprises.

A Short Checklist Before You Hit Enter

• Am I on a personal account or a business tier? Sensitive work belongs on the business tier.
• Does this text contain customer data, credentials, code, or regulated information? If yes, redact or move it.
• Could I replace the real details with placeholders and still get the help I need? Usually yes.
• Is training turned off on this account where that option exists?
• If accuracy matters, who is verifying the output before it is used?

Run that list once and it becomes automatic. Most accidental data leaks come from skipping it in a hurry, not from anyone acting in bad faith.

If your team is already using AI and you want it to be safe by default rather than by everyone's best guess, that is exactly what Managed AI is for. You are welcome to talk it through with us, no pressure, and we will give you an honest read on what is fine to keep doing and what is worth tightening up.