Three questions to ask before any work data goes into an AI tool. They take less than five minutes, and they protect you, your clients and your job.
Most workplaces say they want innovation, but what they reward is predictability. Do the work the way it has already been approved. Do not break anything. Do not create risk. That is why the only experiments that survive approval are the ones that feel boring. They are small enough to ignore, constrained enough to control, and reversible enough that no one has to defend them.
That is the purpose of the box. It creates a narrow space where work can be examined without triggering fear. Think of the box as a boundary around work, not a tool.
Inside the box is where all the preparation happens. Drafting, summarizing, organizing, classifying, and cutting through noise all live here, and the reason this space is safe for automation is that nothing final happens inside it. Work can be wrong and it does not matter because there is always a chance to catch it and fix it before anyone else ever sees it.
Outside the box is where commitment lives. Sending, publishing, approving, making decisions, and changing records are all actions that create real impact the moment they happen and cannot be quietly undone after the fact. The rule that separates the two is simple. Automation can operate as freely as it wants inside the box, where mistakes are safe, observable, and easy to correct. It stops at the edge of the box, right where real responsibility and real consequences begin.
This is the principle that separates teams who use AI well from teams who use it recklessly. Automation does not replace judgment. It makes judgment sharper and more effective, but only when it is applied to the right part of the process. When the preparation, the organization, the pattern recognition, and the noise reduction are all handled before a human being ever has to make a call, that person gets something they almost never have in most workplaces. A clear signal at the exact moment it matters most. The goal is never to pull people out of the process entirely. The goal is to strip away everything that was preventing them from thinking clearly while they were inside it.
There is another boundary that operates alongside the box, and this one is non-negotiable. Not every piece of data belongs in every tool, no matter how convenient it might be to put it there. Before connecting a system or uploading a file, three questions need to be answered honestly.
Inside. Draft. Sort. Summarize
At the boundary. You check the work. Outside. Send. Approve. Decide.
Question 1
The first question to ask is whether the tool you are using is approved by your company to handle business data. If your organization has not given the green light on a specific AI tool, do not paste customer names, deal sizes, internal strategy, or employee details into it. Saving a few minutes is never worth the risk of putting sensitive information somewhere it was never meant to go.
Shadow AI is any AI tool people use for work without their company’s approval, including a personal chatbot account, a browser extension or a free transcription app.
In an Okta-commissioned survey, 52% of knowledge workers admitted using unapproved AI tools, while 90% of executives said they were confident they could see which tools were in use, according toThe Register. Most people are trying to get their work done without an approved option.
Start with an open conversation.
Ask, without blame. “Which AI tools are you using for work, including free ones?” You’ll learn more in one honest team meeting than from any scan.
Check with IT. Sign-in logs, browser extensions and expense reports show which AI apps are already in use.
Give people an approved option. Shadow AI shrinks fastest when there’s a sanctioned tool and a clear one-page rule about what data can go into it.
Question 2
The second question is whether the data you are working with is protected by any laws or industry rules. Health records, financial data and personal information often fall under laws like GDPR and HIPAA, and security standards like SOC 2 that many companies require from vendors. These requirements call for approved tools and controls suited to the data. Running patient data through a tool that has not been approved for that kind of information puts people and your organization at risk.
It depends on which ChatGPT you’re using. A personal account and a company-managed workspace have different data terms, so check which one you’re signed into before you paste anything. If it’s a personal account, treat it as outside your company. No customer names, deal details, health or financial records, or employee information. If your company provides ChatGPT through a business plan, follow its rules on what data is allowed.
Microsoft 365 Copilot works inside your company’s Microsoft 365 environment and respects the file permissions you already have. Microsoft explains how Copilot handles your data on itsprivacy page.
Safe use still depends on your company’s setup. Copilot can find anything you have access to, including files that were shared too widely. Ask the three questions on this page before you use it with sensitive data.
Question 3
If this data got out tomorrow, what would happen? If the answer is lawsuits, fines, lost customer trust or a competitor gaining sensitive information, ask security and compliance to review the workflow before that data touches an automation.
Start with data that would cause little harm if something went wrong. Prove that the workflow does what it should, and move toward more sensitive information only after the responsible people have approved it.
Print it, keep it by your keyboard, and run any new AI task through it before you paste anything in.
If your team keeps asking “Is this allowed?”, it’s time for a one-page AI policy. Ashley helps teams write one that people actually follow, as part of her AI data security and acceptable use training.