Safe Day-to-Day Use of AI at Work
Literacy becomes real at the moment somebody is about to press send. This lesson sets out a short routine that applies to any assistant, any tool and any task, and that will still work when the tools change.
The routine begins with the tool itself. Use approved tools for work, and know why the list exists. An approved tool has usually been through a contract that governs what the supplier may do with your data, a security review, and a decision about where processing happens. A personal account for the same brand of assistant, opened with a personal email address, sits outside all of that. The interface may look identical while the contractual position is entirely different, which is why “it is the same model” is not an argument. If a tool you need is not approved, the correct action is to ask for it to be assessed rather than to use it quietly, and organisations that make that request easy have far less shadow usage than organisations that make it slow.
The second step is to look at what you are about to send. Before any material goes into a prompt, an upload or a connected folder, decide what class of information it is. Personal data about colleagues, customers, patients, pupils or claimants deserves particular care, and so does material covered by confidentiality obligations to a client, unpublished financial information, security details, legally privileged material and anything covered by a non-disclosure agreement. The working question is simple, which is whether you could name where this data is going and what the supplier is permitted to do with it. If you cannot answer that, you are not ready to send it. Removing names is often not enough, because a case description can identify a person as reliably as a name does.
The third step concerns what you ask for. Give the model the material it needs rather than expecting it to know. Asking a general-purpose assistant for your organisation’s refund policy invites fabrication, while giving it the policy and asking it to answer a customer’s question from that text is a task it can do well. Asking for sources when the model has no access to sources produces invented ones, whereas asking it to work only from a document you have provided is verifiable. Specific tasks with supplied material fail less often than open questions about the world.
The fourth step is what you do with the output, and it is where most incidents are prevented. Treat everything that comes out as a draft written by a capable but unreliable colleague who never says “I do not know”. Check anything you would be embarrassed to be wrong about, and check it against something that did not come from the model. Numbers, names, dates, quotations, legal references, medical statements and anything you are about to send outside the organisation all need that check. Read the whole output rather than the first paragraph, because errors cluster where the material thins out. If the output will inform a decision about a person, the check is not optional and the reasoning should be recorded.
The fifth step is disclosure and records. Follow your organisation’s rule on telling people when AI was involved, which the next lessons cover in detail. Keep enough of a record that you could reconstruct what happened, which for most staff means saving the prompt and the output alongside the work product when the output materially shaped a decision. This costs seconds and is the single most useful habit for the evidence side of Article 4.
Some situations should stop you outright. Do not use a general-purpose assistant to make or effectively make a decision about a person’s employment, credit, benefits, care or education. Do not use it as a source of legal, medical, tax or safety advice for someone else without a qualified human check. Do not use it to generate content that impersonates a real person, and do not use it in a way that an emotion-inference or biometric rule might catch without checking first, because the prohibited practices in Article 5 have been in force since 2 February 2025 and carry penalties of up to 35 million euro or 7 percent of worldwide annual turnover. Do not paste material you were trusted with by someone who has not agreed to this.
A few habits deserve a mention because they cause quiet trouble. Meeting transcription tools that join calls capture everything said, including the part after the client thinks the recording stopped, so tell participants. Browser extensions that read pages read every page, including internal systems. Connected assistants that can send email or edit files on your behalf turn a bad output into a bad action, so grant those permissions narrowly. Personal devices blur the boundary between an approved and unapproved account, which is why the account matters more than the device.
Finally, report problems rather than working around them. If an assistant produces something strange after reading an external document, if it reveals information it should not have, or if you discover that a colleague has been feeding customer records into a personal account, the organisation needs to know quickly. Incident reports are also evidence of a working programme, and a system with no reported issues is usually a system nobody is watching.