Workplace AI Privacy: The Information to Keep Out
A practical workplace AI privacy guide for deciding which customer data, payroll details, contracts and public copy should stay out of AI tools.
Workplace AI privacy starts before anyone presses send. The practical question is not whether an AI tool feels clever. It is whether the information you paste into it is public, approved, necessary and safe to process in that tool.
The Short Version
- Red information stays out of general AI tools: customer records, payroll, passwords, private contracts, board papers and unpublished plans.
- Amber information needs cleaning first: remove names, exact figures, email chains and anything that identifies a person or client.
- Green information is usually safer: public copy, blank templates, fictional examples and prompts with sensitive details removed.
- If the tool is not approved for that kind of work data, treat it like an outside service and ask before pasting.
For most teams, the simplest rule is a traffic-light test. Red information stays out. Amber information needs permission, masking or an approved company tool. Green information is usually safe because it is already public, low-risk or stripped of identifying details.
This is not legal advice, and it is not a substitute for your employer’s data protection, security or client confidentiality rules. It is a working guide for everyday judgement: when to pause, when to ask, and when AI can help draft without becoming the place where sensitive work data leaks. For the broader consumer version, see Cristoniq’s explainer on what information should never go into an AI tool.
Workplace AI privacy in plain English
AI tools can feel like smarter search boxes, but many of them are closer to cloud services. Text you paste may be processed outside your direct control, stored for a period, reviewed under service terms, or handled in ways your company has not approved. That is why ICO guidance on AI and data protection focuses on principles such as purpose, minimisation, security, transparency and accountability when personal data is involved.
At work, the safer habit is to treat AI as a drafter, organiser or second pair of eyes, not as a private filing cabinet. If the tool does not need a name, account number, salary figure, contract clause or unreleased product plan, do not include it.
The red list: never paste this into an AI tool
Red information should stay out of general AI tools unless your organisation has explicitly approved that tool and that use case. The list includes customer records, personal addresses, phone numbers, email addresses, payment details, payroll data, health information, HR case details, disciplinary material, passwords, API keys, access tokens, confidential contracts, board papers, unpublished financials, source code from private systems and commercially sensitive plans.
It also includes anything you would hesitate to forward to an external supplier. A customer complaint with names removed may still reveal who the person is if it includes dates, locations, order details and unusual circumstances. A contract clause may reveal more than its wording if the counterparty, price or negotiation context is included. A spreadsheet can expose sensitive information even when the visible task looks routine.
If you are unsure, assume red until someone with the right authority says otherwise. That is especially important for personal data, client work, supplier agreements and material your employer has marked confidential.
The amber list: use only after cleaning it up
Amber information can sometimes be used, but only after you reduce the risk. This includes draft emails, meeting notes, support tickets, internal process documents, project summaries, sales call notes, research notes and anonymised examples. The question is whether the AI tool needs the sensitive parts to do the job.
A useful amber habit is to replace real details with safe placeholders. Use “Customer A” instead of a customer name. Replace exact revenue figures with ranges. Remove employee names from a performance discussion before asking for a clearer structure. Summarise the issue in your own words rather than pasting a full email chain. If the task is to improve the tone of a message, the AI does not need the full attachment, the full client history or every internal comment.
This is where a small business AI policy helps. A simple policy can say which tools are approved, which data types are banned, who can approve exceptions and how staff should anonymise examples. Cristoniq’s guide to creating a simple AI policy for a small business is a useful companion if your team is still relying on informal judgement.
The green list: low-risk prompts that still need judgement
Green information is usually safe to use because it is public, generic or created specifically for the prompt. Examples include public website copy, a blank template, a made-up customer scenario, a generic project plan, a job advert already published by your company, or your own rough notes after confidential details have been removed.
Even then, keep the prompt narrow. Ask for a structure, a checklist, alternative wording, a summary or questions to consider. Do not ask the tool to make the decision for you. If you use AI to draft a reply, check the facts, tone and missing context before sending it. The same human review habit applies to any output, as covered in Cristoniq’s guide to checking an AI draft before you send it at work.
How to decide in ten seconds
Before pasting anything into an AI tool, ask four questions:
- Would I be comfortable sending this exact text to an external supplier?
- Does the AI tool need the real names, numbers, dates or documents to help?
- Has my organisation approved this tool for this kind of work data?
- Could this harm a person, customer, colleague or business relationship if it leaked?
If any answer is no, stop and simplify the prompt. Strip out identifiers. Turn the material into a fictional example. Use a company-approved tool. Ask your manager, data protection lead, IT team or policy owner. The point is not to make every prompt slow. It is to make the risky ones visible before they leave your control.
The NCSC’s guidance on using SaaS securely is a useful reminder that organisations need confidence in how cloud applications protect data, handle access and manage provider-side risks. For workers, that translates into a simple habit: do not treat an unapproved AI tool as if it has already passed your company’s security review.
Where shadow AI fits in
Shadow AI happens when people use AI tools outside the approved process because the tool is convenient, fast or already open in a browser tab. It is often not malicious. It is usually a sign that staff need clearer rules, better approved tools or examples that match real work. The risk is that confidential information moves into systems the organisation cannot monitor or govern.
If your team is already using AI informally, the answer is rarely to pretend it is not happening. Start with the red, amber and green list. Then agree what work belongs in approved tools, what should stay out completely and what needs human review. Cristoniq’s explainer on shadow AI at work sets out why this matters for ordinary teams, not just large technology departments.
What this means for you
Good workplace AI privacy is a habit, not a slogan. Keep personal data, confidential documents and sensitive business information out of general AI tools. Clean up amber examples before using them. Prefer public, generic or fictional material where possible. Use approved tools for approved tasks. When the line is unclear, ask before pasting.
AI can still be useful. It can turn rough notes into a clearer outline, suggest questions, tidy wording and help you spot gaps. But the human should choose the input, check the output and own the final decision. In a workplace, privacy starts before the prompt is sent.
In plain English
Do not paste anything into an AI tool that you would not be allowed to share with an outside service. For red information, do not use it. For amber information, remove sensitive details or use an approved process. For green information, keep the task narrow and check the result. AI should help you draft, not become the place where private work information goes to live.