Insights
What your team should never paste into an AI tool
The practical risk of AI at work is not a dramatic breach. It is ordinary staff pasting customer records, unsigned contracts, salary data and board material into consumer tools, because nobody told them where the line is. A workable boundary fits on one page: three categories of information, four questions to ask before pasting, and a named person to ask when the answer is unclear. This guide, by Ahmed Amrousy — Executive Education Instructor at the American University in Cairo and founder of Menova — sets out that page, and explains why a blanket ban produces the worst outcome of all.
Last updated:
The risk of AI at work is not the one that gets discussed. It is not a model leaking your strategy to a competitor. It is an account manager pasting a customer list into a consumer chatbot to tidy up the formatting, because it was faster and nobody had told them not to.
That happens in organisations that have an AI policy. It happens because the policy is four pages long, lives on an intranet, and does not answer the question the person actually had.
A boundary that works fits on one page.
Three categories, not a list of rules
Red — never. Material you cannot put into a third-party tool without breaking an obligation you have already made:
- Anything that identifies a customer or an employee by name, account or record
- Anything covered by a confidentiality agreement you have signed
- Unreleased financial results and board material
- Credentials, access keys, tokens
- Anything in a live legal, regulatory or disciplinary matter
Amber — only with the identifying details removed. The bulk of real work. A contract becomes “a supply agreement with a vendor”. A customer complaint becomes “a complaint from an enterprise customer in the Gulf”. The work gets done; the identity does not travel.
Green — freely. Published material, your own drafts, public data, general questions, anything you would put in an email to someone outside the organisation without thinking about it.
The point of three categories rather than a list is that staff meet situations your list did not anticipate. Categories travel to new cases. Lists do not.
The four questions
Before pasting anything, four questions, in order:
- Would I put this in an email to someone outside the company? If no, it is not green.
- Can I take the identifying details out and still get a useful answer? Usually yes. Do that.
- Have I promised someone I would not share this? A signed NDA, a customer’s expectation, an employment obligation. If yes, it is red regardless of how harmless the request feels.
- If I am not sure — who do I ask, and will I get an answer in ten minutes?
Question four is the one organisations get wrong. Boundary questions are rarely difficult, but they are always urgent, because someone is mid-task. If the answer takes two days, people will guess. The guessing is the risk, not the technology.
Where anonymisation fails
Removing names is the most useful habit to teach, and it is not sufficient on its own. The combination of details can identify even when no name appears: the only customer of that size in that market, the acquisition everyone in the industry is already speculating about, the employee who is obviously the one with that job title in that office.
Teach the habit, and teach its limit: if removing the names leaves something that only one party could be, treat it as red.
Why banning it is the worst option
A ban does not remove the tool. It removes your visibility of it.
Staff under deadline who believe the tool will help them will use it on a personal phone, in a personal account, with whatever default settings that account has. You lose the logging, the enterprise settings, the plan terms you negotiated, and any chance of knowing what went where. The organisation has not reduced its exposure; it has moved it somewhere it cannot see.
A narrow written permission with one clear boundary beats a broad ban nobody follows. This is not an argument for permissiveness. It is an argument for being able to see what is happening.
On “is our data used to train the model?”
It depends on the vendor, the plan and the terms in force — and all three change.
Do not take this from an article, a trainer, a vendor’s marketing page or an internal memo written last year. Including this one. Have someone read the terms for the specific plan you are on, write down the date they read them, and re-check whenever the plan or the vendor’s terms change. Any source that tells you the answer without naming the plan and the date is guessing.
The review gate nobody sets up
One more rule, and it is the one that prevents the embarrassment rather than the breach: anything that leaves the organisation gets read by a human who is accountable for it.
Proposals, client emails, regulatory filings, public statements. Not because the output is usually wrong, but because it is occasionally, confidently wrong in a way that reads perfectly well. The failure mode of this technology is not gibberish. It is plausibility.
What to put in place first
Not a policy document. In order:
- The one page above, adapted to your obligations.
- A named person, reachable in minutes, who decides the unclear cases.
- One session where people bring their actual work and find out where the line falls for their own tasks — because the boundary only becomes real when someone discovers that the thing they do every Tuesday is amber.
A policy that is read once changes nothing. A boundary that someone has used on their own work changes how they work.
This is the first session in every Menova programme, before any tool is opened. If you are weighing providers, the related questions are nine questions to ask before booking and what the training should cost.
Ahmed Amrousy is an Egyptian AI-for-Business Expert, Executive Education Instructor at the American University in Cairo’s Onsi Sawiris School of Business, Head of Marketing & PR at Nile Air, and the founder of Menova.