What Not to Paste Into AI: Employee Data Rules for a Small Employer

Sources checked: November 9, 2026. AI-assisted draft, pending editor review.

Cover labeled AI and Employee Data over a locked drawer of records — AI-generated image

It's Friday at 4 p.m., and 15 raise letters have to go out on Monday. Marcus, the office manager at a small landscaping company, has the payroll sheet open: names, Social Security numbers, pay rates and a column of review notes. An AI chat window is open in the next tab. He could paste the whole sheet, type "write a raise letter for each person," and be done before five.

A laptop with a chat bubble beside a padlocked filing cabinet and papers behind glass — AI-generated image · Illustration made for this article

Marcus and his company are made up. The decision in front of him is the one this article is about, and the figures and cases below are real, with sources at the end. No AI product or plan was tested for this article.

What leaves the company when you paste

The moment Marcus pastes that sheet, three questions decide how bad the day could get, and he can't answer any of them. Is what he pasted used to train the vendor's models? How long is it kept, and can he delete it? Who at the vendor can see it? The answers differ by tool and by plan, and the free consumer version he has open may not match the business version his company never bought.

This habit has a name, shadow AI: workers using AI tools nobody approved. IBM's 2025 Cost of a Data Breach report measured it across 600 organizations. A high level of shadow AI added about $670,000 to the global average cost of a breach. In the same study, 63% of organizations had no AI governance policies to manage AI or prevent shadow AI, and 97% of the breached organizations that had an AI-related security incident lacked proper AI access controls.

Those are large organizations, and the dollar figure isn't an estimate for a 15-person company. For a small employer the relevant finding is the missing rule. Most of the organizations studied had none, so each employee made the call Marcus is about to make, alone, on a Friday afternoon.

What one ordinary spreadsheet did to a whole workforce

A file doesn't need bank details to do damage. In August 2023 the Police Service of Northern Ireland answered a freedom-of-information request by releasing a spreadsheet. A hidden tab in that file held the surname, initials, rank and role of all 9,483 officers and staff, and the file was online for about 2 hours 20 minutes.

In October 2024 the UK Information Commissioner's Office fined the PSNI £750,000. It said the fine would have been £5.6 million had the Commissioner not applied a public-sector approach. Account details were not exposed. An ordinary personnel spreadsheet, shared without a check of what was inside it, was enough.

Marcus's sheet has the same problem in plain view. He needs three facts per person for a raise letter. The file carries far more, and a paste sends all of it.

What a breach costs, in IBM's figures

Measure	Figure
Global average cost of a breach, 2024	$4.88 million
Global average cost of a breach, 2025	$4.44 million
U.S. average cost of a breach, 2025	$10.22 million
Added by a high level of shadow AI (global average)	About $670,000

Average cost of a data breach, millions

View the numbers as a table

U.S. average, 2025	$10.22M
Global average, 2024	$4.88M
Global average, 2025	$4.44M
Added by high shadow AI	$0.67M

From the table above (IBM Cost of a Data Breach Report 2025). The last bar is the extra cost, not a total.

The U.S. figure was an all-time high for any region, and IBM's 2026 edition puts the global average at $4.99 million, about 12% higher and a record. IBM's averages aren't estimates for any particular small employer, and real costs vary widely. The cost categories are the same at any size, though: investigation, notification, lost business and legal fees.

Sort the data before anything goes into a chat

The quickest fix is to decide once, ahead of time, which kinds of data can go into a general-purpose AI chat. This table is a starting point.

Data	In a general-purpose AI chat?	Safer option
Job description, public policy text, public handbook sections	Usually fine	Still check for anything confidential
An applicant's resume	Think twice	Remove name, contact details and anything not needed, or use a tool your company approved
Pay data tied to named people	No	Use made-up or aggregated numbers
Social Security numbers, bank details, tax forms	No	Keep inside the payroll or HR system
Medical, leave and accommodation information	No	Keep inside the HR system with limited access
Performance reviews, discipline, investigation notes	No	Keep in the HR system; use a generic example if you need drafting help
Immigration and identity documents	No	Keep in the secure system you use for them

Run Marcus's sheet through it, column by column, and almost nothing survives.

Column in Marcus's sheet	Row in the table above	Goes into the chat?
Employee name	Pay data tied to named people	No
Social Security number	Social Security numbers, bank details, tax forms	No
Current and new pay rate	Pay data tied to named people	No
Review notes	Performance reviews, discipline, investigation notes	No
Effective date of the raise	Not personal by itself	Yes, as a placeholder

How Marcus gets 15 letters without pasting 15 people

The letters don't need the data. They need a good template, and writing a template is the part an AI tool does well.

Ask for one letter with placeholders. Marcus describes the situation in general terms and asks for a short raise letter with [NAME], [OLD RATE], [NEW RATE] and [EFFECTIVE DATE] where the facts will go.
Edit the template. He reads it against how the company actually talks to its people and fixes the tone.
Fill in the facts inside the company's own files. A mail merge in his word processor pulls the names and rates from the payroll sheet. The sheet never goes into the chat.
Check three letters by hand against the sheet before sending all 15.

It takes a little longer than the paste would have, and nothing about any employee went into the AI chat.

Questions to ask before any AI tool touches people data

If the company wants to use an AI tool for more than templates, these are the questions to put to the vendor, starting with the three Marcus couldn't answer.

Is my data used to train the vendor's models, and can I turn that off?
How long is my data kept, and can I delete it?
Who at the vendor can see it, and where is it stored?
Is there a business plan with different data terms from the free consumer version?
Can I control who in my company uses it, and see what they did?
What happens if the vendor has a breach? How will I be told, and how fast?
What happens to my work if the service is down?

Write the answers down, and don't approve the tool until the vendor has answered.

Five principles that already apply to employee files

None of this needs a new framework. The U.S. Federal Trade Commission (FTC) guide on protecting personal information sets out five principles, and they fit an AI chat as well as a filing cabinet.

Take stock. Know what personal information you have in your files and on your computers.
Scale down. Keep only what you need for your business.
Lock it. Protect the information that you keep.
Pitch it. Properly dispose of what you no longer need.
Plan ahead. Create a plan for responding to security incidents.

Marcus's sheet fails the second one before any AI is involved. Social Security numbers and review notes sit in the same file as pay rates because nobody ever split them up.

A one-page rule sheet you can copy

A small company doesn't need a long policy to stop the Friday-afternoon paste. It needs a short rule sheet that everyone has read.

Use only approved tools for anything about real people.
Never paste identifiers, bank details, medical information or discipline files.
Redact or replace names with placeholders when you need help drafting.
Do not let AI make or send decisions about pay, hiring, discipline or termination.
Tell your manager or HR right away if something sensitive was pasted by mistake.
Review the rule sheet when tools or laws change.

The fifth line is what makes a mistake fixable. A paste nobody reports can't be dealt with. This is where the FTC's "plan ahead" principle comes in: decide now who gets told and what happens next, so nobody has to work it out on the day.

If AI is part of hiring or other employment decisions

Drafting letters is one use. Screening, ranking or scoring people is a different one, and several places add notice and fairness rules for it. Illinois, for example, requires employers to notify employees when they use AI for covered employment decisions such as hiring, promotion or discipline. New York City requires a bias audit of a covered tool and notice to resident candidates at least 10 business days before it is used in hiring or promotion. AI in Hiring covers those rules. Telling employees what you use and why is a good habit even where it is not required.

Want the next guide when it's published? Join the list.

General information, not legal advice. Reported figures come from the cited sources.