Incident sheet · /security

Filed
Read
4 min · 848 words
Severity
Routine

One ChatGPT spreadsheet cost a therapist her job — what your staff can safely tell AI

Pseudonymized spreadsheets, hidden columns, and the wrong AI tool have already cost jobs. Engineers and DPOs explain how to decide what data can safely reach an AI model.

What can you safely tell AI at work?
FeedWhat can you safely tell AI at work? — Firmbee / pixabay

Event log

  1. A US mental health therapist was dismissed after using ChatGPT instead of her employer's approved AI tool, despite entering no client data.
  2. A health tech employee uploaded treatment notes keyed by customer number to an unapproved AI service — the numbers still linked back to identifiable patients.
  3. Group 1001's CTSO pairs AI principles with data classification, DLP, access controls, monitoring and incident response rather than policy documents alone.
  4. Cornerstone Building Brands' Sawant rebuilds errors with fictional data before sharing them with AI, or keeps them internal when context is too sensitive.

A mental health therapist used ChatGPT to build a billing spreadsheet — entered no client data, admitted the mistake, and was fired anyway. Her case, recounted by Chicago litigation attorney Howard Krauskopf, who fields such questions on AskaLawyerOnCall and JustAnswer, shows how steep the price of picking the wrong AI tool has become at work, even when nothing confidential leaks.

The story starts with an error message. Bhaskar Bharat Sawant, a lead engineer and solution architect at Cornerstone Building Brands, troubleshoots legacy and modern applications daily, and generative AI can explain what's gone wrong. But the input an AI tool needs to help often carries personally identifiable information, a customer's competitive secrets, or proprietary internals.

"Logs or error messages can reveal internal service names, configuration details, or business processes," Sawant said. "You have to protect that information whether it belongs to your employer or a customer."

What stays identifiable after you strip the names?

The practical problem is that employees can strip obvious identifiers and still leak plenty. Julian Gage, who provides fractional data protection officer services through Engage Compliance, recalls a health technology company where an employee uploaded a spreadsheet of treatment notes and clinic customer numbers to an unapproved AI service. No patient names appeared — but each customer number linked back to an identifiable person.

"Replacing a name with a customer number doesn't make the information anonymous," Gage said. "Someone with access to the clinic's records could still establish whose treatment notes they were."

The blind spots multiply quickly:

  • Meeting transcripts can identify someone through context — an absence, a medical treatment, distinguishing circumstances — even with names removed. As Gage puts it: "You have to look at the content, not just the names."
  • Screenshots can carry identifying details outside the passage an employee wants analyzed.
  • Spreadsheets can retain hidden columns in the uploaded file.
  • Automated connections can pipe transcripts to an AI service before anyone reviews them.

Sawant's workaround is deliberately labor-intensive. He reproduces the problem in a controlled development environment, isolates the failing component, then builds a fictional miniature: "I create a small code example with made-up data that reproduces the issue." He asks whether an outsider could reconstruct proprietary information from it; if so, he simplifies further or keeps it out of the AI tool entirely — and sometimes just continues troubleshooting with internal resources.

Is an approved tool a free pass?

No, said Sawant: "An approved AI tool isn't a free pass to share every kind of information." Employees still need to know what data it's approved to handle, how inputs are stored or used, and who can access the conversation.

Gage argues employers should settle those questions before staff ever face them — vendor contracts, data-processing terms, model-training arrangements, access controls, retention settings — then hand employees the usable output: which account to use, which tasks are permitted, what can go in. "There are ways to use AI with personal and confidential business information, provided the appropriate contracts and safeguards are in place," he said.

The wrong route carries consequences regardless of intent. The therapist behind on billing broke a workplace instruction by using ChatGPT instead of the sanctioned tool, and Krauskopf notes US employers generally don't need proven misconduct to dismiss an at-will employee. His advice: "Ask before you use it, and put the question in writing." If things go wrong, a calm appeal works best — "Don't attack. Explain: 'I followed the company's instructions. Here is what I did.'" Larger organizations offer formal reconsideration procedures; smaller ones hinge on an individual owner, and neither guarantees a sympathetic hearing.

How much judgment should sit with the employee?

For Chris Kennedy, chief technology and security officer at insurance group Group 1001, a written policy alone won't carry the load. "How many people go hunting for a policy, rather than asking, 'What do I do in this situation?'" he said. "I never expect a policy statement to do all the thinking for the employee."

His approach pairs principles with examples and training, backed by data classification, access controls, data leakage prevention, monitoring, and incident response — plus an ongoing effort to consolidate data and label sensitive information so controls can recognize it. For narrowly defined operational roles, Group 1001 wants the tooling to decide what data goes where. Developers and product teams, whose work generates proprietary information of unforeseen value, keep continuing access to managers and governance specialists instead.

Kennedy flags a creeping frontier: existing third-party software suddenly arriving with new AI-enabled functionality an employee never signed up for. Companies need to explain permitted uses, watch how AI is actually used, and update guidance as features emerge.

What does all this change for the people doing the work? Sawant's answer is support, not slogans: clear guidance with real examples, someone in security or data governance for hard cases, and managers who accept that "a task is taking longer because they're waiting for guidance or approval." People shouldn't feel stuck — or worried about making the mistake that ends a career.

via linkedin.com (Original)

Filed under

  • workplace-ai-safety
  • data-privacy
  • ai-governance
  • data-leakage-prevention
  • employee-training
Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

News editor covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.

207 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›