Here’s what we tell every client who asks about AI: your employees are already using ChatGPT. The real question is whether anyone is managing it. In many organizations, staff paste client emails, spreadsheets, and contracts into personal accounts that leadership never sees.
The short answer: ChatGPT is safe for business when it’s governed: the right plan, a written policy, trained people, and clean data permissions. Without those, it’s an unmonitored exit for sensitive data.
Windes has helped businesses navigate new risks for 100 years. This guide covers what happens to your data, the biggest risks, the 2026 rules, and the five-step roadmap we use with clients.
What Happens to What You Type Into ChatGPT?
Every prompt, uploaded file and response is stored with your account, along with your name, email, IP address, device and usage patterns. On personal plans, conversations can be used to train OpenAI’s models unless you opt out, and some are read by human reviewers. Rating a response with a thumbs up or down can also send that conversation for review.
Which plan you use matters more than any other setting. According to OpenAI’s enterprise privacy commitments, business plans are excluded from training by default:
Plan | Trains on your data by default? | Business controls |
Free, Plus, Pro (personal) | Yes, unless you opt out | None |
Business (formerly Team) | No | Admin console, single sign-on |
Enterprise and Edu | No | Admin-set retention, advanced controls |
API | No | Optional zero data retention |
“Deleted” doesn’t mean gone right away. OpenAI normally removes deleted chats and Temporary Chats within 30 days. Courts can override that: in the New York Times copyright lawsuit, OpenAI was ordered to preserve consumer chats from parts of 2025, including deleted ones.
Our rule of thumb: assume anything typed into a personal account could be stored, reviewed or subpoenaed.
The “Red List”: What Never Goes Into a Public AI Tool
Some information should stay out of public AI tools no matter how your settings are configured:
- Social Security numbers, birth dates and home addresses
- Patient health information
- Payment card and bank account numbers
- Passwords, API keys and access codes
- Source code, financial statements and deal documents
- Client tax return information, which tax preparers can face federal penalties for disclosing
Need AI to analyze a document? Replace names with placeholders like “Client A,” strip out identifying numbers, and paste only what the task requires.
The 7 Biggest Risks, With Real Examples
The red list covers what goes in. The bigger picture includes what AI can reach and how attackers use it.
- Shadow AI. Employees use personal accounts outside IT’s view. In 2023, Samsung engineers pasted confidential source code into ChatGPT, and the company restricted its use.
- Overshared files. A copilot connected to Google Drive or SharePoint can surface anything a user can technically open, including the salary spreadsheet someone shared with “everyone” years ago.
- Unvetted plugins and extensions. Attackers recently compromised dozens of popular browser extensions used by millions of people, letting them read open browser tabs, including AI sessions.
- Stolen logins. Researchers have found more than 225,000 ChatGPT credentials for sale, harvested by malware on personal devices. One stolen login exposes an entire chat history.
- Vendor breaches. In late 2025, a breach at Mixpanel, an analytics vendor OpenAI used, exposed some API users’ names and email addresses.
- AI-powered phishing and fake apps. Scammers now write flawless emails, clone executives’ voices and promote fake “ChatGPT” apps that steal passwords. Learn more about AI-powered social engineering and deepfake scams.
- Unreliable output. ChatGPT can invent facts, suggest unsafe links, and follow hidden instructions planted in documents or web pages (known as prompt injection). Verify before you act.
The 2026 Compliance Picture
ChatGPT can support compliance, but it doesn’t make you compliant. That responsibility stays with your organization. Here’s what matters now:
- EU AI Act. Transparency rules took effect August 2, 2026. The heavier “high-risk” rules, covering AI used in decisions like hiring or credit, were pushed to December 2027.
- The state replaced its original AI Act with a narrower law taking effect January 1, 2027, focused on disclosures when automated tools influence significant decisions.
- New privacy rules on automated decision-making are phasing in.
- Rules already in force. HIPAA, the FTC Safeguards Rule for financial and tax businesses, and federal tax-preparer confidentiality rules all apply to what goes into AI tools.
Before approving any AI vendor, ask for its SOC 2 report, the same kind of independent review we perform for service organizations.
Windes’ 5-Step AI Governance Roadmap
We don’t recommend banning AI. Bans push it underground, where you can’t see it at all. Instead, we help clients govern it in five steps:
- Inventory current use. Survey staff and review network and expense data to find which AI tools are already in play. You can’t govern what you can’t see.
- Approve business-tier tools. Choose plans that don’t train on your data by default, and require single sign-on and multi-factor authentication. Make the approved tool easier to use than the workaround.
- Write an AI acceptable-use policy. Keep it to a page or two: approved tools, the red list, when human review is required, and who to contact when something goes wrong.
- Clean up data permissions first. Before connecting AI to Drive, SharePoint or Slack, find sensitive files and tighten who can open them. This is the step most organizations skip.
- Train, monitor and review. Run short, practical training, watch for new plugins and unusual usage, and revisit the policy at least twice a year. AI tools change monthly; your policy should keep up.
Quick Settings for Anyone Using ChatGPT
Governance starts at the top, but every user can tighten their own account in five minutes:
- Turn off training: Settings > Data Controls > “Improve the model for everyone.”
- Review or turn off Memory under Personalization.
- Use Temporary Chat for anything sensitive.
- Delete old chats and remove shared links you no longer need.
Myth: a VPN keeps ChatGPT private. It hides your location, not what you type.
If Something Goes Wrong
Mistakes happen. A fast, calm response limits the damage:
- Account compromised? Change the password, turn on multi-factor authentication, sign out of all sessions, and review the chat history for exposed data.
- Employee shared sensitive data? Delete the chat, document what was shared and when, and bring in your IT and legal advisors to decide whether notification laws or client contracts apply.
The Windes Verdict
ChatGPT isn’t the risk. Unmanaged AI is. Organizations that govern it gain real productivity; those that ignore it are already exposed, whether they know it or not.
The good news: getting control takes weeks, not years. If your organization is also exploring AI agents that take actions on their own, read our guide to agentic AI security.
Not sure where your organization stands?
Talk to the Windes Technology & Risk team to schedule an AI readiness review, or call our office at 562.435.1191.
Frequently Asked Questions
Does OpenAI sell my data?
OpenAI says it doesn’t sell personal data, but it shares data with service providers and must respond to legal demands.
Can my employer see my ChatGPT use?
On business plans, administrators manage accounts and settings. Personal accounts sit outside company visibility, which is exactly why work belongs in approved tools.
Is ChatGPT HIPAA compliant?
Only certain business offerings, and only under a signed business associate agreement. Never put patient data into a personal account.
Is ChatGPT more private than Copilot or Gemini?
It depends on the plan, not the brand. Business tiers generally exclude your data from training by default; consumer versions vary.
About Windes
Founded in 1926, Windes is a relationship-driven CPA firm providing tax, audit, accounting and advisory services to privately held businesses, nonprofit organizations and high-net-worth individuals, with offices in Long Beach, Los Angeles and Orange County. Windes Technology & Risk helps mid-market organizations with cybersecurity, IT governance, SOC reporting and emerging technology risk. Learn more at windes.com.
