top of page
Search

AI Is Eating Your Data—Here’s Exactly What Happens When You Paste Confidential Information (NPI) into ChatGPT, Claude, Grok, or Gemini (2026 Edition)

AI is eating your data. Do you know where it's going?
AI is eating your data. Do you know where it's going?

In the rush to boost productivity, many RIAs, accounting firms, legal practices, and healthcare admins have started using generative AI tools for everything from summarizing notes to drafting client communications. But as the Summit Team's recent article highlighted, the moment someone pastes a client spreadsheet, tax return snippet, or financial projection into a consumer-grade AI chat—like free ChatGPT, Claude.ai, Gemini, or Grok—sensitive non-public information (NPI) leaves your controlled environment.


The real question isn't whether these tools are "safe" in a vacuum. It's: What actually happens to that data? And more importantly, does it align with your fiduciary duties, the SEC's Reg S-P safeguards, GLBA requirements, HIPAA obligations, or cyber insurance underwriting standards?


Here's the straight 2026 picture, based on each provider's current published policies (always double-check the latest HIPPA, DPAs (Data Processing Agreement), enterprise addendums, and SOC 2 (System and Organization Controls 2 for CPAs) reports—policies can shift).



Consumer vs. Enterprise: The Critical Divide


Most teams start on consumer/browser interfaces (chatgpt.com, claude.ai, gemini.google.com, grok.x.ai, or personal Microsoft Copilot). These are convenient but risky for NPI.


  • Consumer/Browser Tier (Free or Individual Paid Accounts): Data is typically transmitted to the vendor's cloud, processed for a response, and often retained for service improvement, abuse monitoring, or model training. Retention can be indefinite until deleted (with 30-day windows common even in "temporary" modes). Training use varies: some default on, some opt-in/opt-out.

  • Enterprise/Team/Business Plans (Organizational billing, SSO, admin controls): Almost universally no training on your data by default, shorter/custom retention, customer ownership of inputs/outputs, and stronger contractual protections (DPAs, zero-data-retention options where available).


Pasting NPI into a consumer session—even "private" or "temporary"—creates shadow data that vendors may retain or use, with limited audit trail back to you.



Provider Breakdown (2026 Snapshot)


  • OpenAI (ChatGPT): Consumer defaults to training on your data (opt-out via settings); chats stored indefinitely until deleted. Enterprise/Business/Team: No training by default, admin-controlled retention (including auto-delete), zero-data-retention API options for qualifying orgs, customer owns data. HIPAA BAA available on select plans.


  • Anthropic (Claude): Consumer: Opt-in required for training (if allowed, up to 5 years de-identified retention; opt-out keeps ~30 days). Deleted chats excluded. Enterprise/Work/API/Gov: Separate terms—no training on business data under standard DPA; zero-retention options common.


  • Google (Gemini): Consumer: May use for improvement (opt-out via activity controls); history up to 36 months. Enterprise (Gemini in Google Workspace/Cloud): No training without explicit permission; prompts/responses often not retained post-session; inherits Workspace DLP, data residency, HIPAA/FedRAMP support—data stays in your org boundary.


  • xAI (Grok): Consumer: May use interactions for training/improvement (opt-out available); Private Chat deletes within 30 days. Enterprise/Grok Business/Enterprise: Explicit "no training on it, ever"; advanced audit, Vault for customer-managed encryption keys and isolation.


  • Microsoft (Copilot in M365): Consumer: May use for training (opt-out). Enterprise (M365 Copilot with Entra ID): No training on prompts/responses/Graph data; stays in your tenant; encrypted activity history deletable via Purview; full GDPR/EU Boundary support.


The pattern is clear: Enterprise tiers from the major players provide strong governance—no training, controlled retention, and audit-friendly logging—while consumer tiers introduce ambiguity that regulators and insurers increasingly scrutinize.



One Additional Idea: Building a Simple Compliance Layer for AI Use


Beyond banning consumer-tier use for NPI (step 1), consider mandating a lightweight "AI Compliance Wrapper" policy in your firm's Written Information Security Plan (WISP) or employee handbook. Example elements:


  • Require enterprise-tier accounts only (with signed DPA where possible).

  • Use browser extensions or DLP rules to block pasting of sensitive patterns (e.g., SSNs, account numbers) into non-approved domains.

  • Route all production AI tasks through a centralized, monitored virtual desktop/environment so data never touches unmanaged devices.

  • Conduct annual AI usage audits: Log approved tools, review access logs, and document any incidents.

  • Train staff annually: "If it's client NPI, treat it like emailing unencrypted tax returns—don't do it casually."


This approach turns AI from a governance headache into a controlled productivity booster, satisfying cyber insurers (who now probe AI data flows) and regulators focused on existing rules like Reg S-P (updated safeguards) or GLBA.



Bottom Line for Fellow Professionals Including RIAs Like Me


AI adoption is inevitable and powerful—but unmanaged browser use creates drift: shadow data movement, evidence gaps, endpoint risks, and vendor confusion during incidents. The winners pair innovation with discipline: enterprise agreements, centralized access, documented policies, and predictable accountability.


Productivity without governance is expensive drift. Exposure becomes cost. Let's keep AI accelerating your firm while strengthening— not weakening—your compliance posture.


 
 
 

Comments

bottom of page