Personal vs Work AI Account: What Free, Plus, Business & Enterprise Plans Keep and Train On

Run a signed-out US Google search for "is it safe to use ChatGPT at work" or "is ChatGPT Plus safe for confidential information", as our desk did on October 29, 2026 (Google US SERP), and page one skips the single table an office worker needs. Security vendors pitch monitoring sensors, OpenAI summarizes general habits, and a 2023 column cites old menus. Even IBM's Shadow AI explainer treats chatbot use as one single risk without separating a free Gmail login from a business workspace. Millions of workers pay $20 a month for ChatGPT Plus, Claude Pro, or Google AI Pro assuming a paid receipt buys business privacy. Across all four major AI vendors in October 2026, it does not.

Two separate annual breach studies quantify the stakes and must stay separate. In IBM's 2025 Cost of a Data Breach study (n = 600 breached organizations studied by Ponemon, IBM Newsroom, Jul 30, 2025), 20% of surveyed organizations reported a shadow-AI breach, correlating with $670,000 in higher average breach costs in that observational sample. In IBM's separate 2026 Cost of a Data Breach study (n = 602 breached organizations, Help Net Security, Jul 30, 2026; IBM Newsroom, Jul 29, 2026), unapproved AI tools figured in 43% of security incidents across the 602 organizations, with about half causing data loss and roughly 20% of those incidents triggering a fine.

Part of our program hub on AI safety at work at ed.cafe (Boston 501(c)(3) No Worker Left Behind; general education, not legal advice), this guide compares Free, $20 Plus/Pro, $25-to-$30 Business/Team, and Enterprise plans across OpenAI, Anthropic, Google, and Microsoft, closing with The ed.cafe Personal-vs-Work AI Account Privacy Matrix & 60-Second Settings Card.

Why paying $20 a month for Plus or Pro does not buy business confidentiality

Paying $20 a month for ChatGPT Plus, Claude Pro, Google AI Pro ($19.99/month), or personal Copilot buys higher rate limits while keeping you on consumer Terms of Service. None of those $20 plans include a commercial Data Processing Addendum (DPA), workspace isolation, or default exclusion from model training.

Every AI vendor divides accounts into two legal tracks:

1. Consumer Terms (Free, Plus, Pro, Max): Built for one individual. You grant rights to train on prompts and uploads unless you flip an opt-out toggle. Even after opting out, abuse logs retain flagged or deleted chats for 30 days or more, with zero enterprise breach-notice obligations.
2. Commercial Terms & DPA (Team, Business, Workspace, Enterprise): Built for an organization as data controller. A binding DPA blocks public model training, isolates tenant storage, and gives admins seat and retention controls.

Picture testing a secret bakery recipe. Using a Free or $20 Plus account is renting a counter inside a public culinary school: $20 buys a faster oven, yet instructors copy recipes left out unless you file an opt-out card. Leasing a $25-to-$30/seat Business workspace is renting a locked commercial commissary under an NDA: the landlord never copies your recipes, though your shop manager holds a master key.

Where that kitchen analogy breaks is retention speed. Grabbing a paper recipe leaves zero trace, whereas pasting a client contract into a consumer AI tab logs text immediately for 30 days (opted out) up to three years at Google or five years at Anthropic (opted in). In Harmonic Security's Q3 2025 customer telemetry report (Business Wire, Nov 13, 2025), 26.4% of employee file uploads in that report carried sensitive data, and 12% of sensitive exposures came from personal free accounts. Pair this comparison with what not to paste into an AI tool at work.

What Free, $20 Plus, Business, and Enterprise plans keep and train on across four vendors

Across OpenAI, Anthropic, Google, and Microsoft in October 2026, every Free and $20 consumer plan enables model training by default, whereas every commercial Team, Business, Workspace, and Enterprise plan disables public model training out of the box. Server retention spans from 30 days to five years depending on the tier:

Watch three traps in those vendor terms:

• Anthropic's five-year consumer split: Personal Free, Pro, and Max users who leave Help improve Claude active allow up to five years of de-identified training retention, versus 30 days after deletion when turned off (Anthropic Privacy Center). Flipping it off protects only future chats.
• Google Gemini's three-year human review: Google's Gemini Apps Privacy Hub warns @gmail.com users not to enter confidential text you would not want a reviewer to read; reviewed samples stay up to three years even after deletion. Paid Google Workspace blocks outside training and human review.
• Microsoft's shared Copilot branding: @outlook.com logins use consumer terms, while corporate Entra ID logins enable Enterprise Data Protection. See Is Microsoft Copilot safe for confidential work?.

Who can read your prompts: vendor safety reviewers versus your own employer's admins

Moving from a personal AI account to a work AI account swaps external vendor exposure for internal employer visibility. Consumer accounts hide your cloud sidebar from managers while feeding vendor pipelines, whereas Enterprise accounts lock out vendor trainers while giving company compliance staff transcript export tools.

In KPMG and the University of Melbourne's global study (n > 48,000 respondents across 47 countries, KPMG, Apr 2025), 58% of surveyed workers used AI at work and 57% of those AI users hid that use. Yet typing inside a personal Gmail account on a company laptop still exposes prompts to corporate TLS decryption or endpoint DLP (Can my employer see my ChatGPT history, Copilot, or Claude chats?).

Admin visibility also differs between $25-to-$30 Team/Business and Enterprise:

• Self-serve Team/Business (ChatGPT Business, Claude Team): Admins manage seats and billing without a message-body Compliance API or private-sidebar reader, though the employer owns the workspace.
• The @company.com domain-claim trap: Registering a personal Free or $20 Plus account with [email protected] lets your employer claim or force-migrate that account upon verifying the domain for Enterprise. Use personal email only.
• Enterprise (ChatGPT Enterprise, Claude Enterprise, M365 Copilot, Google Workspace): Compliance admins export full prompts and replies via Compliance APIs, Purview DSPM/eDiscovery, or Google Vault even after sidebar deletion.

Why regulated work hits a wall on $25 Business tiers: HIPAA BAAs, client privilege, and DPAs

Buying $25/seat ChatGPT Business or Claude Team disables model training, yet pasting Protected Health Information (PHI) there still violates HIPAA because neither OpenAI nor Anthropic signs a Business Associate Agreement (BAA) on self-serve Team/Business plans. Small clinics and social-service teams under 50 staff get low-seat BAA coverage through Google Workspace or Microsoft 365:

Test a 14-person operational example (assumptions stated). Suppose a fictional 14-person Worcester, Massachusetts physical therapy clinic pays $14/user/month for Google Workspace Business Standard ($2,352/year) and six staff draft insurer appeals and patient instructions:

1. Six reimbursed $20 Plus/Pro cards ($1,440/year): Violates HIPAA (zero BAA), leaves training on by default, and strands patient data in personal accounts at turnover.
2. Six ChatGPT Business or Claude Team seats at $25/seat/month ($1,800/year): Safe for de-identified SOPs under a DPA, yet still unlawful for identifiable patient charts without a BAA.
3. Accepting Google's Workspace BAA on existing Business Standard seats ($0 added fee): Admin accepts the BAA in Admin Console -> Account settings -> Legal and compliance, blocks unvetted add-ons, and keeps PHI inside Workspace Gemini. (On Microsoft 365 Business Premium, Entra Copilot Chat carries Microsoft's automatic DPA/BAA at $0 extra; full M365 Copilot adds $30/seat/month after a 3-hour SharePoint permissions cleanup.) Nonprofits handling 42 CFR Part 2 or VAWA records should check our case manager AI notes and BAA guide.

How to lock down a personal AI account in 60 seconds when your job provides zero tools

When your employer supplies zero AI tools and you use a personal account strictly for public-text rewriting, Excel formulas, or career prep, disable model training across all four apps today (verified October 2026; never substitutes for a DPA):

1. ChatGPT (Free, Plus, Pro): Settings -> Data controls -> Improve the model for everyone -> Off -> Done. Delete old public links under Shared links -> Manage.
2. Claude (Free, Pro, Max): Settings -> Privacy -> Help improve Claude -> Off (cuts post-deletion retention from five years to 30 days).
3. Google Gemini Consumer (gemini.google.com): Activity -> Turn off -> Turn off and delete activity (stops human review and three-year snippet storage, leaving up to a 72-hour buffer).
4. Microsoft Copilot Personal (copilot.microsoft.com): Avatar -> Privacy -> Off for Model training on text and Model training on voice.

Need an approved work tier? Share our guide on picking one approved AI tool. Learners forming Circle #1 at ed.cafe (our free 30-chair Massachusetts cohort closing December 1, 2026 at ed.cafe/start/) practice these checks on synthetic files first.

The ed.cafe Personal-vs-Work AI Account Privacy Matrix and 60-Second Settings Card

In the ed.cafe safety workflow, staff route every task by matching data sensitivity to the contract behind the login. Print or bookmark The ed.cafe Personal-vs-Work AI Account Privacy Matrix & 60-Second Settings Card for onboarding.

THE ED.CAFE PERSONAL-VS-WORK AI ACCOUNT PRIVACY MATRIX & 60-SECOND SETTINGS CARD
Version: Oct 2026 | Neutral Reference Card (General Education, Not Legal Advice)

1. THREE-QUESTION ACCOUNT ROUTING FLOW
Q1. Does the input hold PHI (HIPAA), FERPA, 42 CFR Part 2, VAWA, export data, or API keys?
    -> YES: Never paste keys/passwords anywhere. Regulated data is BANNED on Free,
            $20 Plus/Pro, and $25 ChatGPT Business / Claude Team plans. Use only an
            authorized Enterprise or BAA-configured Workspace / M365 tenant.
    -> NO:  Go to Q2.

Q2. Does the input hold non-public work data (clients, pricing, code, contracts, memos)?
    -> YES: Approved Work Tier ONLY (ChatGPT Business/Enterprise, Claude Team/Enterprise,
            Workspace Gemini, Entra Copilot Chat, or M365 Copilot). NEVER paste into
            Free or $20 Plus/Pro, even with training turned off.
    -> NO:  Go to Q3.

Q3. Is the prompt about YOUR personal pay, job search, union rights, or health?
    -> YES: Opted-out Personal Account + Personal Device + Personal 5G / Home Wi-Fi.
            Never register personal accounts with an @company.com address.
    -> NO:  Public learning or generic syntax tasks fit either opted-out account.

2. TIER SUMMARY (OCTOBER 2026 DEFAULTS)
TIER                   | PRICE        | TRAIN? | BAA?         | ADMIN PROMPT LOGS?
Free / $20 Plus & Pro  | $0-$200/mo   | ON     | No           | No (check laptop)
ChatGPT Business/Team  | $25-$30/u/mo | OFF    | No           | Seat/billing only*
Claude Team            | $25-$30/u/mo | OFF    | No           | Shared proj/seat*
Workspace Gemini (Biz) | In Workspace | OFF    | Yes (accept) | Yes (Google Vault)
Entra Copilot Chat     | $0 w/ M365   | OFF    | Yes (DPA)    | Audit metadata
M365 Copilot (Paid)    | +$30/u/mo    | OFF    | Yes (DPA)    | Yes (Purview/eDisc)
ChatGPT / Claude Ent.  | Custom sales | OFF    | Yes          | Yes (Compliance API)
*Employer owns workspace & can inspect traffic on company-managed devices.

3. 60-SECOND PERSONAL OPT-OUT CHECKLIST
[ ] ChatGPT: Settings -> Data controls -> "Improve the model for everyone" -> OFF
[ ] Claude:  Settings -> Privacy -> "Help improve Claude" -> OFF (5 yrs -> 30 days)
[ ] Gemini:  Activity -> Turn off -> "Turn off and delete activity" (no 3-yr review)
[ ] Copilot: Avatar -> Privacy -> "Model training on text" & "voice" -> OFF
[ ] Email:   Confirm personal AI logins use personal email, never @company.com

Where this stops being true

Vendor terms shift quarterly, and four conditions override the defaults in The ed.cafe Personal-vs-Work AI Account Privacy Matrix:

1. Managed laptops override personal cloud privacy: Typing into an opted-out personal Plus or Pro account on an employer laptop exposes text to corporate TLS proxies and endpoint DLP.
2. Third-party connectors leave your DPA: External plugins, custom GPT actions, or unvetted MCP servers inside Business or Enterprise accounts transmit prompts under third-party policies.
3. Legal holds and safety flags apply everywhere: Subpoenas, preservation orders, and severe-harm classifiers override standard deletion timers on every tier.
4. Sovereign and clinical rules set higher floors: ITAR/EAR export data (requiring FedRAMP High or GCC High), 42 CFR Part 2 substance-use notes, GDPR residency, and legal privilege require specialist review. ed.cafe provides workforce safety education, not legal or compliance representation.

Questions people ask

Key terms

• Data Processing Addendum (DPA): A binding commercial contract governing how a vendor processes, secures, audits, and deletes organizational data.
• Business Associate Agreement (BAA): A mandatory US HIPAA contract requiring a vendor to safeguard Protected Health Information (PHI) before patient data is shared.
• Model training opt-out: A privacy toggle instructing an AI vendor not to feed future prompts or uploads into foundation-model updates.
• Zero Data Retention (ZDR): An enterprise or API mode where prompts process in memory without writing message text to persistent disk logs.
• Enterprise Data Protection (EDP): Microsoft's commercial baseline keeping Entra Copilot Chat and M365 Copilot work prompts inside the tenant boundary.
• Domain verification claim: An admin DNS check bringing cloud accounts registered under @company.com under corporate workspace governance.

Sources

1. IBM & Ponemon, "Cost of a Data Breach 2025: AI Access Controls" (Jul 30, 2025): IBM Newsroom 2025.
2. IBM & Help Net Security, "Cost of a Data Breach 2026" (Jul 29-30, 2026): IBM 2026 and Help Net Security.
3. Harmonic Security, "26% of Uploads to GenAI Contain Sensitive Data" (Nov 13, 2025): Business Wire.
4. KPMG & Univ. of Melbourne, "Trust, Attitudes and Use of AI 2025" (Apr 2025): KPMG Study.
5. OpenAI, "Data Controls FAQ" & "Enterprise Privacy" (Oct 2026): Data Controls and Enterprise Privacy.
6. Anthropic, "Privacy Center" & "Claude Enterprise" (Oct 2026): Privacy and Enterprise.
7. Google, "Gemini Privacy," "Workspace Privacy," & "Workspace HIPAA" (Oct 2026): Gemini, Workspace, and HIPAA.
8. Microsoft, "Copilot Privacy FAQ" & "Enterprise Data Protection" (Oct 2026): Consumer FAQ and Learn EDP.
9. IBM Think, "What Is Shadow AI?" (Oct 2024): IBM Shadow AI.