What Actually Happens to Your Data
| Stage | What happens | Practical risk |
|---|---|---|
| Transit | Prompt travels encrypted to the provider's datacenter | Low for most — TLS is standard |
| Processing | Provider decrypts and runs inference — their model, their hardware, their operators | Moderate-high — the provider sees it now |
| Retention | Logs kept for abuse monitoring, improvement, or compliance | High — windows of 30 days to years, tier-dependent |
| Training | Prompt may be used to train future models unless excluded by your tier | High — irreversible and effectively unremovable |
| Subcontractors | Anthropic uses AWS/Azure; OpenAI uses Azure; each adds a party | High — your data chain grows without your consent |
Nothing in this table is an accusation — it's how hosted inference works. The point is to be clear-eyed: hosted = the provider can access it. Whether they will is a contractual question; whether they can is physics.
Consumer Tiers vs Enterprise Tiers: Read the Contract
| Consumer / free tier | Enterprise / API (no-training) | |
|---|---|---|
| Training on your data | Historically allowed by default (toggles vary) | Contractually excluded |
| Retention | Longer, less transparent | Shorter, defined in the DPA |
| Human review | Possible for quality/abuse samples | Excluded by contract (mostly) |
| Data residency | Provider-chosen regions | Negotiable regions |
| Audit trail | None | Logs, SOC reports, sometimes DPA audits |
The famous 2025 backlash over a major provider's policy flip — quietly adding training-on-user-data back to consumer terms — is the reminder that these defaults change without notice and without a migration path for your history. Enterprise tiers are more stable, but the contract, not the checkbox, is what protects you.
The Business-Specific Risks
- Regulatory exposure. Pasting client data into consumer AI can violate professional obligations — attorney-client privilege, medical confidentiality, GDPR/financial duties. The breach isn't hypothetical; it's the act of sending the data. See Offline AI for Regulated Industries.
- Trade secrets. A prompt describing your proprietary process is a disclosure, full stop. Training clauses make it permanent.
- Cross-border transfer. Your data may process in jurisdictions you didn't choose, complicating GDPR compliance and any residency requirements.
- Breach surface. Every provider you use adds a third party that can be breached. 2024–2026 saw high-profile breaches of AI-adjacent infrastructure — the risk is real, not theoretical.
- Unproducible audit trail. When a regulator asks "what data went where?", a consumer tool gives you nothing to show.
The Honest Trade-Off
Cloud models remain ahead on capability — frontier models beat anything self-hostable on hard reasoning and long context. That is real, and this guide won't pretend otherwise. The decision framework that actually works:
| Data sensitivity | Recommendation |
|---|---|
| Public or low-sensitivity | Cloud is fine — capability wins |
| Personal but not regulated | Cloud with enterprise/no-training tier, or local |
| Client, regulated, or proprietary | Local AI — see the honest private-LLM guide |
| Classified / air-gapped requirement | Local, offline — see Air-Gapped AI |
💡 The hybrid that most teams end up on: local models for anything sensitive (drafting contracts, analyzing patient data, reviewing HR files) and cloud frontier models for public or synthetic work. Lawyer Assistant is exactly this thesis: legal-grade answers with zero data leaving the machine.
If You Must Use Cloud AI: The Checklist
- ✅ Use the enterprise/API tier with a signed no-training clause for business data
- ✅ Disable chat-history saving on consumer tools before pasting anything real
- ✅ Assume retention happens; never paste what you couldn't survive being published
- ✅ Redact names, IDs, and client identifiers before sending
- ✅ Choose data-residency regions that match your compliance needs
- ✅ Document which tool processed which document — your future audit self will thank you
Frequently Asked Questions (FAQ)
Do cloud AI providers train on my prompts?
It depends on the service tier you choose. Consumer tiers historically allowed training on prompts by default; enterprise and API tiers with "no training" commitments generally do not. The obligation is contractual — and the practical point is that your data leaves your control either way, so the risk assessment must assume exposure.
How long do cloud AI providers keep my prompts?
Retention varies by provider and tier: from 30 days for some API traffic to months or years on consumer tiers, with exact windows changing as policies evolve. Encrypted-at-rest storage does not protect against a provider with the decryption keys — retention is about access, not just storage.
Is it safe to paste client data into ChatGPT or Claude?
For strictly personal, non-sensitive text, consumer tools are generally acceptable. For anything regulated — client legal work, patient data, financial records, employee PII — pasting into consumer chat is a genuine compliance and liability risk, regardless of the "don't train" toggles. That is the use case for local AI.
What are the biggest risks of using cloud AI for business data?
Data exposure to the provider and its subcontractors, retention beyond your control, training-data clauses, cross-border transfers, insider or breach access, and audit trails that you cannot produce. Each is manageable contractually for enterprise tiers; none disappear.
Can cloud AI providers see my data if it's encrypted?
Encryption in transit and at rest protects against external attackers, not the provider itself — the provider holds the keys needed to process your data. The newer confidential-computing options (TEEs) narrow this gap but require the provider to cooperate; the only unconditional answer is not sending the data at all.
When does local AI make sense over cloud AI?
When the data is sensitive enough that any off-machine exposure is unacceptable (legal, medical, financial, HR), when compliance demands data residency or auditability, when you need reliability without network dependence, or when per-token costs at scale beat a one-time hardware purchase.