Where Your Data Goes When Employees Use ChatGPT
Prompts sent to public AI tools leave your control. This guide maps retention, training use and sub-processors — and what to tell employees this week.
Where does my data go when i use chatgpt is the question every employee asks the day after signing up — and most cannot answer it beyond “the cloud”. For a company, the answer determines whether pasting company data into chatgpt risks is a tolerable convenience or a reportable leak: prompts leave your controlled environment, land on a third party’s servers under their terms, and may be retained, reviewed and used in ways your security policy never contemplated. This guide maps the actual data flow — retention windows, do ai tools train on my prompts across tiers, subprocessors and cross-border access — then gives the controls and the one-week communication plan for is it safe to use chatgpt at work with company data conversations.
The data flow, step by step
When an employee submits a prompt to a public AI service, five things happen:
- Transmission to the provider’s infrastructure. The prompt (and any files or images attached) travels over TLS to servers typically hosted in the United States or the provider’s chosen regions. Your data is now processed under their privacy policy, not your DPA.
- Short-term storage for abuse detection and service operation. Providers keep prompts for a period measured in days to months — historically 30 days for public tiers as a default retention window, with abuse-review queues possibly longer. Some providers offer opt-out from any retention; the setting usually defaults to off unless you switch it.
- Potential human review. Retained conversations may be read by human reviewers for abuse or quality purposes unless excluded. This is the fact that makes pasting a client’s personal data into a public tier indefensible: you have disclosed it to a third party’s reviewers without a lawful basis.
- Training pipeline (tier-dependent). On public consumer tiers, providers may use conversations to improve models unless the user opts out — and the employee’s personal opt-out does not help the company, because the next employee’s prompts still flow. Business and enterprise tiers contractually exclude training on your data by default.
- Subprocessors and analytics. Content delivery, logging and monitoring subprocessors handle fragments of the flow; each is a party your security team has never assessed.
Nothing in this flow is inherently malicious — it is ordinary SaaS behaviour that becomes a compliance problem when the data was never the AI provider’s to receive.
Tier comparison: what actually differs
| Consideration | Public / free tier | Consumer paid tier | Enterprise / API tier |
|---|---|---|---|
| Training on your prompts | Yes by default (opt-out available) | Opt-out default varies by provider | Contractually excluded |
| Retention | Default window (often 30 days), opt-out possible | Same mechanics | Shorter/configurable; API often zero-retention options |
| Human review | Possible for abuse/quality | Possible under same conditions | Minimised by contract |
| Data residency / DPA | Generic privacy policy | Same | DPA, regional hosting, SSO, admin controls |
| Admin visibility | None | Minimal | Central billing, usage logs, domain controls |
The delta between rows one and four is the entire business case for enterprise tiers: not better models — identical models — but contractual and administrative control over where data sits and who can see it.
The real risks beyond retention
Retention is only the headline. The full risk set for chatgpt data privacy for companies:
- Training contamination. Prompts used to improve models may surface, in altered form, in outputs to other users. Proprietary strategy pasted as context becomes part of someone else’s advantage — a form of disclosure that no confidentiality clause contemplates.
- No privilege, no deletion guarantee. Confidentiality relationships (legal privilege, NDA scope) do not extend to a provider you have no contract with. Deletion requests are governed by their policy, not yours.
- Cross-border access. Data stored abroad can be subject to foreign legal processes; your DPA and DPIA must acknowledge this honestly rather than assume it away.
- Context window copying. Employees paste whole documents “just for summarising” — board packs, contracts, customer exports — creating copies outside every DLP control you run.
- Integration creep. Browser plugins, desktop apps and email connectors extend the tool into places your network controls never see. Shadow integrations, not the website, are where most exposure actually starts.
The tier test in practice: three real cases
A sales team pastes a public tender document into a free chatbot to summarise requirements. The document itself is public — the tier test passes at the content level, but the free tier’s retention and review settings still apply to the summary prompt; the same task in the approved enterprise tool passes fully. A support agent pastes a customer complaint containing names and order numbers to draft a reply — restricted tier, public tool: a breach, reportable within the hour under the incident route, because that personal data has now been reviewed by a third party’s systems with no lawful basis. A developer pastes an error log containing internal hostnames into a free tool to interpret it — internal tier, wrong tool; the enterprise version of the same model was one click away and acceptable. Note the pattern: two of three failures were solvable by tool choice, not by forbidding AI. When employees can name the tier and the approved tool for their own examples, the policy is working; when they escalate every borderline case, it is not.
The compliance hooks: why this is not just an IT issue
Three regimes make prompt hygiene a compliance topic, not a preference. GDPR: pasting personal data into a public AI tool is a disclosure to a new controller with no lawful basis, no DPA and no transfer mechanism — the kind of finding that draws both regulator attention and customer-contract breaches. Run a DPIA if any approved tool processes personal data at scale; record the transfer safeguards honestly. AI Act: where your organisation deploys AI, Article 4 literacy means employees understand these data rules, not just acknowledge them; data governance duties bind provider-side systems directly, and deployer-side instructions for use flow down from vendors. Contractual: customer DPAs and NDAs frequently prohibit third-party disclosure outright — one paste of a contract excerpt can breach a clause you signed months earlier. The integration point matters: log the opt-out and tier decisions where your DPIA, AI register and vendor records can all see them, because auditors cross-reference these three documents and notice when they disagree.
Controls that work without slowing anyone down
In descending order of leverage:
- Tier procurement. Move heavy users to an enterprise tier with no-training terms, SSO and admin controls. This single step closes the training and human-review questions for the covered population.
- Tier discipline in policy. The three-tier test (public / internal / restricted) from your acceptable-use policy, with your own examples. Employees self-serve the right answer.
- Technical guardrails. Block unapproved AI domains at the proxy or CASB, allowlist approved ones, and monitor for new unsanctioned tools quarterly — not to punish discovery, but to see the estate.
- DLP sensitivity on paste. Patterns for the obvious restricted classes (customer identifiers, financial account data) trip before submission rather than after.
- Disclosure workflow for incidents. An employee who realises they pasted restricted data must have a blameless route to report within the hour. Speed of disclosure is the only thing that limits blast radius — and it is the metric that tells you the culture is working.
How to check your own position
This week’s thirty-minute audit: enumerate every AI tool your staff can reach (SSO logs, expense claims for AI subscriptions, a one-question departmental survey), mark which are enterprise-tier with no-training terms, and check whether any browser extension or integration grants broader access than you would approve knowingly. Then take the last conversation any employee had with a public AI tool — ask them to show you, blamelessly — and run the tier test against it. Finally, verify your opt-out posture: where your organisation’s prompts are excluded from training by contract rather than by an individual setting, note it; where only personal settings exist, that is your top gap. Answer those five points and you can respond to “is it safe to use ChatGPT at work with company data” with specifics instead of a shrug.
Frequently asked questions
Does ChatGPT train on my prompts?
On public consumer tiers, conversations may be used to improve models unless the user opts out — and that opt-out is personal, not organisational. Business and enterprise tiers contractually exclude customer data from training by default. For companies, the safe answer is tier-dependent: verify which tier your employees actually use, because the model is identical and only the data terms differ.
Where is my data stored when I use an AI tool?
On the provider’s infrastructure, usually in the United States or their contracted regions, under their privacy policy rather than your company’s agreements. Public tiers keep prompts in a short-term retention window (often 30 days by default) for abuse detection, with human review possible; enterprise tiers shorten or configure retention, add data-residency options and formalise the relationship with a DPA.
Is it safe to paste company data into ChatGPT?
It depends entirely on the data and the tier. Public-tier prompts may be retained, reviewed and used for training — never acceptable for personal data, secrets or regulated records. Internal-tier material belongs only in enterprise-tier tools with no-training terms; restricted-tier data stays out of general-purpose AI entirely. The tier test turns “is it safe” from a mood into a rule.
How do companies stop employees leaking data to AI tools?
Layer it: enterprise tiers with contractual no-training and SSO for approved tools, an acceptable-use policy with a concrete data-tier test, blocking of unsanctioned AI domains, DLP rules on paste patterns, quarterly discovery of new tools — and a blameless one-hour incident-reporting route. Blocking alone fails; employees routed around it and you lost visibility too.
What should we tell employees about AI data privacy this week?
Three facts, in this order: what actually happens to their prompts (stored on a third party’s servers, possibly reviewed, possibly used for training on public tiers); which tools are approved and what may be pasted into each; and who to tell within an hour if something restricted went in — with no blame for reporting. Three sentences you can send today beat a policy nobody reads.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.