Key takeaways
“Data protection is not a property of a tool. It is a property of the terms, settings and controls around the tool — and those differ between the free account and the contract.”
An employee pastes a client contract into a public chatbot to get a quick summary. The summary is good. The contract has now left your environment, under terms your legal team never reviewed, through an account you do not manage. Nothing dramatic happens, which is exactly why it happens again tomorrow, and the day after, across many people. Shadow AI turns routine productivity into a steady flow of information beyond your control. This guide is for CISOs, data protection leads and business owners who want to understand what that flow means for privacy and how to bring it under control.


Where Your Data Goes After You Press Enter
We have already defined Shadow AI in What Is Shadow AI? A Business Leader's Guide to Unsanctioned AI Tools, so we will start from the moment that matters: the point at which an employee submits something. What happens next depends on facts that the employee almost never checks and that vary from product to product.
Broadly, an AI service receives your text or file, processes it on infrastructure operated by the provider or its sub-processors, and returns a response. Along the way several things may or may not occur, depending on the service, the plan and the settings:
- Storage. The input and the output may be kept in the user's history, and in the provider's systems, for a period set by the provider.
- Human review. Some providers allow staff or contractors to review conversations for safety or quality purposes, under conditions described in their terms.
- Model improvement. On some consumer plans, conversations may be used to improve models unless the user opts out. Business and enterprise plans commonly offer different commitments.
- Onward processing. Providers rely on hosting and tooling partners, so your data may pass through more than one organisation and more than one country.
- Account-level exposure. A personal account sits outside your identity controls. If the credentials are compromised, shared or left logged in on a personal device, the history is exposed with them.
The point is not that any one of these is certain to occur. It is that you do not know which of them applies, and neither does the employee. Privacy depends on terms and settings that your organisation has not evaluated, and cannot enforce.
Why Good People Paste Bad Data
It is tempting to frame this as carelessness. In practice, the mental model most people hold is wrong in an understandable way. A chat window feels like a private conversation or a search box. It does not look like a data transfer to a third party, and nothing in the interface suggests otherwise.
The reasons it keeps happening
- The task is real and the deadline is closer. Summarising a long document, tidying a spreadsheet or drafting a reply are exactly what these tools do best. The sensitive content is simply what the task happens to contain.
- Sensitivity is not obvious in the moment. A table of 'customer issues' may include names, contact details and account references. A meeting transcript may include salary discussions. The person pasting rarely stops to inventory what is inside.
- Training happened once, in general terms. A line in an induction deck saying 'do not share confidential data with external tools' is hard to apply to a tool that nobody described as external.
- No approved alternative exists. If the only compliant option is 'do it by hand', many people will choose speed.
Why the common fixes fall short
Reminders and awareness campaigns help at the margin, but they rely on individual judgement under time pressure. Blocking known domains is easy to work around and does not touch AI features embedded in software you already use. And a blanket prohibition removes the benefit without removing the demand. Privacy outcomes improve when the safe route is also the easy route, and when controls operate at the moment of submission instead of in a policy document.


What Is Actually at Stake
The consequences fall into four groups: legal, contractual, commercial and operational. Note that this section is general commentary, not legal advice, and the position in your own jurisdiction and sector needs to be confirmed with qualified counsel.
Legal and regulatory exposure
Data protection regimes such as India's Digital Personal Data Protection Act and the EU and UK GDPR generally place obligations on organisations that determine why and how personal data is processed. Those obligations typically cover having a lawful basis, limiting use to defined purposes, keeping data secure, controlling who processes it on your behalf and, in some circumstances, restricting transfers across borders. When an employee submits personal data to an unvetted service, it is at least arguable that these obligations are engaged, and that you could not demonstrate you met them. Depending on the facts, that may also raise questions around breach assessment and notification. Sector rules can add further layers.
Contractual and commercial exposure
- Client confidentiality. Many client agreements and non-disclosure agreements restrict disclosure to third parties and sub-processors. Submitting client material to a public AI service may breach those terms, whatever the employee's intentions.
- Trade secrets and IP. Source code, pricing models, product plans and negotiation positions are valuable because they are not widely known. Weakening control over them can complicate your ability to protect them.
- Procurement and audit. Enterprise customers increasingly ask suppliers how they govern AI use. 'We are not sure' is a damaging answer in a security questionnaire.
Operational exposure
Even if no harm ever materialises, you lose the ability to answer basic questions: what was shared, by whom, when, and with which service. Without records, an incident response becomes guesswork, and demonstrating good faith to a regulator or customer becomes harder.
Finding Out What Has Already Left
Before designing controls, establish a factual baseline. The aim is to understand categories and patterns, not to reconstruct every message.
- Map data types to teams. Which departments handle personal data, client confidential material, financial information or source code? Those teams carry the highest exposure if they are using AI tools.
- Establish which services and account types are in use. A work-email account on a business plan is a very different proposition from a personal free account. Identity and application-access records can show the difference.
- Review the terms that apply. For each service found, read the retention, training and sub-processor terms for the relevant plan and region. Record what you learned and what you could not determine.
- Ask what is being pasted. Short, non-punitive conversations with team leads usually reveal the typical use cases and the types of content involved.
- Escalate by sensitivity. Where regulated or contractually restricted data appears to be involved, involve legal and data protection leads promptly, before attempting informal remediation.
Treat the output as a risk register, not a list of offenders. It tells you which use cases to replace first and which controls will matter most.
Building Privacy Controls That Work at the Point of Use
Effective privacy controls for AI share a design principle: they sit close to where the data is submitted, they are simple for users, and they leave a record.
The practical building blocks
- Classify data in plain terms. A few categories — such as public, internal, confidential and regulated — with concrete examples for each team. Link each category to the AI routes it may use.
- Provide approved routes on proper terms. That may mean enterprise-plan assistants with the contractual commitments your counsel is comfortable with, or AI built inside your own cloud. Our AI strategy and advisory work helps organisations choose between these on risk, cost and capability.
- Keep sensitive workloads in your environment. For high-sensitivity tasks, an autonomous AI agent connected to your own systems and deployed on your cloud, using a model accessed under enterprise terms, avoids sending raw records to a consumer service.
- Apply policy at submission. Rules that check a prompt before it goes out, and that can pass, flag or escalate it, are more reliable than guidance alone.
- Isolate and restrict access. Separate data and configuration by business unit or client, and give people only the access their role requires. For a related example of tenant isolation and role-based access, see Making a Multi-Tenant SaaS Platform Enterprise-Ready from the Ground Up, delivered for an enterprise SaaS company in the USA.
- Log and review. Keep an audit record of who submitted what category of request and what the outcome was, and review exceptions on a schedule.
Securing the underlying platform matters just as much as the policy. Cloud managed services with monitoring and access management help ensure those controls continue to function after launch. For sectors holding patient records, see our work with healthcare and life sciences organisations.
Making a Multi-Tenant SaaS Platform Enterprise-Ready from the Ground Up
Privacy controls fail when data and permissions are not separated properly. Prospiretech was the end-to-end development partner for an enterprise SaaS company in the USA that builds AI governance tools for large enterprise customers, responsible for both the product and the infrastructure. The platform is multi-tenant, with each customer's data, users and configuration separated at the database level, and an authentication architecture designed to be Okta SSO/SAML-ready from the start. It was containerised with Docker and released through an automated GitHub Actions pipeline. This is an example of architecture, not a data loss prevention tool, but the isolation and access principles are the same ones your own AI controls will need, whatever platform you build them on.
- New enterprise tenants can be onboarded without infrastructure changes, because isolation is built in
- The platform passed enterprise security review without any architectural changes after delivery
- Enterprise customers can connect their own identity provider without requiring additional engineering
Frequently Asked Questions
Does ChatGPT train on the data our employees enter?
It depends on the product, plan and settings, and these change over time, so check the provider's current terms rather than relying on a general answer. Consumer accounts and business or enterprise plans are commonly governed by different data commitments, and users may be able to change certain settings. The practical problem for Shadow AI is that you do not know which plan or settings individual employees use, so you cannot give assurance either way.
Is pasting personal data into an AI tool a data breach?
Not automatically, and the answer turns on the facts and your jurisdiction. Whether an incident counts as a reportable breach depends on what happened to the data, who received it, the terms involved and the applicable law. That is why it is important to establish the facts and involve legal and data protection advisers promptly. This article is general information, not legal advice, and you should take specific advice for your circumstances.
Are enterprise versions of AI tools automatically compliant?
No. Enterprise plans typically offer stronger commitments on retention, training and security, which is valuable, but compliance is a property of how you use the tool, not only of the tool. You still need a lawful basis, data minimisation, access controls, a processing agreement, appropriate transfer arrangements and a record of use. Treat the enterprise plan as a good foundation and assess it against your own obligations. Revisit that assessment whenever the provider changes its terms.
Can we allow AI use but still protect confidential information?
Yes, by combining several measures. Classify data so people know what can go where, provide approved tools on proper contractual terms, keep the most sensitive workloads inside your own environment, apply checks at the point of submission, and keep audit logs. No single control is sufficient, but together they let employees keep the productivity benefit while narrowing the routes by which confidential data could leave. Review the controls regularly, as tools and terms change.
What should we do first if we suspect sensitive data has already been shared?
Establish the facts before acting. Identify what category of data was involved, which service and account type received it, and under what terms. Involve legal and your data protection lead early, because notification and contractual duties may need to be assessed. Avoid informal clean-up that destroys evidence. In parallel, provide an approved alternative so the same exposure does not continue while you investigate. Keep a clear record of each step.
Keep the productivity. Keep control of the data.
Prospiretech helps organisations replace risky workarounds with governed AI: approved tools, enforced policies at the point of use, and the infrastructure to run them securely. Begin with a complimentary 30-minute consultation.
