A practical answer to the “ai chatbot gdpr” question
Using an AI chatbot on a UK website is not inherently unlawful. No chatbot, model or vendor badge is GDPR compliant in isolation. Compliance comes from how the business defines its purposes, limits the data flow, configures providers and operates the system.
A useful ai chatbot gdpr review has three layers. UK GDPR covers personal-data processing, PECR covers cookies and similar access to a visitor’s device and consumer law can require clear disclosure that customers are dealing with AI.
UK GDPR becomes relevant when messages, account details, identifiers, IP addresses or conversation logs can relate to an identifiable person. That can happen even when the assistant was intended to answer simple FAQs.
Consent is not automatically required for every processing purpose. The business needs to assess a lawful basis for each one, then deal separately with any PECR consent required by the widget.
Map the chatbot’s data flow before choosing a vendor
Start with the system rather than a provider’s compliance page. A typical custom assistant looks like this:
Browser widget → business or ZeroPixel backend → conversation database or vector store → model API → optional CRM or helpdesk
Record what happens at each step:
- The data received, generated and stored
- Which organisation is the controller, processor or sub-processor
- Where processing and hosting take place
- Who can access conversations
- Any restricted international transfers and their safeguards
- When each copy is deleted or anonymised
The business will usually determine why the assistant exists and how its data is used. That makes it responsible for the main controller decisions. Hosting companies, model providers and development partners may process data on its instructions, depending on the arrangement.
For many website assistants, we favour retrieval-augmented generation. The system searches approved business documents, selects passages relevant to the visitor’s question and sends those passages with the current question to the model.
Raw conversations should not be added to the knowledge base by default. Product improvement needs its own defined purpose and controlled dataset.
Retrieval-augmented generation does not remove UK GDPR obligations. It can reduce unnecessary access to customer records, narrow what the model sees and make incorrect source material easier to trace. Our AI development approach uses these architectural controls when building assistants and AI features into client products.
Choose a lawful basis and minimise what the model receives
Answering a requested question, opening a support ticket, measuring usage, improving prompts, capturing a lead and sending marketing are separate purposes. Treating them as one vague activity makes the lawful-basis assessment difficult to defend.
Legitimate interests may suit some requested support interactions where the impact on the visitor is limited and safeguards are documented. Contractual necessity may apply where the processing is objectively needed to provide something under a contract. Consent may suit optional uses where the visitor has a real choice and can withdraw it.
There is no universal answer. Special-category data also requires an Article 9 condition, alongside an Article 6 lawful basis.
Design the interface to discourage unnecessary disclosure. A general website assistant should not request payment details, passwords, health information or similarly sensitive material unless the use case and controls require them.
Controls before a model call can include:
- Field-level allowlists for CRM and account data
- Pattern detection and redaction for known sensitive formats
- Validation of structured fields
- Removal of old conversation turns that are no longer needed
- Separate tools with narrow permissions for support actions
Sending an entire CRM profile because an integration makes it easy is poor data minimisation. Pass the few fields needed for the current task.
AI does not trigger a data protection impact assessment in every case. The ICO says a DPIA is required where innovative technology is combined with relevant high-risk criteria. A low-risk FAQ assistant may fall below that threshold, whilst profiling, sensitive data or decisions with serious effects can change the assessment.
Chatbot cookies and local storage: when PECR consent is needed
Chatbot cookie consent cannot be assessed by searching the browser for conventional cookies alone. Audit local storage, software development kits, analytics scripts, session replay and any other technology that stores information or accesses the visitor’s device.
Under PECR, these technologies require prior UK-GDPR-standard consent unless an exception applies. This is separate from the lawful basis used to process the resulting personal data.
The ICO says storage or access may be strictly necessary when an informed visitor actively requests and uses a chatbot. Advertising, cross-site tracking and optional analytics do not become necessary merely because they ship inside the same widget.
Prevent non-essential tracking from loading before consent. Document which strictly necessary exception you rely on, what technology it covers and why the requested chatbot feature cannot work without it.
What a website chatbot privacy notice should say
A website chatbot privacy policy needs enough detail for visitors to understand the real processing. Avoid burying one vague sentence in a general policy.
Cover the following:
- The identity and contact details of the controller
- Each processing purpose and its lawful basis
- The message, account, device and usage data collected
- The model provider, hosting provider and other recipient categories
- Overseas processing and relevant transfer safeguards
- Retention periods or the criteria used to set them
- Data protection rights and how to exercise them
- How to reach a person or challenge an important outcome
- Whether conversations are used to test or improve the assistant
Pair that detailed chatbot privacy notice with a short message beside the input box. For example:
This is an AI assistant. Messages are processed to answer your question and may be handled by [provider]. Please do not enter payment card details, passwords or health information. Read our [chatbot privacy notice].
The short notice helps at the point of collection. It should link to the fuller information rather than trying to compress every recipient, safeguard and retention rule into the widget.
Identifying the assistant as AI is not a blanket GDPR formula. The CMA says businesses should tell customers they are dealing with AI where concealing it could mislead them or affect their decisions. Clear labelling is also a sensible interface choice because visitors need to know the limits of the interaction.
Check model providers, contracts and ChatGPT settings
Where a provider processes personal data on the business’s behalf, UK GDPR requires a written controller-processor contract with equivalent protection for sub-processors.
Review the provider’s terms and technical configuration for:
- Model training and product-improvement use
- Provider-side log retention and deletion
- Security controls and incident reporting
- Sub-processor changes
- Hosting regions
- Restricted transfer safeguards
- Support for access and deletion requests
The answer to “is ChatGPT GDPR compliant?” depends partly on the plan and configuration, then on the customer’s own purposes and controls.
| OpenAI service | Published model-improvement position |
|---|---|
| Personal ChatGPT Free, Plus and Pro workspaces | Sharing is enabled by default and can be disabled. |
| ChatGPT Business, Enterprise, Edu, Healthcare and Teachers | Organisation data is not used to train or improve models by default. |
| OpenAI API platform | Organisation data is not used to train or improve models by default. |
OpenAI’s business Data Processing Addendum describes it as a processor acting on the customer’s behalf. The customer still has controller responsibilities. It must decide what data enters the service, establish a lawful basis and provide suitable privacy information.
A “not used for training” setting is valuable. It is one control rather than a GDPR certificate. Retention, transfers, access, security and the necessity of sending the data still matter.
A custom website assistant would normally use a managed API or comparable business arrangement. Staff copying customer messages into personal chatbot accounts creates a less controlled data flow and can expose information to settings the business has never reviewed.
Set retention, support rights and keep consequential decisions human
UK GDPR provides no universal retention period for chatbot conversations. Set periods according to purpose and delete or anonymise information when it is no longer needed.
Conversation history, model-provider logs, support tickets and security logs can have different periods. Record each one rather than applying an indefinite default to the whole system.
Build a process for access, correction, erasure and objection requests. A conversation reference, account identifier or carefully chosen verification step may help locate records without collecting excessive new identity evidence.
Give visitors an obvious route to a person when the assistant is uncertain, an answer is disputed or the interaction could affect a refund, application or eligibility decision.
Law position from 5 February 2026: Articles 22A to 22D use the test that a decision is solely automated where there is no meaningful human involvement in taking it. A reviewer needs authority, relevant information and enough time to change the outcome. A rubber stamp does not provide meaningful involvement.
The ICO’s principal AI guidance remains useful, although it is under review following the Data (Use and Access) Act 2025. Older summaries of automated decision-making rules may therefore be stale.
The pre-launch ai chatbot gdpr checklist
Before releasing the widget:
- Write down every purpose, necessary data field, lawful basis, provider and sub-processor
- Record storage locations, international transfers and safeguards
- Audit cookies, local storage, scripts and other device access under PECR
- Publish the full privacy information and add a just-in-time notice beside the widget
- Sign the required processor terms and review sub-processor arrangements
- Decide whether a DPIA is required and document the reasoning
- Configure minimum retention for each data category
- Test access, correction, deletion and objection workflows
- Restrict transcript access by staff role
- Test redaction, prompt-injection resistance and failure handling
- Provide human escalation and monitor unanswered or disputed interactions
- Confirm the assistant cannot approve applications, promise refunds or take other consequential actions outside its authority
Prompts are behavioural guidance, not an authorisation boundary. If an assistant must never approve a refund or application, remove that permission from its tools and APIs.
This is general technical information, not legal advice. Use the ICO guidance linked above and seek appropriate legal advice for your organisation’s processing.
If you are planning a custom assistant and want to map the technical data flow before development starts, discuss the build with us.