AI Policy
Last updated:
This policy governs the design, development, deployment, and governance of AI systems to ensure they are safe, ethical, secure, and compliant with the EU AI Act and other relevant legislation.If you have any questions regarding our AI policy, please contact us.
Definitions
AI (Artificial Intelligence): Machine-based systems that perform tasks typically requiring human intelligence, including natural language processing, data summarisation, insight generation, or predictive modelling.
Inference: The processing of data using a pre-trained AI model to generate outputs without modifying the model.
RAG (Retrieval-Augmented Generation): An architecture that combines pre-trained AI with retrieval of external data sources to inform responses.
Customer Data: Any data submitted, uploaded, or processed by the customer or participants within the Ballpark platform.
Scope
This policy applies to:
All AI features developed or integrated into Ballpark products and internal systems.
All staff involved in the lifecycle of AI models, including design, testing, deployment, and monitoring.
Third-party AI systems integrated with Ballpark services.
Core Restrictions
Inference-Only Operations
Ballpark restricts all AI system usage to inference operations only. This means:
AI systems may only be used for inference: processing data and generating outputs from pre-trained models
Prohibited: Using any customer data, user-generated content, or customer-specific business data for training, fine-tuning, retraining, model evaluation, or performance testing
Retrieval-Augmented Generation (RAG) Systems
Ballpark may use RAG systems to enhance AI capabilities with internal knowledge and expertise.
Permitted content:
Ballpark’s internal documentation and procedures
Team member professional knowledge and expertise
Industry best practices and publicly available technical documentation
Ballpark’s proprietary methodologies and frameworks
Prohibited content:
Customer data
Customer-generated content
Customer-specific business information
All RAG knowledge bases must be reviewed and approved by the CTO before implementation.
Data Protection
Customer Data Isolation
Customer data must remain completely isolated from AI training pipelines and RAG systems
Technical controls required to prevent inadvertent customer data usage
All customer data processing must be transient with no storage for improvement purposes
Lawful basis, transparency and participant choices
Ballpark and its authorised service providers may use AI for disclosed study-related activities, including interviews, transcription, translation, analysis, summaries, research reports and response-quality checks. Only the purposes and input data described for the relevant activity may be processed. Participants must be told when they interact with an AI interviewer.
Each use of personal data must have an appropriate lawful basis identified and documented by the responsible controller before processing begins. Where Ballpark acts as a processor, it processes data on the customer’s documented instructions within the applicable agreement. Where Ballpark conducts its own research as controller, it is responsible for that assessment. Customer authorisation and acceptance of Participant Terms do not, by themselves, establish a lawful basis.
A separate AI opt-in is not required solely because a study uses AI. Disclosed AI processing may rely on another valid lawful basis, including legitimate interests where the required purpose, necessity and balancing assessment supports that use. The relevant privacy information must explain the actual basis, purposes, data types, recipients, retention and applicable rights. Merely including AI in a contract does not make the processing objectively necessary for that contract.
Where consent is relied on, or is required by applicable law, an applicable agreement or a binding commitment to participants, it must be freely given, specific, informed and unambiguous, and explicit where required. It must be obtained through a clear affirmative action before the relevant processing. An appropriately presented opt-in button or other valid method may be used; a checkbox is not mandatory. General acceptance of Participant Terms is not treated as AI consent. Distinct optional purposes must have distinct choices where required.
Participants must have accessible ways to exercise their rights. Where processing relies on consent, withdrawal must be as easy as giving consent and must stop future processing that depends on it. Where processing relies on legitimate interests, participants must be informed of their right to object on grounds relating to their particular situation; objections must be assessed and processing restricted where required by law. Stopping participation does not waive rights concerning information already collected.
Record the lawful-basis assessment, participant notices and contractual acceptance, together with any required consents and subsequent requests. Processing controls must honour applicable restrictions across collection, analysis, reporting and later processing. New purposes, materially expanded input data or new uses require a fresh assessment, updated information and new consent where required before they begin. A broad reference to possible AI use is not permission for undisclosed future processing.
This approach does not remove separate requirements for recording, sensitive information, children, international transfers or other regulated activities. It does not relax Ballpark’s restrictions on using Study Data for training, fine-tuning, retraining, model evaluation or performance testing. Existing consents and contractual or participant commitments remain binding unless lawfully changed; this policy update does not retrospectively authorise new uses of previously collected data or allow withdrawal to be bypassed.
Compliance Framework
Risk classification
All AI systems must be assessed and classified according to the EU AI Act’s risk-based framework:
High-risk systems: Require pre-deployment assessments, documentation, and registration (e.g. recruitment tools, critical infrastructure decisions)
Limited/minimal-risk systems: Implement proportionate safeguards with transparency and user control
Ballpark will maintain documentation of all AI use cases, classification outcomes, and applicable controls.
Human Oversight
All high-risk AI systems must include documented human-in-the-loop or human-on-the-loop mechanisms to allow for oversight, correction, or override of automated decisions. Responsibility for each AI system must be assigned to an accountable individual or team.
Transparency Requirements
Users must be notified when interacting with AI systems
Clear communication that data is used for inference only, never for training
Explanation of AI decision-making logic and limitations, especially for high-risk cases
Third-Party AI Services
Contracts must explicitly prohibit using Ballpark/customer data for training or improvement
Service agreements must guarantee inference-only usage
Regular compliance audits required
Participant Study Data and authorised providers
Authorised AI providers may process participant Study Data only for the instructed, disclosed research functions under appropriate contractual and security safeguards. They must not use that data for their own model training or improvement. Ballpark’s existing prohibitions on training, fine-tuning, retraining, model evaluation and performance testing apply equally to participant responses, recordings, transcripts and other customer data.
Evaluating the quality of a research response is a different activity from evaluating an AI model. Any AI-assisted response-quality assessment must nevertheless be specifically disclosed and covered by the required permissions. An automated flag or AI-detection score is not proof of fraud; disputed adverse participation or incentive decisions must receive appropriate human review. Ballpark’s zero-tolerance rule for prohibited participant use of AI is distinct from the disclosed and lawfully authorised use of AI to conduct or analyse a Study. Ownership of study intellectual property does not override these AI restrictions or the applicable participant permissions.
The Participant Terms do not authorise biometric identification, biometric-based emotion recognition or consequential decisions about an individual. New use cases outside the approved research scope require separate legal, security and product review before deployment.
Quality Assurance
Fairness and Security
Regular bias testing for discriminatory impacts
Pre-deployment security testing and ongoing performance evaluation
Protection against adversarial attacks and model drift
Fallback mechanisms for service degradation
Ethical Impact Assessments
Required for high-risk systems, including:
Societal and human rights impact assessment
Potential harm identification and mitigation strategies
Confirmation of customer data protection compliance
Annual review and registration in AI systems register
Incident Management and Redress
Users must have access to a channel to report:
Negative effects or harms from AI usage
Errors or inappropriate decisions made by AI systems
Reports will be reviewed under Ballpark’s incident response framework, and affected parties will be informed of outcomes and potential redress options.
Sustainability and Social Benefit
Teams are encouraged to evaluate the environmental impact (e.g. energy consumption of model training) and consider more efficient alternatives.
Priority will be given to projects where AI applications contribute to social good, accessibility, or public benefit.
Monitoring and Policy Review
AI systems must be regularly audited to ensure:
Accuracy and reliability
Ongoing compliance with evolving legal standards
Annual policy review or upon significant legislative changes
Ethical performance
This policy will be reviewed annually or upon significant changes in legislation or AI practices.
Roles and Responsibilities
Role | Responsibility |
|---|---|
CTO | Policy owner and final authority on AI governance decisions |
Engineering Team Leads | Classification, documentation, and testing of AI systems |
Security & Compliance | Legal risk assessment, GDPR/Data Privacy alignment, and ethical impact review |
Support & Customer Ops. | Communicating AI usage and managing incident reporting and redress |
Exceptions
Any deviations from this policy must be approved by the CTO in writing, with a supporting risk justification and mitigation plan.
Internal exception approval cannot override applicable law, a signed customer agreement or a commitment made to participants. Any proposed change affecting participant purposes, recipients, retention or permitted uses must be assessed before deployment and communicated or separately consented to where required. No internal exception creates permission to use Study Data for a purpose prohibited by the applicable participant notice or agreement.
Related documents: Participant Terms · Privacy Policy · AI Policy · Terms / MSA.