
Estimated reading time: 4-5.5 min.
A UK-based charity supporting individuals recently released from custody or probation aimed to create an AI chatbot to serve as the first point of contact for its users. The chatbot’s role was to offer immediate support and triage referrals to human caseworkers. Given the sensitivity of the user base and the type of information potentially disclosed, such as mental health, housing status or criminal history, the project required strong guarantees around privacy, transparency and compliance with UK GDPR, alongside adherence to Ethical Technology principles.
Requirements
The charity needed a solution that was:
- Available 24/7, offering immediate, consistent support
- Accessible, especially for users with limited digital literacy or high levels of stress
- Emotionally neutral, avoiding parasocial bonds or addictive engagement
- Integrable with existing internal systems
- Low maintenance, both technically and environmentally
- Cost-effective and compliant from day one
Options Considered
Two tools were initially explored:
- Pi (https://heypi.com/): an empathetic chatbot known for its conversational tone, which can improve accessibility for stressed or anxious users. However, Pi does not offer an API, which makes it unsuitable for integration with other digital tools used by the charity.
- ChatGPT (https://openai.com/chatgpt): known for high-quality language generation and ease of implementation.
Is ChatGPT Appropriate for a Rights-Based, Sensitive Context?
⚖️ Privacy and Data Protection
- ChatGPT is a cloud-based model hosted by OpenAI in the United States. This involves international data transfers, subject to UK GDPR’s transfer restrictions.
- To use it legally, the charity would need to conduct both a Data Protection Impact Assessment (DPIA) and a Transfer Risk Assessment (TRA) to assess how user data is processed and if additional safeguards are needed.
- As the charity deals with vulnerable individuals and potentially special category data (e.g. health, criminal history), any risk of misuse, re-use or surveillance must be taken seriously.
- There is limited transparency over what happens to data submitted via ChatGPT, particularly in free versions. Although OpenAI claims not to train models on API inputs by default, concerns remain about how long data is retained and who can access it.
🔒 Cybersecurity and Control
- OpenAI’s infrastructure is likely secure, but the charity cannot independently verify or configure the protections in place.
- A self-hosted alternative offers more control over access policies, encryption and user data governance.
🌍 Environmental Impact
- Large models like ChatGPT are computationally intensive, relying on energy-hungry cloud infrastructure, often outside the UK or EU.
- For a chatbot whose primary tasks involve simple triage and referral, such a large model may be overkill, increasing environmental impact and costs without improving service.
Are there any alternatives to ChatGPT?
The answer is yes. For example, Rasa and Botpress are two open-source frameworks that are more aligned with Ethical Technology principles.
Rasa (https://rasa.com) is an open-source conversational AI framework that allows organisations to build highly customisable chatbots. One of its main strengths is that it can be fully self-hosted, which means no user data needs to leave the organisation’s servers. This makes Rasa particularly well-suited for organisations that handle sensitive information, as it allows for complete control over data privacy and security. Its open-source nature also provides transparency, a strong community and the flexibility to tailor every aspect of the system. With Rasa, it is possible to meet UK GDPR obligations, including implementing detailed data retention policies and pseudonymisation strategies. However, deploying Rasa typically requires moderate to high levels of technical expertise.
Botpress (https://botpress.com), also open source, is another ethical alternative designed for building conversational interfaces. Like Rasa, Botpress supports on-premise deployment or private cloud hosting, enabling full control over user data and compliance with privacy regulations. While it is generally easier to set up than Rasa thanks to its intuitive user interface, it is slightly less flexible when it comes to advanced customisations. Nonetheless, Botpress includes important cybersecurity measures such as built-in encryption and role-based access control. It also benefits from active community development and rich documentation.
From a security and compliance perspective, both Rasa and Botpress offer more control than ChatGPT. Because they can be deployed on your own infrastructure, you avoid the surveillance risks and data transfer issues associated with US-based cloud services. This means better compliance with GDPR, the ability to localise data and customisable security measures.
Below you can find a table comparing the characteristics of each alternative.
Security and Compliance Comparison
| Feature | ChatGPT (API) | Rasa (Self-Hosted) | Botpress (Self-Hosted) |
|---|---|---|---|
| GDPR Compliance | ⚠️ Complex | ✅ Full control | ✅ Full control |
| Encryption Control | ❌ Limited | ✅ Customisable | ✅ Built-in |
| Data Localisation | ❌ No | ✅ Yes | ✅ Yes |
| Surveillance Risk | ⚠️ High | ✅ Minimal | ✅ Minimal |
Besides this, taking into account Environmental Considerations, using on-device or self-hosted solutions reduces energy consumption associated with cloud-based large language models. When cloud services are necessary, selecting providers with renewable energy commitments or carbon offsetting schemes can help mitigate environmental impact.
What about the cost?
While ChatGPT may seem simple and affordable at first, it can become expensive as usage grows. ChatGPT operates on a pay-per-use model, meaning that as usage scales, expenses can rise unpredictably. This is especially problematic in services where user volume fluctuates or grows over time. In contrast, Rasa and Botpress have no licensing fees, as they are open source. Costs are primarily related to hosting and maintenance, which are more predictable and can often be managed internally. Although the technical knowledge needed to implement these systems would be higher than ChatGPT, for small to medium usage (up to 1,000 daily interactions) a self-hosted solution is generally more economical after the first year, particularly if the organisation already has some server infrastructure in place.
While mainstream AI tools like ChatGPT offer convenience, they raise serious privacy, compliance and ethical concerns for organisations supporting vulnerable populations. Open-source, self-hosted alternatives such as Rasa and Botpress provide greater control, lower long-term costs and align better with ethical technology practices.
If you’re building an AI solution for a sensitive context and want to explore privacy-respecting, responsible AI, Pau&Company can help you design, deploy and maintain ethical digital tools tailored to your mission.


