Caddy Q&A copilot

Caddy is an AI co-pilot developed by Citizens Advice in England to support legal advisers, helping them deliver faster, supervised answers with trusted information.
Project Description
Caddy is a behind-the-scenes AI assistant developed by Citizens Advice and the UK government’s i.AI team, in collaboration with volunteer developers and public sector partners. It supports frontline advisers by helping them draft and research responses to complex legal and public service queries. Caddy is not a chatbot for the public—it is a copilot for humans, built to enhance the productivity, accuracy, and confidence of those working directly with clients in high-pressure, high-need environments.
Caddy was launched in response to a major service delivery challenge. Citizens Advice—a nationwide legal and social support organization serving millions of people across England and Wales—has seen a sharp rise in demand during the cost-of-living crisis. Queries about energy bills, welfare entitlements, and personal debt have become more complex, while staff and supervisory capacity remain limited. Traditionally, junior advisers would escalate tough questions to experienced supervisors. But with caseloads growing and expertise stretched thin, this model led to bottlenecks and delayed responses, putting clients at risk of going without timely help.
That’s where Caddy comes in. Built originally by Citizens Advice Stockport, Oldham, Rochdale & Trafford (CASORT) as a prototype, and then scaled up with i.AI, Caddy provides draft responses based on trusted information from two sources only: Gov.uk and the Citizens Advice internal knowledge base. Using retrieval-augmented generation (RAG), semantic routing, and curated prompts, Caddy produces legally grounded answers designed to match the tone and accuracy of the organization’s traditional outputs.
Critically, every AI-generated message is routed through a human-in-the-loop process. Responses are reviewed and edited by supervisors before being shared with clients. This safeguard ensures both legal accuracy and a high standard of empathy, tone, and clarity. The system is tightly governed by ethical principles: clients always speak to a human, AI responses are never delivered directly to the public, and transparency and human oversight are required at all times.
Caddy is integrated into advisers’ existing workflows—responding to staff via their normal internal chat systems. It doesn’t require a separate interface or new training. For advisers, particularly new or junior ones, Caddy serves as a second brain and a safety net, supporting them in delivering more confident, comprehensive answers without increasing their cognitive load.
Evaluation and Impact
In a 2025 randomized controlled trial embedded in Citizens Advice operations, Caddy was tested across over 1,000 support interactions. Each time an adviser requested help, the system randomly assigned the question either to a human supervisor or to Caddy. Advisors then completed surveys at the end of each session, reporting their confidence, satisfaction, and case resolution status.
Results showed that 80% of Caddy’s responses were approved without changes, and that answers were returned in under 4 minutes—cutting drafting time in half compared to traditional supervisor workflows. Staff who used Caddy reported being twice as likely to feel confident in their answers and 1.5 times as likely to say the client’s issue had been resolved. Importantly, these improvements were delivered without skipping human oversight.
Design Challenges and Governance
The pilot surfaced several opportunities for improvement. Because Caddy only pulls from a small set of trusted documents, it sometimes fails to respond well to vague or unusual questions. Follow-up questions are limited in the current version, which can make it hard to gather all the needed facts. Additionally, the mandatory supervisor review process, while key to safety, can still create delays when staff are unavailable.
In response, the team is exploring options such as:
- Expanding the trusted source base beyond the initial two websites
- Adding conversational follow-up capabilities for better fact-gathering
- Allowing experienced advisers to approve low-risk responses themselves
- Building a vetted repository of reusable answers for common scenarios
Caddy’s development has been shaped by a commitment to inclusive and ethical AI. Citizens Advice worked with the Manchester Metropolitan University public AI advisory panel to incorporate community perspectives into Caddy’s governance and evaluation. In addition, Caddy is being released as open-source software, supporting broader transparency and collaboration across the public and nonprofit technology ecosystem.
Looking Ahead
Caddy is now live in six local Citizens Advice offices, with more locations adopting it as the rollout continues. Beyond Citizens Advice, the i.AI team is preparing to pilot Caddy inside UK government departments, where frontline workers face similar constraints in delivering guidance on immigration, tax, and other complex matters.
Future iterations will explore how Caddy can power AI-assisted training for staff, help build repositories of accurate answer templates, and support live interactions in real time. With each of these developments, the core commitment remains the same: to responsibly integrate AI into public service delivery without replacing the human connection that defines effective legal and social support.