# Draft answers for advice-line staff and have a supervisor approve each one

Caddy drafts answers for Citizens Advice staff from trusted guidance, and a supervisor approves each one before the adviser uses it with a client.

Organisation: Citizens Advice Stockport, Oldham, Rochdale and Trafford (SORT), with the UK government's Incubator for AI (i.AI)
Status: Piloted; deployed in six local Citizens Advice offices in England and Wales, with Caddy 2.0 announced in November 2025 and an open beta that more than 100 offices had signed up for
Place: United Kingdom
Categories: Financial inclusion & benefits, Justice & legal aid
Built with: Claude
Try it with: Claude, ChatGPT / OpenAI models, Gemini, Grok
Time to try: 25 minutes

## Why it matters

Advisers at Citizens Advice help people with benefits, debt, housing and work problems, and when they get stuck mid-call they wait for a supervisor to look the answer up. Caddy, built by Citizens Advice Stockport, Oldham, Rochdale and Trafford (SORT) with the UK government's Incubator for AI (i.AI), does that research first and hands the supervisor a draft to check. In a trial covering more than 1,000 calls, about 80% of its drafts were good enough to pass to advisers without changes, usually within four minutes, and advisers using it were more than twice as likely to feel confident giving advice than a control group. The team kept it away from clients on purpose, because many people who call are vulnerable, and every answer still goes through a person.

## How it works

An adviser types a client's question into Caddy through the chat tool the office already uses. Caddy searches only trusted sources, such as the Citizens Advice adviser guidance (AdviserNet) and GOV.UK, and writes a draft answer using retrieval augmented generation; Citizens Advice in Greater Manchester said in 2025 that it runs on Anthropic's Claude model. The draft goes to a supervisor, who checks, edits or rejects it before the adviser sees it, so the service never gives unchecked AI advice to a client. Questions are routed by topic, and the team tracks queries on a dashboard. In November 2025 Citizens Advice SORT announced Caddy 2.0, which splits each answer into separate claims and checks every claim against the trusted sources, and said more than 100 offices had signed up for an open beta.

## Who it's for

Advice charities, council helplines, housing and debt advice teams, and public contact centres where front-line staff wait for a supervisor to check an answer.

## The prompt

```
You are helping me design a supervised answer service for advisers at an advice charity or public helpline, where an AI drafts answers and a supervisor approves them before an adviser uses them. Do not ask me for any client's name, contact details or case records.

Our organisation: [for example, a local advice charity, a council helpline or a housing support team]
Topics we advise on: [for example, benefits, debt, housing, employment]
Trusted sources we may use: [list the guidance, websites and internal manuals]

1. Write the instructions for the drafting assistant: answer only from the trusted sources, quote and link the passage behind each statement, say plainly when the sources do not cover the question, and never guess eligibility or amounts.
2. Write a template for the draft a supervisor reviews: the adviser's question, the proposed answer in plain English, each claim with its source, and a confidence note.
3. Write a checklist the supervisor uses to approve, edit or reject a draft in under two minutes.
4. List the questions that must skip the assistant and go straight to a supervisor (for example a risk to someone's safety, a court deadline, immigration, or anything the sources do not cover), with the exact handover text.
5. Suggest how to log rejected drafts so we can find gaps in our sources each month.
6. Suggest the measures for a four-week trial: share of drafts approved unchanged, time to an approved answer, and adviser confidence, compared with offices that do not use it.

Do not invent rules, benefit amounts or legal deadlines; leave placeholders. Mark every rule that a supervisor or the charity's quality lead must sign off before use.
```

## Caveats

The trial figures come from i.AI and Citizens Advice SORT, reported in March 2025; no independent evaluation was found, and the published summaries do not give the size of the control group or show client outcomes. The deployment described in the sources is a pilot in six local Citizens Advice offices; Caddy 2.0's open beta and the national rollout targeted for 2026 had not been confirmed as complete in the sources checked. In March 2025 the team said future development and funding were still a live question. The service depends on supervisors checking every draft, so the time saved is in research and typing; the supervisor still has to read and judge every answer. Any similar service should keep the AI away from vulnerable clients unless it has been tested for them, and should log every rejected draft.

## Sources

The trial figures come from i.AI and Citizens Advice SORT themselves, relayed by UKAuthority and Computing. The SCVO write-up and the Computing interview describe the design from the team's own presentations. The UK government's AI Knowledge Hub page is a government source, but i.AI co-developed the tool. No independent evaluation was found.

1. UK Government, AI Knowledge Hub (DSIT, i.AI), UK Government Web Archive copy: "a human-in-the-loop copilot for front-line contact centres"; "6 local Citizens Advice offices". Caddy (AI Knowledge Hub tool page, archived) (https://tnaqa.mirrorweb.com/ukgwa/20260702051918mp_/https://ai.gov.uk/knowledge-hub/tools/caddy)
2. UKAuthority: over 1,000 calls; 80% passed to advisers without revisions; "humans in the loop". i.AI to extend use of Caddy customer service AI assistant (https://www.ukauthority.com/articles/iai-to-extend-use-of-caddy-customer-service-ai-assistant)
3. SCVO (Scottish Council for Voluntary Organisations): "It's built to support advisers, not replace them"; uses Anthropic's Claude; AdviserNet and GOV.UK. Building Caddy: an AI support tool for advisor teams (https://scvo.scot/p/97562/2025/03/10/building-caddy-an-ai-support-tool-for-advisor-teams)
4. Computing: "cannot be client-facing"; supervisor time per query cut from 10 to four minutes. Why Citizens Advice built a chatbot, then made sure citizens couldn't use it (https://www.computing.co.uk/interview/2025/why-citizens-advice-built-a-chatbot)
5. UKAuthority: verification engine checks claims against trusted sources; over 100 offices signed up for open beta. Citizens Advice SORT to launch Caddy 2.0 (https://www.ukauthority.com/articles/citizens-advice-sort-to-launch-caddy-20)

Fact-checked: 31 Oct 2026

Page: https://escapevelocitysociety.com/picks/citizens-advice-caddy-supervised-answers

See Selected Projects: https://escapevelocitysociety.com/picks.md
