Citizens Advice SORT Caddy

Watch this video by Stuart Pearson from Citizens Advice – Stockport, Oldham, Rochdale and Trafford (SORT).

Download this video clip.Video player: course_4_interview_video_clip_1.mp4

Show transcript|Hide transcript

Transcript

FRANCINE RYAN

So, please can you introduce yourself and tell us a little bit about your job and your role?

STUART PEARSON

So I'm Stuart Pearson. I'm the head of innovation at Citizens Advice SORT, which is part of the Citizens Advice Network. My role is to look at how we can use technology to help deliver on our mission at Citizens Advice, which is to provide free, confidential, independent, impartial advice to people, whether that is through making access to advice easier or making advice-giving for advisers more straightforward and more helpful.

FRANCINE RYAN

I know it's part of your role. You've developed something called Caddy. So would you tell us a little bit about what Caddy is and how you're using it within Citizens Advice SORT?

STUART PEARSON

Yes, so Caddy… I mean, one of the drivers for Caddy was that with the cost of living crisis, demand for advice has never been higher. And also, the issues people are presented with have never been more complex. And so, like a lot of probably organisations, when ChatGPT dropped on us all in November 2022 both myself and my much smarter colleague Curtis thought, is there an opportunity where we could use generative AI somewhere within the advice process which could help us meet some of this demand?

And that was the starting point really. And one of the things we thought of is, well, how do you use it safely? How do you mitigate the risks? Because everyone will be aware now around the risk of hallucinations with generative AI, where it will provide you with an answer that is very convincing but completely made-up.

So before we wrote a line of code, we sat down with our advisors and supervisors, talked about the opportunities that generative AI might provide for us, and asked what the pain points were that we had within the service. One of the things we decided very early on was that we wanted to make sure clients spoke to people, so that any kind of AI solution would not take away from that kind of human-to-human experience that for Citizens Advice and for the clients who approach us and for our advisors is so important.

But one of the areas that we did identify was… we have kind of support spaces, chat spaces, where an advisor, when we have a person (we call them clients) approached us with an issue and maybe the advisor's not too sure or they want some validation. If it's something they've never come across again or if the issue’s extremely complex, they'll drop into one of these shared chat spaces the question for one of our kind of pool of supervisors to kind of give their expert opinion on.

So what was happening with that kind of surge of demand I was talking about is that the spaces are getting very busy with multiple inquiries from our advisors. And the result of that was it was often taking longer for supervisors to respond.

Because issues were more complex, it was taking longer for them to research and respond. And sometimes, adviser having to put the question in a couple of times because it had been missed. The impact then was that it was taking longer to get the answer back to the client.

So we thought, well, here's a space where we could potentially, to paraphrase Keir Starmer, we could inject mainline AI into the issue. So we created an LLM-powered chatbot which is now called Caddy. And Caddy sits within that supervisor space. So what will happen with the process now is again, to remember the client will always speak to a human advisor. If the advisor is not sure, they will ask Caddy the question.

Another thing we did to help improve the accuracy of Caddy and its responses and try and mitigate some risks of hallucinations was to use a system called RAG, which is Retrieval-Augmented Generation, which is a very fancy term for saying restrict the information sources. So Caddy only goes to three trusted sources. One of them is gov.uk. One of them is the Citizens Advice public site. And the third one is the Citizens Advice private back-end site, which is targeted for expert advisors. It's not available to the public.

So when Caddy’s asked the question, it will go away and look at those information sources. Then when it returns the answer, it doesn't return it straight back to the advisor, it returns it to a supervisor first. So a supervisor gets the answer, checks it's got the sources, the links are in there, the sources.

Then the supervisor approves it. Now that is a really important part of the process. Because one, we're assured that information has had human oversight, and that when that information goes back to the adviser, the advisor's confident that information is correct for them to pass on to a client.

The other reason it's important is because of compliance. What that positive approval does, when there's a button you press to say this is correct, it creates an audit trail within the system to say this has had human oversight. This is a human approval. Therefore, in terms of compliance and liability, the advice given, even though the research has been done by Caddy, the advice given has been by that human supervisor. They're the person who's approved that. So by doing that, we've got kind of human guard rails on either side.

End transcript

Download

 

Stuart is Head of Innovation at Citizens Advice SORT, and his team have developed Caddy which is an AI Copilot being used by advisers to provide answers to questions that are validated and signed off by a supervisor. Caddy is an LLM, but it has been custom-built by inhouse technologists at Citizens Advice SORT and draws on content from trusted websites. It produces AI-generated advice based on existing documentation directly within the messaging interface (Google Chat, Microsoft Teams, etc).

Some key points about the Caddy use case:

• Caddy answers are regulated by ‘human in the loop’ validation: answers are not presented directly to the public, but only after they have been approved by a supervisor.
• The responsibility of ensuring the answer to the question is correct lies with the supervisor not Caddy.
• The purpose of Caddy is to support Citizens Advice to improve its processes in order to help more people and is an example of AI implementation for public good.