How AI can improve public services
Many people are interested in how AI is going to reshape public sector services. There has been tremendous discussion about the values, methods and ethics of using AI in the public sector. In particular, there is a large open question about if and how AI could be used to deliver better user journeys. At Register Dynamics, we know that to spot the real value of an emerging technology, you need to watch the places it is being applied successfully by early adopters.
Here we'll introduce a case study we've seen in Cape Town where a homeless charity was trying to improve their homeless shelters. They employed AI in their design system in three main ways:
automatic facial recognition,
automatically transcribing conversations
and automatically summarising those conversations as answers to the questions that the shelter needed to know.
How the service works
The homeless shelter has a process that an applicant goes through at the point they arrive at the shelter. A representative from the shelter collects information about the applicant, their needs and any risks they might pose so that the charity can make a decision about the best way to help them whilst keeping everyone safe.
Someone from the homeless charity works through the screens on a tablet with the newly arrived homeless person. Initial iterations of this interaction were implemented using a design system as commonly seen in digital journeys online. A design system is a set of standards, guidelines, and reusable components that help teams design and build digital products.
This can often be a tricky interaction for lots of reasons. The applicant could be vulnerable. The people representing the shelter are often drawn from the community, so may have previously been homeless themselves, may know the applicant, either in a positive or negative way, or may be vulnerable to influence from gangs or crime. Either person may have a low level of literacy. The design of the interaction needs to take these factors into account and the charity in Cape Town was trying to make the process easier and safer so that their shelters could offer help to more people.
How AI helped
Filling in the application form
Early prototypes of the service had a lot of structured questions that took a long time to work through. However, user testing showed that the participants preferred to have a more freeform conversation and even when specific prompts were given, they often got sidetracked talking about separate but important issues. It was difficult for the representative to keep track of everything and capture it in the service in the correct way. This is where AI really supported the process.
The service was redesigned to transcribe the spoken conversation automatically, synthesise and summarise the information that the service required and then present it back to the participants for final review. These applications of AI were found to be both effective and safe at improving the overall functioning of the service. They're also AI features that Large Language Models have been shown to be pretty good at. Moreover, they do not require any context that is not provided by the user. In this application, LLMs are not being used in an exploratory manner in lieu of a search engine or domain knowledge: they're being used to process natural language data that is provided then and there by the users.
Understanding who is applying to the shelter
The other application of AI in the service used an even older AI technique but proved to be a bit more sensitive.
When an applicant first arrives at the shelter, the representative needs to know who they are. People in this segment of society are often "thin file" users. By definition they have no stable place of abode. Often they do not have, or have lost, any official government ID documents and they don't always have bank accounts. Nonetheless, the shelter needs to know who this person is so that they can keep the shelter safe. They need to know if this person has a history of disruptive behaviour, substance abuse or even a gang membership that might make them incompatible with someone else already admitted to the shelter. Often there is an adversarial part of the identification: those who pose the biggest safety challenges to the shelter are those who are hardest to positively identify.
The service design employed AI image recognition to match a photo of the applicant’s face with people who had previously applied to the charity's shelters. This allowed them to understand the history of the applicant.
Automatic face recognition has been a staple part of products such as Google Photos for a long time now. It makes it easy to find all the pictures of your mate Dave. But in this application things are a bit more sensitive. These are important issues so we'll cover this more in a future blog post.
What this means
I think this application of AI in the service design of this Cape Town homeless charity is a really positive story about how far the design of user facing services has come over the last decade and how much better we can still make them by applying new technology as it becomes available.
AI can be an incremental addition to what you’re already doing. It doesn’t need to be groundbreaking. We’ve seen it in this example and it has made a real impact. It isn’t risky and it works really well. Even if everyone you know is vibe coding and getting side tracked with AI hallucinations, there's still plenty of opportunity to use new AI technologies such as LLMs in a safe, productive and useful manner to iterate on and improve your existing services. Be on the lookout for cool, safe and seamless applications of AI and get in touch if you see any particularly good ones!
Author
