Creating Digital Assistants
How we’re designing with AI tools to help our customers
At Barclays, we’ve been developing digital assistants to help our customers manage their financial lives. And we’ve been combining AI technologies and human centred design practices to do it. Here’s a quick overview of what we’ve done so far, what we’ve learnt, and where we’re headed next.
Helping people achieve their ambitions
Barclays is a very large organisation, as I found out when I joined. And in the UK alone, we have a huge customer base. Our CEO frequently points out that 24m people interact with our services in the UK. That’s 1 out of every 2 adults. So it’s essential that we’re always thinking about our customers. Our aim is clear too — we’re focused on helping people achieve their ambitions. That ambition can be many things. To buy a house or plan for a comfortable retirement. Or just to be able to check your account calmly, free from anxiety. At the root of this mission is the idea of connecting people and money.
Making banking completely personalised
We want to make the experience of banking completely personalised, which is where AI comes in. We want to be your financial partner, helping you understand your money, predict the future and work in synch with you. We want to be proactive and reactive. Looking at the data that makes up your financial life and giving you the insights you need. And we want to tailor this to everyone’s unique circumstances. As a company, we’re data rich. Banking transactions are, at their simplest, just data. Using AI and machine learning, we can interpret this data to help empower our customers. But there is so much data. The question we’re facing now is — what should we focus on? What will AI be good for, and how will it augment our existing services.
Solving human problems and answering core human needs
This is where design comes in. Design is essentially the business of solving human problems and answering human needs. By applying human-centred design, we can focus on what is most important to customers. Where their pain points are, where the friction is. And this isn’t a case of us simply asking people what they want.
We’re not trying to make a faster horse.
Henry Ford said if he’d asked people what they wanted, they would have said they wanted a faster horse. Human centred design is not about asking people what they want. It’s not about trying to make a faster horse. It’s more akin to a doctor. We don’t ask “what do you want?” Instead, we want to know “where does it hurt?” We do this in many ways. We do copious research. We study reams of data and analytics. We run regular diary studies. We build rapid prototypes and test with real customers. We test and tweak and iterate and retest constantly. We’re trying to get closer to the pain of our customers. And we’ve identified some clear, powerful user needs. One stands out for me.
How can we help people be better off?
People want to feel better off, and want help to do this. As a bank, we’re ideally placed to assist here. We’re good with numbers, we’re the experts when it comes to money. And, all our research tells us that our customers are desperate for this. It comes right back to our core mission — helping people achieve their ambition. Better off can come in many ways. From small things like letting people know when they’ve gone overdrawn, to big things like planning for their retirement. But how can we do this — how can we help people be better off? And how can we use AI tools to achieve this? Furthermore, how can those AI tools work with our branch and call centre colleagues to remove some of the admin so they can provide the real empathy and human service that our customers want and need. We started by looking around for a comparable model outside of banking.
Luckily, there’s a huge offering around this in the area of physical health. There’s a wealth of apps in this space, focused on trying to help you get fitter, get healthier, improve your lifestyle and achieve your fitness goals.
We learnt a few things from these health apps, and our experiments with them. We can alert people and remind them about their financial health. We can give them information about how to run their finances. We can nudge them in the right direction. We can present personalised information about their own performance. We can listen and react.
Design Sprints and demos
So we created a very rapid prototype of what the service might look like, and tested it with real customers. We built a decision tree engine which allowed us to structure conversations with our customers. We used a conversational interface so we could make it feel like a genuine assistant. And we used a series of buttons that allowed the user to respond to the questions easily. We’d previously experimented with NLP and intent training for full AI agents, but the process at the time was lengthy, and the training process was complex. But the worst outcome was that the user interface was frustrating — you had to work out the right questions to ask before you got a sensible response.
When we initially experimented with NLP-only interfaces, the experience felt effortful and error-prone. By limiting ourselves and our users to pre-formatted responses, we could take the effort out but still have a conversation. At first, we experimented with the concept of a finance coach. An assistant that gives you insight into your spending habits, your recent activity, spend categorisation, and tips about how you could improve your financial health. We tested and iterated the prototype over 4 weeks, testing each week. We then tested it with larger groups to see the reaction people had to the experience. And we discovered that people were desperate for this sort of service. They loved it. About one third of people wanted to be able to write their own questions and have more than just buttons, but they reported it was a minor frustration. The rest of our testers didn’t even notice they couldn’t write their own questions. They just didn’t need to.
One thing that became clear with designing a digital assistant is that we were designing a relationship over time. Normally when we design an experience we’re designing in space. We map out the flow, and we make it as painless as possible for users to travel through this flow. But here, the flow was completely dependent on the customer information, choices, and triggers. We’d be designing services that would reintroduce themselves into people’s lives at pertinent moments.
Here’s an example. We alert users that it’s payday (interesting challenge — how do we know it is pay day, and not just an influx of money? We can use an algorithm to work it out, based on frequency and repetition and size. Or we can simply ask users directly). And then, we change the design slightly — it’s more positive, more upbeat. And we give the user some helpful information — they had money left in their account on payday. This gives us the perfect opportunity to talk about pay day saving. Behavioural economics tells us that the best time to talk to someone about saving is on payday, when they have money. The worst time is a week before payday when they’re just trying to stay in the black. So we give some insight about it, and then help them set it up. In a few weeks time, we can contact them again about some other element of their finances. But the triggers are based on their personalised information, and we can score the triggers to ensure relevancy and importance to the user.
We’ve introduced the digital assistant to Launchpad, our beta app used by 10,000 customers. Launchpad gives us the chance to trial new ideas with our customers, and get feedback and reactions as well as behaviour and analytics. Here’s a short video that shows how the digital assistant works:
We’ve had some interesting responses so far. Firstly, people love it as an idea. All ages get the idea, and especially with older, more traditional customers, we get some great feedback. One user, a 75 year old ex-miner who doesn’t bank online, but does use his phone to message his kids and grandkids, was able to make payments faster than using the normal in-app method. But his comments were telling. He said the regular payment journey through the app felt you were the customer and the cashier. But with our digital assistant, we brought the cashier back. It felt more helpful, familiar and reassuring, and much more human.
Our next exploration was our Unlock Britain app.
We designed this for users overseas either visiting or living in the UK or looking to invest in the UK. Usually our relationship managers either in branch or in call-centers help these customers work through things such as education or buying property in the UK. But that often meant our highly talented staff were stuck processing repetitive information. What if we could digitise these conversations so we could answer the initial questions users had, but in the same friendly, human way they would with a colleague in branch?
Again, using a decision tree engine and a button-driven conversational interface, we created an easy-to-use, information rich service that qualified interested customers. It allowed them to speak to the right people, and provide them with a great level of service.
What have we learned?
Users love it… but it needs to do more
On first contact, users enjoy using our digital assistant. But they quickly learn it’s limitations. And this is likely to be a problem for AI solutions for a while — as inquisitive humans, we try to find the limits, the boundaries. And maybe that’s because we want it to be really clever and powerful. Or perhaps because we’d rather it wasn’t…
But our users, especially older users who are more used to a regular in-branch relationship, find this sort of interaction really useful. It’s a great way to introduce an older user base to a new format. And that’s because everyone can text, everyone can talk, and the interface isn’t alienating or confusing. There is very little to learn by way of tools.
And there are some clear accessibility benefits here too. By using interfaces that are familiar and very easy to grasp, we make using our services better for everyone. This isn’t just about providing for people with specific access issues — it’s about making the services we provide easier to use across the board. At Barclays we believe accessibility has a beneficial effect for all users. So we’ll continue to test with a broad audience, with different levels of skills and experience to ensure we can take everyone with us. By thinking of edge cases and those who might get left behind, we believe we’ll create services that are better for all users.
NLP and intent engines are great when they work, but frustrating when they don’t.
Thus far, NLP tools need either a greater degree of training, or they need to get better at inferring. Currently our experiments have meant that we’ve avoided using them. We’ll get there, for sure, but the button-based decision tree interface, despite it’s limitations, is less annoying and disruptive for our users than trying to guess how to ask the right question.
Deep AI is not required yet — decision trees can work well as a proxy
We haven’t yet come across a need for something really clever. We can plot out most of the customer needs and things they want to access, so we can map successful paths to the right end result. It’s possible that as we grow in capability, we’ll need a more autonomous, flexible, intelligent system. If we really want to partner with our customers, and be a vital part of their lives, we’re going to have to get there. But at these early stages, it’s not essential.
Tone of voice is crucial — we need to get permission in order to make it feel OK
We believe it’s fundamental to ask permission from the outset. Your bank talking to you via a bot can feel uncomfortable. A bot trawling through your account and presenting insights can feel intrusive. So it’s crucial to we ask “Is this ok?” Even though we already have all this information to hand, and our customers know we do, the act of asking permission is really important. It says clearly that this is their information, and we get to look at it and interrogate it only with their say-so. Our customers are in charge.
It can free up branch and call centre colleagues to add real help.
By taking away some of the repetitive form filling and simpler activities, we hope this can free up our teams in branch and in call centres to add real value to people. This isn’t a replacement to human interactions — it should be an enabler. There is, quite rightly, a lot of concern around how AI technology will make jobs redundant. We’ve seen this before with other levels of automation throughout industrial and technological developments, and the fear is often over-stated. Barclays has experience in this — we launched the first ATM, the first credit card in the UK, were quick to launch online banking, and were first to market with our mobile app and ideas like Pingit. And yet we have more employees now than ever before. And we’re delivering a better service than ever before too. We know that these advances are great at handling things that can be automated. And they free up our in-branch and on-call colleagues to do what humans are innately good at — empathy, reason, problem solving.
We’ve also looked at how we can deliver a seamless handover from AI to human interaction. We want our digital assistant to be humane, but not human. It doesn’t have a name, or an avatar, and that’s deliberate. That way, when we hand over to a human colleague to take over a conversation, we can announce it clearly: “I’m handing you over to Paul now. But unlike me, Paul is a real person, so please be nice 😄”
Proactive & Reactive
The key thing people really want and respond well to is when we’re proactively monitoring their account with them. When we’re pointing out how we might help save them money, or make more money, or avoid charges. And when something changes, or is about to change, we react appropriately. If they tell us they don’t want to hear from us so much, we quiet down. If they’re checking their account more regularly, we offer to automate the checking process. The key is to be monitoring and assessing all the time, always on, to create a reliable, meaningful digital assistant. And it’s by employing the oceans of data that we already have, unique to us as a bank with often life-long relationships with our customers, we can help our customers predict and anticipate. By doing that we can again reduce the friction, ease the pain and partner effectively with our customers’ lives.
We’re just starting, and there’s a lot more to come, but we know that being proactive and reactive is the way forward as our customers really appreciate it.
Where Next
This is a huge space and we’re growing our capabilities all the time.
Voice
We are, like everyone else, looking at voice-driven interfaces. Whether this becomes the new normal, or if it remains just an additional way of accessing information remains to be seen. But the accessibility use case is clear, and we’re focused on being as accessible to all as we can. Voice interfaces can provide that greater level of accessibility, and we’re experimenting to deliver this.
Other channels and platforms
We want to be where our customers are. Which means we need to look at how we can be useful on other platforms. We’re interested in how we work with large platforms such as Facebook, Whatsapp, Skype and others to be there for our customers.
Increasing assistant features based on customer needs
Whatever we do next, it’s vital that we continue to focus heavily on our customer needs. What is going to add the most value for them, and will lead to a demonstrably better experience.
It’s an ever-changing, growing area, and we’re developing our capabilities all the time. And we want to be part of the discussion around AI and design, how it will affect our lives and how it should be used. We’re figuring out the answers to this by designing and trialling with our customers, to ensure we’ve got their interests at heart. And we are being open about our work so we can engage in the conversations about this technology. We don’t have all the answers, and we’re finding our way by doing. We’d love to hear from anyone else doing the same so we can share learnings and best practice in order to deliver better, easier, more personalised services for our customers.
