What kind of personality should Lucie have?

In a couple of months we’re planning to start live testing of Lucie with a small group of people. The prospect is equally exciting and daunting as there’s a lot still to be completed in time, including something that’s been exercising my mind for ages: Lucie’s ‘personality’.

Much of what our first users will see is already decided. Lucie’s experience is conversational, so you just talk rather than tap and navigate through app screens. And the first functions that Lucie will support are also decided: everything to do with bank accounts, transactions and the myriad of questions that people will have about them.

But what I’ve been thinking about is different: it’s the sense people get of Lucie itself.

When we deal with another person, we pick up something about their motivations and priorities. A great deal of that comes from their mannerisms and behaviours. But a lot also comes from the way they listen to us and the way they talk back.

People will do the same with Lucie. Irrespective of the fact that Lucie is an AI agent, between the lines, they’ll form their own conclusions about Lucie. What’s its true underlying purpose? Is it honest and can it be trusted? Is it reliable? What are its priorities? Do its actions accord with its words? In other words, they’ll form a view about its personality.

So I’ve recently been working through what I want that personality to be.

Finding Lucie’s voice

Everyone’s familiar with the stereotype of the cheery customer service operator at the fast food outlet who has to greet you by name, ask you how your weekend was and ask if you’d like fries with your cappuccino. I know why that behaviour exists; it’s an expectation of the job. None of that is in any way appropriate for Lucie.

Cleo, the US money app, is an interesting comparison: some good, some bad. Its voice is casual, slightly Gen Z and occasionally snarky. It sometimes includes popular culture references in its replies and has a mode that will roast you about your spending¹ (Cleo reportedly had 15 comedy writers working on its voice;² why they needed 15, I’ve no idea). That’s kind of cute, and it has got them attention, but I don’t think it quite fits what we’re building; we’re not a ‘Cleo alternative’. What I do like, though, is how short Cleo’s messages are; that’s right for a conversational interface.

I’d like Lucie to be just as distinctive, but in its own way: easy to understand, every time. Short messages help, but what’s important is understanding, so the answer should come first and you shouldn’t have to read anything twice. Its warmth should come from paying attention to you and your money rather than from a cheery greeting, and its tone should suit the moment: light for a quick question, serious when something matters.

Lucie has to be honest

My starting point has been that Lucie has to be truthful and honest.

Much of this objective is achieved through something that has nothing to do with language: the verified data architecture that we’ve been working on for the last year. When Lucie ingests your bank data via the CDR, it passes through a verified data pipeline so that the data you end up seeing is directly produced in code rather than by a model. This is pretty foundational and allows us to have confidence about the accuracy of account and transaction data.

But getting the underlying data right is just the start. We also have to deal with how Lucie handles it. One of my great frustrations working with a large language model (I live in Claude Code these days) is that it will say one thing, and I’ll say, “No, we should do the other.” It immediately says, “Oh, you’re right, we should do the other.” Then I’ll say, “No, no, no, we should do the first thing”, and it says, “Oh, you’re right.” It’s more than frustrating. With Lucie we can’t afford that, because it’s all to do with money. Lucie needs to be evidence-based.

So when someone and their bank data disagree, Lucie is designed to show what it’s looking at and work out where the two differ.

Speaking of LLMs, another frustrating trait is their cheerful willingness to agree that they’ve been completely wrong, and their immediate happiness at moving on to the next thing, without acknowledging the harm the original error may have caused. This is completely unacceptable when it comes to conversations about money. If Lucie does get something important wrong (say it answers about the wrong account, or counts payments it shouldn’t have), it can’t just say “Oh yes, you’re right” and move on. It needs to stop, take stock and think about what that means for you. And when a mistake is made, Lucie should say so and apologise. Apologising when something really matters is a sign of actually wanting to act in that person’s interests.

Listening

If you care enough to listen, care enough to remember and care enough to ask, that’s the hallmark of a sensitive and concerned person. To the extent we can build something like that into Lucie, that’s what I’d like it to be.

Part of that is using what Lucie can already see. Access to your transactions is one of the main things Lucie has going for it. If it didn’t use that to understand you better, it would end up sounding just like a vanilla chatbot. The payments that go out every fortnight are things it should already know. Important facts about yourself that you’ve told it, it should remember. Where it’s guessing at what you want, or why you spend the way you do, it should ask. And what you tell it should count for more than anything it works out for itself.

And it can ask as many questions as it needs, one at a time. I think that’s a central part of wanting to know what’s important to you.

What Lucie is for

Underneath all of this is motive. Lucie is meant to work for the person it’s talking to, and to help with what they’re trying to do. Even so, a good advisor also thinks about the broader context and brings to your attention what you’d want to know: in an appropriate way, at an appropriate time, with appropriate formality and vigour.

Lucie isn’t an advisor yet, though that’s what we intend it to become, as our licensing allows. For now its job is your everyday money, so when a question crosses into financial advice or credit assistance, it should say so plainly and only help with what it can. It has to know both where it’s heading and what it can do today.

And if Lucie says it’s going to do something, it should do it. That applies to everything.

Shown in what it does

At the end of the day I think it comes down to the difference between someone who really does care about you and someone who’s performing. The second will say they care and do nothing about it. I want Lucie to be like the first: to show it in what it says and, most importantly, in what it does.

All of this is now extensively documented and is being built into how Lucie answers. The details and wording will change, probably many times, but I feel it’s a good start. The testers will soon tell us what they think.