Where are finance firms actually using AI?

Amelia (00:16)
Hey there, I'm Amelia.

Paul Wood (00:17)
I'm Paul.

Pat (00:17)
I'm Pat.

Amelia (00:18)
And this is Fin the Week. So just the three of us today — Russell is away, probably recovering from the marathon he ran at the weekend.

Paul Wood (00:27)
Yeah, when this episode goes out, it'll probably be a few weeks old. But he did a great job finishing the marathon, and especially the Guernsey marathon. I was watching some videos and they start on this epic hill climb that would just — I mean, I was saying to my wife, if I got to the top of that hill, I'd think, that was good, I'm done.

Amelia (00:48)
Has anyone checked in on him? Is he still standing?

Paul Wood (00:50)
Yeah, he's still going. He's still designing websites.

Pat (00:54)
Yeah, he'll be basking in the glory of his medal for months, I think.

Amelia (00:59)
Well, too right, too right. So we've made it to the end of Q2 and it's time to wrap up the Pragmatic AI Edition. In a slight change to the agenda, we're gonna go beyond the topic that we trailed last week, which was, my AI thinks it's a financial advisor. And instead, we're gonna be looking at AI in the wild. So we're gonna be seeing how the financial services sector is actually making use of AI. So Paul, what have we got in store?

Paul Wood (01:29)
Yeah, so I've spent some time this week hunting around for examples of where finance firms are actually implementing AI solutions, because over the course of the last few weeks we've spent a lot of time discussing quite theoretical stuff, and ways you could use AI, but I think it's going to be really interesting to look at, well, actually, how is it making a difference for firms in real life and how are they overcoming potential challenges, say regulatory challenges or challenges with data quality and things like that. So what I've done is I've lined up a load of examples and I've come up with a game that I've called Yay-aye or Nay-aye. So we can all discuss these examples and see, do we think they're genuinely good or are they just a bit of PR to try and get attention? And I was thinking Pat, you can bring the AI implementation expertise, and then Amelia, you're sort of our resident non-techie. So perhaps you can be a normal member of the human race and give your reaction on the ground, day to day. Do people notice this stuff or is it all just noise? So yeah, that's the plan for today.

Amelia (02:35)
I love it. Yay-aye or nay-aye. Sounds good. So let's get going. What's example one?

Paul Wood (02:52)
So the first example — actually, there's two examples in one here really. So Morgan Stanley has been partnering with OpenAI, the makers of ChatGPT, to deliver a couple of tools. The first one is Morgan Stanley Assistant. They've got this sort of business unit called AI @ Morgan Stanley — that's an at with the email at sign, so it's all very trendy. But they've essentially delivered an internal chatbot for answering financial advisors' questions. When AI first became known in the public sphere — generative AI, that is — everybody was talking about chatbots, and there was a lot of noise about creating your own chatbots, and the way people delivered that was often just to funnel in all of their company knowledge into a chatbot, and then the general public, or their customers, their users, could access that. But what Morgan Stanley have done here is they've created an internal chatbot that is basically to — I guess I'm putting words in their mouth — superpower their financial advisors. So there's a quote here from Jeff McMillan, the head of Firmwide AI, which says, this technology makes you as smart as the smartest person in the organisation. Each client is different and AI helps us to cater to each client's unique needs. And continuing: now advisors can engage clients on topics they haven't discussed before because the friction between knowledge and communication has gone to zero. So this is quite an interesting implementation, I think, because it's essentially giving each individual financial advisor a much bigger pool of information to draw from, way more than they could ever fit in their own heads. So yeah, reactions to that one.

Pat (04:49)
Yes. It's not a difficult thing to do. So I think there's some hype around this — Morgan Stanley saying, you know, we're rolling out AI, we've implemented this chatbot, but actually there's loads of platforms out there that make this really easy. And the challenge is actually just getting your knowledge repo structured. But that doesn't mean it's a bad implementation. The quality of the implementation should be based on whether it's actually helping people. And the numbers included in this case study are pretty compelling. It demonstrates that actually there are really easy, accessible ways to roll AI out across your organisation that can have really positive impacts on people's efficiency. So looking at access to documents jumped from 20% to 80%. Advisors can engage on topics they haven't discussed before because the friction between knowledge and communication has gone to zero. Yeah, I'd say it's a yay-aye for me.

Amelia (05:55)
Yeah, it's an interesting one, isn't it? And I know we've touched upon this so much in the past few weeks, and my knee-jerk reaction to all these things is like, but we're losing that human touch, we're losing that personal element. But again, those advisors are spending more time on client relationships because it's taking out so much of the legwork. It's actually allowing more of that. Do you think that's the case?

Paul Wood (06:18)
I can imagine it is. I think my reaction is that in this specific example, financial advisors are highly trained. They're held to high standards. So providing them with access to a super powerful resource that just gives them more information to draw from feels like a really good thing. But I suppose my fear would be that you could also implement something like this in a less structured organisation. So, you know, a typical support call centre for any type of organisation — it could be a finance firm, but just like a bank support centre with staff that, perhaps they're well trained, but they're not regulated in the sense that a financial advisor is. Would giving them access to a really powerful library of information, of stuff that they've never really addressed before, potentially cause more headaches? Because you're prompting them to discuss things that perhaps they only understand at a very basic level. How do you overcome that? It's the kind of rhetorical question I would have.

Pat (07:27)
Yeah, it's valid, isn't it? Is there an element of data overload as well? I guess in your example, Paul, if the chatbot says, you should talk about this thing with the client, and then the financial advisor starts talking about it with the client, and the client starts asking difficult questions the financial advisor doesn't know. I guess the proper response from the financial advisor would be, I don't have an answer to that question, I need to take this offline and come back to you. Some of them might actually just carry on relying on the chat tool, just proxying the questions into the chat tool and getting the answers and giving them straight back to the end user without a deep understanding of whether that is the right answer or not. So there's definitely risk associated with it, especially if you've got junior advisors talking about pretty complicated financial structures and solutions.

Amelia (08:20)
So overall, are we saying yay-aye or nay-aye for that one?

Paul Wood (08:23)
I think it's a yay-aye, but I think this is probably one of the outcomes of the conversation, and probably the whole edition that we've been talking about, which is that AI, or the outputs of AI, are really only as good as the person receiving them and knowing what to do with them. Because one of the other stats from this case study is that it's had 98% adoption in their wealth management team, which means basically everybody on that team is using it. And the law of averages would say that there's got to be someone on the team at any given time who's having an off day. And then is it going to just spread problems more quickly? So yeah, it's only as good as the person receiving it, I think.

Pat (09:11)
Yeah, but then would that person give bad advice anyway, because they're having an off day? I think there'll always be instances where it might cause the odd headache. But if every headache it causes, it solves 10, then the net benefit is obviously really positive. And I think it's important to balance the positives against the negatives with this kind of stuff, because people do like to coalesce around specific instances where AI's hallucinated something or got something wrong, whilst ignoring a lot of the incredible benefits that it can bring. What I would really like to see is actually how many people are truly using it. Because this is straight out of Morgan Stanley's marketing team, and they rolled it out and everyone's using it, but actually it'd be really interesting to get some honest feedback from people on the ground as to whether it's actually useful or not.

Amelia (10:03)
So our next example is another Morgan Stanley one, right Paul?

Paul Wood (10:07)
Yes. Yeah. So I think it's a further development of this suite of tools that they're building with OpenAI. It's AI @ Morgan Stanley Debrief, and this one is possibly more familiar to people. It's where it essentially records Zoom meetings, then with the client's consent, can action outputs like creating client notes. It automatically sends information into a CRM, creates follow-ups, summarises key actions from meetings. And I know that we as a company use solutions like this through Google's Gemini system that, for meetings that are suitable, will annotate them, and will provide recordings if we ask it to, provide transcripts. So this sort of stuff is in use. It's interesting in this case because it's obviously dealing with financial services and things that definitely need to be well looked after. The information that is processed through a Zoom call in this sort of scenario is potentially important information that can't be spread outside of the people who are permitted to see it. And another quote from this one is: one of the first questions we get is, is our information going to be used by OpenAI to train the public ChatGPT? And the response is that OpenAI promises zero data retention. So they say they don't retain any of it. So what's everybody's feelings on that? Would you be comfortable speaking to your financial advisor on Zoom, providing lots of personal data, knowing that an OpenAI system, or any other company's system, is going to use that to create outputs?

Pat (11:53)
Yeah, I mean, it's kind of par for the course these days. There's so many platforms that do this. It's built into Teams, it's built into Google Meet. I think Morgan Stanley promoting the fact that this is a really cool Morgan Stanley development is probably a bit of a stretch, because it's kind of a trivial thing to do. I don't know what extra stuff they've got running across the top of it, but I'd say it's useful. Do you know what I mean? In many ways, we use it all the time.

Pat (12:28)
We record a lot of our client meetings. We save the transcripts down into a client folder. We then use AI to crunch that folder along with all of the other context we've got around that client, and we use that to generate insights and intelligence and briefs and all sorts of stuff off the back of it. So it's incredibly useful. I think the capturing of transcripts and summarising meetings is just a really trivial kind of thing. It's what you do with those transcripts after, to improve the long-term client relationship and add more value. So it's kind of a no-brainer. I'm tempted to give it a yay-aye, because it's just like — do you know what I mean? It's like saying, we've got this cool technology where you can open a text file in your computer and send it to someone. We're calling it electronic post. Do you know what I mean? It's that basic.

Amelia (13:17)
It'll never take off, it'll never take… So what do we think with this one? What are we doing? Pat, are you gonna say yay-aye or nay-aye?

Pat (13:20)
I'm going to be cheeky and give it a nay-aye, because they're trying to hype up something that's just completely run of the mill — maybe because of their positioning on it rather than the technology itself.

Paul Wood (13:35)
Yeah, I think what Pat has said is probably fair. I think this case study was first published in 2024, so it's a little bit old, but even back then this wasn't groundbreaking. I suppose I haven't got the deep understanding of exactly how they use the information and whether they've got any clever processing that goes on that's more specifically geared towards the work they do. But I think the basic tool is actually pretty simple, so let's go nay-aye.

Amelia (14:10)
Nay-aye for that one. What's the next example?

Paul Wood (14:13)
So the next one is moving on to Klarna, who have an AI assistant. Again, funnily enough, this was first published in February 2024 as well. And essentially it's an AI assistant — something that customers can access via, say, their mobile app, and they can ask questions and the AI assistant will do its best to answer those questions. Some of the stats are that when this was published in 2024, the AI assistant had had 2.3 million conversations — two thirds of Klarna's customer service chats. It's doing — this is probably a controversial bit — the equivalent work of 700 full-time agents. So that really taps into the fear of AI stealing your job. The press release said it's on a par with human agents in regard to customer satisfaction score. It's faster at resolving the jobs it needs to — two minutes compared to 11 minutes previously. Some interesting things about it: it's available in lots of different markets, it's on 24/7, and it's available in 35 languages, which is a genuinely interesting thing, because to offer customer support in 23 lang… sorry, 35 languages is no mean feat. You couldn't do that with humans very easily. And it's estimated to drive 40 million US dollars in profit improvement for Klarna. So it's doing things like managing refunds and returns, fostering healthy financial habits. Actually, that's another point — it provides guidance on considerations that a customer might want to keep in mind. Should you be spreading the cost of this purchase? Is this the best way to do it? Stuff like that. So I think there's pros and cons to this one. For me, the pros are it's really smart use of technology to provide global, multilingual customer support, and it's clearly effective. On the downside, it's perhaps fuelling this fear that AI is going to steal your job. That's my view.

Pat (16:35)
I'm really excited about what AI is going to do to customer service. I noticed that it's mentioned there that it's going to drive $40 million USD in profit improvement. I'm hoping that's not just cost reduction off the bottom line, and I'm hoping that's actually more sales. Because I think that's what's going to happen with customer support. I wrote a white paper recently on how to build an actually useful AI-based customer support agent. And the reason I'm excited about customer support is because traditionally, customer support in SaaS has always been a massive compromise. It's kind of universally crap. There's this paradigm where, as a customer, if you need some help, you go onto their website and you click Support, and have to type in what support you want. And then they surface like 50 articles and none of them are relevant to what you want. And you have to go through the articles, and then there's another question at the bottom, like, did you find your answer? And you have to select no. And then you have to go through five or six different layers of radio buttons. What's your issue? And then at the end you get a contact form, and then you fill in the contact form and you type out your long issue, you click save, and then the page crashes and you lose your entire submission. Or if you do manage to submit the form, then it's like 48 hours before a customer support agent gets back to you, because the customer support team are just so overwhelmed. And because they're so overwhelmed, huge amounts of investment is put into trying to ease the load on that customer support team with these kind of crappy, half-baked initiatives to ease the load. Now, the technology is there today to surface an AI agent that's got full access to all of the documentation. So if your question is, how do I do this? It will say, this is how you do it. More importantly, the technology is there today for the agent to securely access your account, your transactions, your data within that SaaS platform, your logs, all that kind of stuff — request that data, crunch it, analyse what's wrong, what's happened, and then get back to you to help you through that process. And a lot of AI assistants sound quite basic — it's the documentation and the knowledge base, and they don't actually have access to the data, and they're not that useful. When you give AI access to the customer's data, suddenly it becomes incredibly useful, because that AI has got access to everything that the human customer support agent has access to. The only area where you might want to draw the line is where you start allowing the AI to change customer data. But as an example — imagine you've got a booking system, and the booking system sends out SMS reminders to your customers. So you've signed up for an account on the SaaS platform, you might have a hair salon or a beauty salon or something, and you have SMS reminders going out to your clients for their appointments. SMS is kind of a ropey technology a lot of the time. And so there'll be times where the SMSes aren't delivered and the customer will miss an appointment, and you'll phone them up and say, why did you miss the appointment? And they'll say, I didn't get my SMS. And then you have to go to the support team for the SaaS platform and say, you know, Jackie Smith didn't get her SMS, this is her mobile number. And then the customer support team will need to look at their SMS logs and go back to you. But actually an AI agent can do that now. You could go to the AI agent and say, Jackie Smith didn't receive her SMS messages. And the AI agent can dig out your SMS logs, analyse them and send you a message back saying, oh yes, I can see that we tried to send her SMS at 11.23 yesterday, but you've run out of SMS credits. Here's a link to go and top up your SMS balance. Done. Or, oh, I can see that Jackie logged in yesterday and she turned off SMS notifications herself, so she's not getting SMS notifications. So actually that's an issue with the client. And that level of personalised support is technically absolutely feasible. It's out in the wild. It's solving problems all the time. And over the next couple of years, we'll see that kind of more useful AI support become much more prolific. And it will start replacing these horrible, drawn-out user journeys where you're having to navigate through this horrible documentation and type out a long message. It'll all be done through a chat interface — chat to our agent, chat away. And actually if the agent can't solve the problem, it will then be escalated to a human who will then solve the problem more quickly because they are genuinely seeing a drop in the load. But actually I think the human teams will probably remain a similar size, because at the moment they're just unable to even cover a fraction of the support queries that come in. And I think the human teams moving forwards will start to see them being able to more effectively deal with the more challenging support requests that the AIs can't deliver, rather than having to handle loads and loads of random support requests for really simple challenges and basic questions. So yeah, it's really an amazing area. It's definitely a yay-aye for me on this one, because Klarna were doing this in February 2024, which is two years ago. So they were well ahead of the game on that.

Amelia (22:19)
It's interesting, because my knee-jerk reaction when I see — you know, if I think I'm speaking to a bot when it comes to customer services — this isn't going to be as good, because I'm not speaking to an actual person. But these stats just speak for themselves, don't they?

Pat (22:34)
Yeah, I've got an anecdote, actually — a really interesting anecdote. I went out for lunch with a friend the other day and he was talking to me about the NHS bot. He said to me, I've been feeling a bit under the weather for a few days, but I really don't want to have to go through the rigmarole of trying to get a doctor's appointment. But I decided I had to, because this throat infection just wasn't going away. So he phoned up the NHS — he phoned up his local doctor's surgery to try and get an appointment, and an AI answered the phone, and he had a conversation with this AI. And he was just like, God. He went through the conversation. It took him about seven or eight minutes, and the AI in the end interviewed him and asked him about his symptoms. And he was just like, I have a sore throat, feeling a bit under the weather. And the AI came back and said, well, actually, the symptoms you've given me fall within the remit of issues that a pharmacist can deal with directly. So would you like me to book an appointment with your pharmacist instead? And he was just like, no, I want to see a doctor. And the AI was like, well, no, the symptoms fall within the bracket of these issues, so you need to speak to a pharmacist. And it wouldn't let him speak to a doctor. He was really frustrated by it. And in the end, he gave in. He was just like, okay. It booked him in with the pharmacist, and said, okay, I'll put you in in an hour's time in the pharmacy down the road. So he turned up to the pharmacy really grumpy, and the pharmacist was just like, yeah, I've got your appointment here, I can see your symptoms, quick chat, here's your prescription, and I'm going to fulfil that prescription now. And he got home with this prescription and he was just like, hang on. I've basically phoned up the NHS, described my symptoms, got a prescription, collected my prescription in under 45 minutes. And it dawned on him that actually that's an incredible end result. So he's converted, and that's a really good example of a really amazing deployment of an AI agent by the NHS, of all organisations.

Amelia (24:46)
Yeah, that sounds amazing. That sounds amazing. I don't know about you, but my doctor's surgery — I still have to do it online. And it's one of these things where there's like a million different options, and if your symptoms don't fall into one of those options, it won't let you go to the next page. You have to just make something up just so you can get to the final page and fill in a box, and it's so frustrating.

Pat (24:47)
Yeah, there's a key word for that. Chest pain.

Amelia (25:10)
Yeah, chest pain and bleeding, done.

Pat (25:13)
Yeah, go to A&E. Chest pain, straight through.

Amelia (25:16)
So we're staying for this one, the Klarna AI Assistant. This is definitely a yay-aye, yeah?

Pat (25:22)
Yeah.

Amelia (25:24)
There's something else interesting that Klarna are doing, right?

Pat (27:33)
Yeah, it's really cool actually. When I first read it, I was just like, God, what a gimmick. But actually, when you're gathering data from users in UX research, there's two types of data you gather. There's quantitative and qualitative. Quantitative is where you've got data from lots of users, real numbers, and it's quite scientific, and you're just crunching different users taking different user journeys and finding out what works and what doesn't. You've got qualitative, which is where you've got small amounts of data but from very interesting, specific users. And that's like user interviews and in-person user testing and remote user testing and stuff like that. And this is a novel, cool way to get some kind of cool qualitative feedback. I'd be really interested to have a look at the transcripts of some of these conversations. I'd imagine there's a few journalists in there, probably a few people taking the mick, asking challenging questions, but there's probably some really genuine conversations in there from people who have good ideas, who have had bad experiences with Klarna, and it gives them an opportunity to articulate those experiences in great detail and fast-track that feedback straight into the product team. Really powerful. Actually, back on the customer support element — if you imagine how powerful it would be to have the AI customer support agent answer in the voice of the CEO, so that the CEO is answering all of your questions over the phone. Another really cool thing — I've pitched this idea a few times, I haven't actually pitched it to Premier Inn, but imagine if when you phone up Premier Inn, like, Lenny Henry answers the phone, and he helps you work through all of your issues. So as a customer, you'd find it difficult to get angry, wouldn't you? Lenny Henry.

Amelia (29:21)
That is an amazing idea.

Pat (29:24)
So technically, 100% achievable, the technology exists now. You just have to agree probably quite a lucrative contract to license Lenny Henry's voice. Aside from that.

Amelia (29:53)
It's an interesting one. And actually, I also do voiceover work. About 18 months ago now, it didn't end up happening, but I got approached — I've got a voiceover agency, and they approached me and said, we've got a company that want to use your voice and essentially clone your voice. And they said to me, what do you think? And I was like, well, what do you think? I don't know, is this where it's going? As somebody who's done voiceovers for years, my initial thought was like, no, I don't think we should be doing this. And then at the same time, well, if this is where the industry's going, I'd rather have the work than not. And this particular job was a big company who owned lots of offices in London, and they wanted to use my voice for internal things in their office, like the lifts and the phones, and kind of essentially be the voice. But it was gonna be like an hour's recording session, recording a bunch of stuff, and then that was it. And for whatever reason, the job didn't end up going ahead. But it was quite an interesting one, because, like I say, initially I was like, no, I don't know about this. And I'm really keen for a voiceover to stay as authentic as possible. But there's definitely been a shift, especially over the last 18 months. There's a lot less voiceover work. And I can instantly tell when I see things that are definitely an AI voiceover. I don't know how you guys feel about that — I mean, you know a lot about AI, but I'm not quite sure if I think the voiceover side of things has kind of got there yet.

Pat (32:55)
Yeah, but they'd still be way more excited to speak to the actual CEO, wouldn't they? And that's the point there, isn't it? Time for a human is a fundamental limitation. And for a human to strive to achieve something, the output of that thing that they strive to achieve has intrinsically a lot more value than if a computer achieved that thing in a couple of seconds, for a tiny amount of compute. And I think that where the desire of the output is to create an emotional response — whether that's in an ad or a movie or listening to an ebook or something like that — having an actual human produce that will just add so much more value to that thing, and humans would be so much more likely to pay for that thing. If you had an ebook read by the actual Stephen Fry or an AI Stephen Fry, you'd always choose the actual Stephen Fry version, even if it was like three or four quid more expensive. Because even if they sounded exactly the same — do you know what I mean? Right now you can probably tell, but it won't be long before it's kind of hard to tell. If you're sat there listening to the ebook and you've got a vision in your mind of Stephen Fry sat there reading this, that's just got so much more to it. I wrote an article on that a little while ago: when you're rolling out AI, when you're using gen AI in your business, think about the output you're producing. If you want it to create an emotional reaction, think about getting a human to do it. You know, the strapline on your website, product copy on your website that is going to encourage people to scroll down and get to the call to action, the copy on the call-to-action button — get a human to write all that, because it's so important and it won't cost loads. It'll cost a lot more than just bashing it out of ChatGPT, but the output of that will be more conversions. If you're writing software, if you're writing a bit of backend software that takes data from system A and system B, and you need to write the code quickly to a high level of quality, just use AI, because AI will do an absolutely brilliant job and literally no one cares. But don't use AI to write all the copy for your website and draft and schedule all of your social media campaigns, because they just won't get as much reach and engagement as if you use a human for that stuff. Because people can tell. People can tell.

Amelia (36:03)
So Sebastian then, Klarna AI — we're giving him a yay-aye.

Pat (36:08)
Yeah, I think so. Even though it'd be pretty cool if he could actually speak to customers. What they should do is like a lucky dip where one in a thousand calls he will actually pick up and have a conversation. And then you wouldn't even know as a human, but he could actually have the odd random call.

Amelia (36:13)
That would be excellent. I like the sound of that. Or just Lenny Henry. So what's next, Paul?

Amelia (38:21)
It's really clever, isn't it?

Pat (38:22)
Yeah, it's definitely really useful. And the data protection concerns — there's always, in data protection, if you're using data to prevent fraud or illegal activity, then that is kind of universally accepted. You don't need the user's permission. If you're collecting IP addresses and names and stuff and using them to detect fraud, that's an acceptable, legitimate use. So from a data protection perspective it's absolutely fine. What would be a problem is if they were collecting all of this user information and then cross-referencing it with sales patterns across different clients and producing profiles of these users and then reselling the users' profiles as targeted advertising to different users and stuff like that. But from a purely fraud prevention perspective, data protection is absolutely fine. This is a massive problem with SaaS software, and has been for decades, ever since SaaS was invented. And Stripe, with their reach — the fact that they've got so much data and so much of the market share — are able to pull all that together and add all this value. It's just a really clever solution. So yeah, a big big thumbs up, yay-aye for me on this one.

Amelia (39:37)
What do you think?

Amelia (39:50)
So, another yay-aye. Lots of examples there of how AI is being used by these huge corporations. Really interesting stuff. And of course, like you've said yourself, these services are only going to get better and better. But for now, shall we play some Jargon Busters? So this is a little game that we play every week where we pick a term from the jargon busters list to see if you guys know what it means — an industry term — and this week it is IPO.

Pat (40:23)
Initial public offering.

Amelia (40:25)
Yes.

Pat (40:25)
So that's when a firm is going from private to publicly listed. It's the process of doing that and launching it onto a public stock exchange.

Amelia (40:36)
Absolutely right. And a nice easy one this week — I'll have to make it more difficult next week. So IPO, initial public offering. Absolutely right. We will have more next week. But what else is coming up next time, Paul?

Amelia (42:17)
So, looking forward to getting into that next week. We will be back then. Have a great week. Take care.

Pat (42:22)
See you later.