Smart Money

2. CHAT-BASED COMMERCE EVOLVES

Not only are banks experimenting with AI personalization to help people manage their money, retailers are also testing out ways to help consumers spend it. In April, German shopping platform Zalando announced plans to integrate ChatGPT so users could ask questions such as, “What should I wear to a friend’s wedding in Greece?” The chatbot would then give shoppable recommendations based on factors such as the weather.
Some analysts imagine that, in the future, users may be able to give natural language prompts that automatically fill whole shopping baskets with relevant goods. Perhaps you would type: “I am planning a dinner for four adults, and we want to eat spaghetti bolognese.” Push a button, and there are your ingredients plus a recipe. Think of it as a new form of retail interface.

An obvious evolution of this idea would be to make it possible to pay by natural language, too. Chat-based commerce already exists. US-based Clickatell has developed a chat commerce platform that enables brands to use chatbots to communicate with customers and send them payment links over popular messaging apps, such as WhatsApp. A brand might use WhatsApp to message a customer if an item on their wishlist is available: “Hello, Marian—the yellow high-top sneakers you wanted are back in stock! You’re a size six, right? Would you like to go ahead and order a pair?” If Marian says yes, the chatbot could send her a secure and personalised payment link in the WhatsApp chat.

Chat commerce is innovating fast. In April, Clickatell launched a new feature so users could permit a company’s chatbot to save their payment details using technology called “card tokenization”. Card tokenization is a security measure that replaces card details with a series of algorithmically generated numbers called a token. That token can then be traded through the internet without revealing a person’s real payment details. For Clickatell, this streamlines the user experience, as they no longer need to leave WhatsApp to make their payment. Instead, the chatbot would ask: “Hi, you have left an item in your shopping cart. Would you like to pay for it now using your card ending in 1395, YES OR NO?” If they answer YES, their card will be charged automatically.

For German software provider Serrala, chats are one of several channels through which it enables the online payment of bills and reminders—others include email and SMS. Whenever a bill isn’t paid automatically, companies need to reach people and get them to take care of payment. “That’s a very different paradigm than shoppers visiting your web shop or customer contact,” says the company’s Solution Architect Jeroen Dekker. “So the machine learning here is not about the payment itself, but about deciding the right message, place, and time for each interaction to drive attention and conversion”.

3. BRINGING THE (SYNTHETIC) FIGHT TO FRAUD

There’s a normal way to behave in a shop. Perhaps you walk through the front door, browse the products, maybe experience a moment of indecision, then make a payment. Online it’s no different. “There are certain characteristics that we can identify as harmless and likely to have been authorised by the individual,” says David Britton, Vice President of Strategy, Global Identity and Fraud at Experian, which uses AI to detect fraud on behalf of its retail or banking clients. It’s therefore potentially significant when standard patterns aren’t followed.
“Perhaps they don’t come through the front door, they enter the store through the equivalent of a side window,” says Britton. “And then instead of doing some shopping and putting some things in their cart and taking things out of a cart, they go right to the most expensive item, grab three of them and then go right to the checkout.”

To identify suspicious behaviour such as this, Experian uses a tandem of supervised machine learning (AI designed to look for characteristics known to be associated with fraud), and unsupervised machine learning (where AI is searching for signs of fraud not yet discovered). To find these patterns, algorithms search through mountains of payments data as well as information on factors such as typing speed and web-page navigation.

The problem is that AI is enabling an army of digital proxies to act on the consumers’ behalf, and this is muddying the definition of what legitimate activity looks like. As we have seen, AI tools can find products we might like, notify us of payment options, or simply tell us what’s going on in our bank accounts. Bank aggregator software, which gathers information from a user’s multiple bank accounts in one place, is also growing in popularity. “Every time those systems go and look at your bank account balance, they’re acting as your proxy and logging in as you,” says Britton. Right now, he is preoccupied with a question: “How do I know that that particular bot or gen-AI entity has truly been authorised to go and act on a person’s behalf?” That unknown makes it challenging to learn what’s normal and use that information to sift out red-flag behaviour.

AI is also going to give fraudsters more tools to beat standard security measures, in turn putting yet more pressure on AI to detect ever subtler anomalies in the data. “Generative AI can be used to develop materials to move fraud forward at an industrial pace,” says Britton, who is already seeing criminals use the technology to generate fake identities, backed up with fake social media accounts and fake ID documents.

Many fraud detection companies are also witnessing criminals quickly adapt to using gen-AI. Fraudsters have been creating 2D and 3D deepfake masks to trick facial recognition security systems into authenticating payments or establishing fraudulent accounts. To try to stop this happening, identity verification company Onfido has been using gen-AI itself. “We use generative AI in the same way as fraudsters do— primarily to generate deepfakes using open source methods that spoofers are already using themselves to generate data,” says Therese Stowell, the company’s Vice President of Identity Verification. These can then be used as training data for anti-spoofing systems. “This helps our AI models distinguish between genuine faces and deepfake spoofs.”

It taps into a wider trend in the anti-fraud world: Using gen-AI to produce “synthetic data” of all stripes to better train machine learning-based tools. This type of work is important to combat overfit, Stowell says. Overfitting is when an AI model relies too much on available training data, and struggles to detect new, emerging anomalies. In deepfake detection, for example, that means some AI models can rely too heavily on characteristics known to be suspicious, such as the position of a fake face within the camera frame. But a position is easy for fraudsters to modify, she says. “So supposedly state-of-the-art defensive AI models often fail. As a result, we are using generative AI to help train our models with broader, more varied data to ensure simple adjustments do not bypass them.”