What AIs Want
Flowers and chocolates or electricity and compute?
Dateline: Budapest, 28th May 2026.
In our 2025 paper “Digital banking in the artificial intelligence era: Strategies for serving nonhuman customers” in the Journal of Digital Banking, Kirsty Rutter of Lloyds and I wrote that while AIs were unlikely to care about which soccer team a bank sponsors or how cool the TV ads of an insurance company are or what the Super Bowl show featuring plugs for a crypto exchange looks look, there are presumably some things that bots would care about. But what?
Active Agents
Jean Luc Di Manno, the Innovation Lead in Payments over at Fime came up with a very nice way to make me think about such things. He observed that while the commerce of today is built on human emotions (such as desire, aspiration and impulse) and then went on to ask a very important and very interesting question: How do you seduce an agent?
(Well, how indeed. I don’t think flowers and chocolates will turn a bot’s head, but memory, compute and connectivity might.)
I think it is possible that some approaches to that fundamental issue are beginning to crystallise. Here is an example. The management consultants at Bain have been looking at what is going on in the travel sector and they have found a few early lessons that I feel might also apply in financial services sector.
Travel is a good sector to look at because it is often used as the canonical example of agentic commerce (“hey computer book me a flight to New York going on either Friday or Saturday and coming back the following Thursday and book me one of the hotels I like near Times Square”) so it is interesting to see what we can learn from the early efforts in the field.
Bain found that AIs frequently went to Online Travel Agents (OTAs) to book rather than direct to airline websites even after figuring out which was the best flight and note the implication: AIs favour sources with the easiest downstream interaction and OTAs currently produce cleaner, more structured and generally more agent-friendly data than airlines do. They were also able to complete bookings reliably using OTAs but not airline websites.
From this and their other experiments, Bain put forward set of sensible suggestions for the travel industry centred on the core proposition that to compete in an AI-agent-mediated world, airlines must start optimising infrastructure that agents care about rather than interfaces that people care about and design for trust not access. I rather like their framing: make agents care where booking happens. This means optimising data for machines which in turn means delivering structured content, stable application programming interfaces and consistent identifiers. If agents cannot reliably parse and compare offers, they will simply route demand elsewhere.
(Put in a pin in that point about trust. It really resonates, so we’ll come back to it later.)
If we think about agentic commerce in general, rather than only in travel, then we can find more early lessons to feed into our strategy formulation. Look, for example, at the OpenAI experiment centred on creating a “product feed” for merchants to share structured product data with ChatGPT so it can accurately surface their products in search and shopping experiences. The basic idea was that the merchants provide a feed containing key details such as identifiers, descriptions, pricing, inventory, media and fulfilment options.
OpenAI’s Instant Checkout was built on the Agentic Commerce Protocol (ACP), which was co-developed with Stripe. The concept was relatively simple. If a user found a product they wanted that was supported by Instant Checkout, all they had to do was click the “Buy” button in ChatGPT and it was all set. After only six months, however, OpenAI have abandoned Instant Checkout and have shifted their focus to product discovery.
It is the same in Europe, where Carrefour has just introduced its ChatGPT service so that consumers can check availability, select items and choose delivery options. When it comes to paying, they are directed into Carrefour’s existing e-commerce platform. In summary, then, agents have so far turned out to be useful for helping to find things, but not to pay for them
(Simon Taylor jokes that agentic commerce has more payments protocols than payments.)
OK, OK
I do not doubt that what is true in general commerce and true in the travel business is also true in the even more commoditised world of financial services. If I ask the agent to set up a savings account where my wife and I are going to build up our holiday fund for next year, then it will search through tens, hundreds, maybe even thousands of options before presenting me with its choice, or a small number of choices.
At which point I will then undoubtedly just press “OK” as I neither know nor care about the difference between a Citi savings account or a Wells savings account (or, for that matter an X Money account or a tokenised money market fund). I will simply tell an agent to go and do it.
The agent will look at various financial services organisations agents to see what they have on offer (ignoring, of course, organisations that do not have comprehensive content, good APIs, fast response times and so on). Marketing and advertising money will be completely wasted since I will never pay any attention to it.
Trusted Transistors
Does this mean that the agent will always pick the highest interest rate or lowest cost? Not necessarily. Remember that point about trust. Since trust is critical in financial services, how will my agent know whether to trust its counterparty or not (that is, is my agent talking to the real Barclays or to a Cambodian scam factory Barclays) and how will Barclays know whether the Martin Lewis Moneysupermarket Bot that they are talking to is the real thing, has not been tampered with and it acting on my behalf and within my authority?
The answer, of course, is cryptography. A key reason why I think that we should move toward agentic finance in the mass market is that cryptography delivers security that is lacking in the “real” world. As I have framed the issue before, you can make a fake Brad Pitt video but you can’t make a fake Brad Pitt digital signature.
When we shift financial transactions from me talking to my bank to my bot talking to my bank’s bot, we are doing a lot more than relieving me from boredom. We are moving from an environment based on logos and OTPs and call centres to an environment based on digital signatures, encryption and frameworks. It should be a matter of public policy to do something about the rise of fraud, the collapse of trust and the subversion of online spaces by getting bots in the loop as soon as possible.