Extend's Jamison talks:
- The role of payment networks in an automated payment world
- Cost as a barrier to AI adoption
- Detecting, and attaching blame, for fraud
The oft-cited 'Holy Grail' of automation, true end-to-end automated transactions, would include payment as part of the process, and frameworks already are coming into place that will allow for AI agents to conduct payment autonomously at a large scale. Mastercard, for example, this summer launched Agent Pay for Machines, and Visa recently has partnered with financial platform Highnote for capabilities around automated invoice payments and vendor spending. Andrew Jamison, CEO and co-founder of payment technology Extend, spoke recently with BTN executive editor Michael B. Baker about the potential for such technology and why autonomous spending will not be an overnight transformation for the industry. An edited transcript follows.
BTN: What are some of the ways payments players are leveraging AI?
Andrew Jamison: People underestimate that when you travel, part of the experience is making the choice. It's not about, 'I want a five-star hotel that's in this price range.' I want to check out if it's a smaller hotel, or has a nice bedroom or a cool bar, and it's all part of the buying process. That's why when it comes to automation, consumers will take longer than businesses.
When you think of everything people do today around the banks and card world, it's all about experiences. Experience has become top of mind for everyone, and part of the experience is not the automation. The curation of 'find me stuff,' but the choosing is not really the part people feel comfortable with automating.
In the context of business, there's been this idea that you constantly replenish items in your stocks when they start to dwindle. So, the fact that you have AI models now looking like an agent, and they're going to do all these different things and they start replacing individuals. We've seen transactions from Visa. We've seen transactions from MasterCard. I think those transactions are just a point that says we can do it. It doesn't mean it's being leveraged and used more broadly at this particular juncture.
BTN: When might that happen?
Jamison: When I speak to investment bankers who are really close to OpenAI and to the different players in this space, their sense is very much more one of, 'Look, 2026 is about building agents.' I don't think the payment capability, the automation of the payment, really comes in until next year.
First and foremost, they need adoption in terms of, 'What is it that you're automating?' And the payment is just one step of it. That's probably the least of their concerns, because they're going to build these agents that are going to be modified before the payment even happens.
You're seeing a lot of feverish work around the agents and trying to justify the value that they bring. In a world where tokens are becoming way more expensive, I don't think people have accounted correctly yet for what the actual cost of doing some of these activities is relative to the return. At some point, outsourcing this to different markets, where the cost of labor is way less, is actually still going to be the right [choice] for some of these decisions. It was fine when we all used a few tokens and whatever else, and then we got into token-maxing and then it's like, 'OK, that was a real cost associated with this, right'?
Each developer is about $110 a month for them to access their tokens. That quickly racks up across the team.
Agents are the No. 1 area of focus. Payments, the rails are there. I think the big challenge also now has to be one of, how do we put controls in place so that people either feel comfortable because the size of the checks and the velocity of the checks, or the funds that are being transferred are such that actually it's within my tolerances?
BTN: How is the trust built for that tolerance?
Jamison: If you go back even a decade or two, people said, 'The credit card networks are in trouble. Real-time payment rails, lower-cost rails are there.' I think people have fundamentally misunderstood the role of the network. The network creates instant trust between two parties that literally have no knowledge of each other. The network had the acquirers go out and speak to the supplier, the merchant. The card issuer went out and gathered data on that individual and then giving them a line of credit. So, there's trust, instant trust. You walk in anywhere around the world with a card, debit or credit, and there's instant trust, and you can walk out with the goods in seconds.
It got more complicated when some people started buying things on the web, because it's card-not-present, and fraud numbers went up. They started realizing we need to control for a few additional things. Andrew can't at the same time be in New York and out in Africa somewhere, right? The fraud models had to evolve.
The same is going to be true here, I think, on the agentic side, until we reach a level of trust and understanding of where this sets out.
BTN: What about AI's capability to detect fraud?
Jamison: Machines are able to identify things. You already hear about it a lot in a medical space. The models are able to detect cancers or pre-cancerous growth way quicker than the naked eye can, and with greater certainty, so less things are missed. The same is going to be true here in terms of doing those things.
Part of what we're really working on, on the AI side, is we want to help. Finance professionals and SMBs need to close their books at the end of the month. How do we help to close their books faster? Part of the task they need to do is somewhat laborious. You take your statement from your bank, and you truly reconcile them against all the charges that you see have come through. Until you've done this one-to-one match, you have no idea if that's truly what you owe.
There's a huge role for an agent to say, 'Give me your statements, and let me go back and look at all the transactions. Let me also very quickly figure out who's in policy, out of policy, if there's a meal on the weekend or someone taking an Uber on the weekend. There's a meal for a big-ticket number, and it doesn't seem like there's any justification behind it.'
From a compliance standpoint, there's just going to be massive benefit to getting this all done. They're going to have either post-transactional or pre-transactions. With the transaction itself, the question mark is always going to be, now you've introduced another person into the equation, which disrupts what has been readily agreed upon between merchant and the acquirers.
It's really around who has the weakest technology, and therefore who bears the cost in the event of fraud. Now the rules are broken, and then they find a way for assigning and attributing blame for incorrect purchases. There's going to be a degree of hesitation. It takes time to sort of figure out where the potential chinks are in the armor. Is it a client that's using the agent? Is it the agent itself that's liable? Is it the merchant that's liable? Is it the bank that's liable? To understand that framework is going to be really, really key.
BTN: When it comes to getting these guardrails in place, who do you think the players are really going to be in doing that?
Jamison: The different AI providers, they almost want to create their own rails. In that respect, they're taking the lead from Apple, when it started introducing the concept of Apple Pay, which created their own rails, and they started to be able to charge for that rail, right? They charge issuers for that rail, basis points on something, where it actually amounts to quite a lot of money.
But if you think about Apple Pay, it still rode over the rails of credit cards. There was economics there to make that all work. Whether you're in the AI realm or you're on the e-commerce realm, I think they're all trying to find these different rails, the challenge being again, what payment rail can you use?
Could you use a blockchain? Stablecoins are great at netting out exchanges. At the macro level, it works amazingly well because it's open 24/7, 365 days a year, and it's exactly what an exchange needs to settle en masse. But when it goes down for the last mile, the individual it's broken. I don't think we get into the world of stablecoins for another 10 years, [with] mass adoption of those kinds of tools and solutions.
The card networks are still playing a truly important role there, and I think they'll continue to do that. You look at the likes of the acquirers, the Stripes of this world are trying to play an important role here because they're already at the checkout with many of these players. So, they have a role to play as well.
There are other interesting players in the telco space, because we leverage virtual cards from the different issues that we work with. I can actually ping you an SMS at the point which the purchase is being made. Did you mean to issue a token, a virtual credit card for this amount? If the answer is yes, we're going to make it valid for five minutes. For this basket. Because you've acknowledged it, we've established now where liability stands. Now, obviously, if the agent screws up, that's going to be between you and the agent, but the others are out of that discussion.
BTN: If this is the year of the agent, what do you think will be the focus next year?
Jamison: In the B-to-B realm, there's already the concept of catalogs with the likes of Coupa and Ariba and those type of tools. I think in B-to-B, you're going to see a lot of things starting to be automated there on the payments front. But again, it's deeply curated lists, with really strong integrations into the existing organization. It's leveraging the technology and the guardrails that have already been put in place through platforms, then just [to] affect the last mile of automation. That's where the first protocol is going to be. It's only valuable if you're saving time, so there's the proof that has to be there. The agent did all of this work, and now it connected. Do I save time? Do I save money in having this last mile automated as well?
BTN: For those who are managing payment at their companies and might be overwhelmed at the rate change is happening, what steps should they be taking?
Jamison: You're also talking to the most conservative audience inside a company. Finance people know where their challenges are. They're trying to solve for those problems, as opposed to taking technology and saying, 'Cool, where can I apply this to a problem, and when can I root out a problem because of this technology?'
There's quite a lot in inertia in this space because there's a lot of sunk cost. Here's why I don't think it goes anywhere in a massive hurry. If you look at the U.S. and the U.S, is a particular beast in the world, how many checks do we still have in the U.S.? The answer's clearly moving to digital forms of payment—there's a massive advantage—but what people forget is, there's so much sunk cost in the technology that digitizes a paper process.
I [get] checks today from insurance companies, but I didn't step into a bank. I'd go to my phone, I took a photo, the boom checks the deposit. I'm done. It might take three days to get to be in the mail and whatever else, but given the size of money and the frequency, it's actually not a problem.
You tend to find innovations take hold faster when there is a degree of a mandate federally, but we know the U.S. doesn't like to mandate things at a federal level. There's a reason why U.K. has no real checks left, because they put a law in place and said we're getting rid of these things. So, guess what? They disappeared. We will see innovations happening faster in places where there is more of an appetite to start to mandate changes. While the technology tends to be invented in the U.S., I don't think necessarily that's where it goes first.