The current purchasing agents (Muse and Dots) will not survive.
In the end, Apple and Google will be the agents as part of Siri/iOS and Gemini/Android.
Apple Siri has access to all the personal information on your phone and can pay with Apple pay which solves the problem you noted.
Google Android can get access to everything on an Android phone and have the user pay with Google pay.
Consumers are already comfortable using Apple Pay and Google Pay. I own a bakery with multiple locations. Our web site accepts Apple Pay and Google Pay.
I get why Meta, OpenAI, Shopify, Stripe, Mastercard, Link, Shop Pay, PayPal, etc are trying jump into the game. I'm guessing medium and large enterprise, will develop their own version of Muse. But in the end, people do most of their shopping on an iphone/mac or Android/Chrome browser platform.
Your article is an evaluation of AI as a product that companies are trying to use to develop a niche. AI is not a product, its a technology. Apple's implementation of Siri is the right answer assuming they can get it to work (which I'm pretty confident they can). Google will do the same with Gemini on Android although they will have a challenge on promoting privacy given they are behind on a hardware platform that can support local AI.
In the mean time, purchasing agents will be a wild west experiment that Apple and Google will use to tune their implementation.
One additional point, Muse works fine with my web site. Including with users who already have an account and credit card on my web site. Users will want to log into my web site so they can get loyalty points. Can Muse (like Doordash et al) create a loyalty system that users prefer over my loyalty system. That's a more interesting question. Apple currently does not have a loyalty system using Apple Pay and neither does Google using Google Pay. I personally believe the war on where loyalty rewards accrue is the more interesting topic.
My main concern would be the fact that advertising space could be bought within these agents. For example, if you search for the best mortgage deals, would the agent recommend the genuinely best option, or prioritise a company with high SEO rankings or one that has paid to appear in its recommendations? In the long run, this could distort competition and potentially push smaller businesses out of the market.
Take Coca-Cola as an example. It’s the same product sold at different prices across shops and supermarkets. If you were comparing options, you’d probably choose the cheapest one, rather than the supermarket that paid the most for advertising. I think the challenge is ensuring that AI agents genuinely act in the consumer’s best interests rather than the interests of whoever pays them the most.
Your point about using different cards for different purposes, such as Amex for flights and a Clubcard for supermarket points, is also a good one. There’s still a lot of functionality that would need to be coded and developed. However, given how quickly AI is advancing, I don’t think we’ll have to wait too long for these solutions to become a reality.
Regarding mistakes, I suspect we’ll still carry some liability, depending on the provider’s terms and conditions. It’s similar to asking your assistant (human 😂) to book a hotel: if they book the wrong dates, you’re still the one who has to sort it out and hope the hotel will refund you. I’m not sure you’d necessarily expect your assistant to reimburse you personally for the mistake. The same question applies to AI agents: where does the responsibility lie when something goes wrong?
Just some additional thoughts. Really interesting article, and apologies for the long comment!
"Buy this" is about to become a miniature constitution. We spent years making checkout one click; agents are putting the committee back inside the click.
The hotel case is an approval-binding failure, not only a liability question. Authorizing "spend $500 on a hotel" lets the agent fill in the dates, and the hotel then delivers exactly the agent's order, which was never checked against what the user meant. A policy that only bounds the amount will pass that. I'd want the approval to name which fields may vary, and to keep the user's instruction next to the booking that actually went out, so a dispute can show substitution rather than a completed purchase.
Energy solved this decades ago with fleet fuel cards. Diesel only, capped volume, approved stations, odometer check. Driver as agent, company as cardholder. The token was never the problem. The fight was always over who writes the policy and routes the volume, because that party keeps the margin.
Loyalty rewards will be a big factor in how agents transact, and who they transact with. Middlemen, for example, are vulnerable to getting cut out by agents who make an end run around them for better pricing or rewards.
The split between helping my agent pay and letting it transact as me is the part I hadn't separated before. Loyalty is a good test of it, because points only follow the person the merchant believes is on the other side.
Terry, excellent distinction between enabling an agent to pay and enabling it to transact on someone's behalf.
There's a third distinction worth examining: a valid credential and an initially authorized mandate do not necessarily establish that authority remains valid at the moment of execution.
What happens when the user's instructions change, authority is partially consumed or revoked, or an agent selects a payment route that satisfies the transaction but violates the user's underlying conditions?
Tokenization establishes important controls, but the broader challenge is maintaining authority continuity across agents, wallets, merchants and payment intermediaries—with a reconstructable record of what was authorized, what changed and why execution proceeded.
This is the problem we're exploring through VNet™ at Persistence Analytics Group.
Authority before execution. Evidence after the event.
The current purchasing agents (Muse and Dots) will not survive.
In the end, Apple and Google will be the agents as part of Siri/iOS and Gemini/Android.
Apple Siri has access to all the personal information on your phone and can pay with Apple pay which solves the problem you noted.
Google Android can get access to everything on an Android phone and have the user pay with Google pay.
Consumers are already comfortable using Apple Pay and Google Pay. I own a bakery with multiple locations. Our web site accepts Apple Pay and Google Pay.
I get why Meta, OpenAI, Shopify, Stripe, Mastercard, Link, Shop Pay, PayPal, etc are trying jump into the game. I'm guessing medium and large enterprise, will develop their own version of Muse. But in the end, people do most of their shopping on an iphone/mac or Android/Chrome browser platform.
Your article is an evaluation of AI as a product that companies are trying to use to develop a niche. AI is not a product, its a technology. Apple's implementation of Siri is the right answer assuming they can get it to work (which I'm pretty confident they can). Google will do the same with Gemini on Android although they will have a challenge on promoting privacy given they are behind on a hardware platform that can support local AI.
In the mean time, purchasing agents will be a wild west experiment that Apple and Google will use to tune their implementation.
One additional point, Muse works fine with my web site. Including with users who already have an account and credit card on my web site. Users will want to log into my web site so they can get loyalty points. Can Muse (like Doordash et al) create a loyalty system that users prefer over my loyalty system. That's a more interesting question. Apple currently does not have a loyalty system using Apple Pay and neither does Google using Google Pay. I personally believe the war on where loyalty rewards accrue is the more interesting topic.
My main concern would be the fact that advertising space could be bought within these agents. For example, if you search for the best mortgage deals, would the agent recommend the genuinely best option, or prioritise a company with high SEO rankings or one that has paid to appear in its recommendations? In the long run, this could distort competition and potentially push smaller businesses out of the market.
Take Coca-Cola as an example. It’s the same product sold at different prices across shops and supermarkets. If you were comparing options, you’d probably choose the cheapest one, rather than the supermarket that paid the most for advertising. I think the challenge is ensuring that AI agents genuinely act in the consumer’s best interests rather than the interests of whoever pays them the most.
Your point about using different cards for different purposes, such as Amex for flights and a Clubcard for supermarket points, is also a good one. There’s still a lot of functionality that would need to be coded and developed. However, given how quickly AI is advancing, I don’t think we’ll have to wait too long for these solutions to become a reality.
Regarding mistakes, I suspect we’ll still carry some liability, depending on the provider’s terms and conditions. It’s similar to asking your assistant (human 😂) to book a hotel: if they book the wrong dates, you’re still the one who has to sort it out and hope the hotel will refund you. I’m not sure you’d necessarily expect your assistant to reimburse you personally for the mistake. The same question applies to AI agents: where does the responsibility lie when something goes wrong?
Just some additional thoughts. Really interesting article, and apologies for the long comment!
"Buy this" is about to become a miniature constitution. We spent years making checkout one click; agents are putting the committee back inside the click.
The hotel case is an approval-binding failure, not only a liability question. Authorizing "spend $500 on a hotel" lets the agent fill in the dates, and the hotel then delivers exactly the agent's order, which was never checked against what the user meant. A policy that only bounds the amount will pass that. I'd want the approval to name which fields may vary, and to keep the user's instruction next to the booking that actually went out, so a dispute can show substitution rather than a completed purchase.
Energy solved this decades ago with fleet fuel cards. Diesel only, capped volume, approved stations, odometer check. Driver as agent, company as cardholder. The token was never the problem. The fight was always over who writes the policy and routes the volume, because that party keeps the margin.
Loyalty rewards will be a big factor in how agents transact, and who they transact with. Middlemen, for example, are vulnerable to getting cut out by agents who make an end run around them for better pricing or rewards.
Interesting read.
The split between helping my agent pay and letting it transact as me is the part I hadn't separated before. Loyalty is a good test of it, because points only follow the person the merchant believes is on the other side.
Excellent posts. We’re working on agentic ID solutions for wallet transactions at @soulboundsecurity
Terry, excellent distinction between enabling an agent to pay and enabling it to transact on someone's behalf.
There's a third distinction worth examining: a valid credential and an initially authorized mandate do not necessarily establish that authority remains valid at the moment of execution.
What happens when the user's instructions change, authority is partially consumed or revoked, or an agent selects a payment route that satisfies the transaction but violates the user's underlying conditions?
Tokenization establishes important controls, but the broader challenge is maintaining authority continuity across agents, wallets, merchants and payment intermediaries—with a reconstructable record of what was authorized, what changed and why execution proceeded.
This is the problem we're exploring through VNet™ at Persistence Analytics Group.
Authority before execution. Evidence after the event.
Neil P. Osnato
Founder | Persistence Analytics Group LLC