Click here to watch a video of this episode.
Four personal AI agents landed in seven weeks, and three of them can already spend your money. Meta's help page puts the liability in eleven words, and Instinct's terms bind you as if you had made the purchase yourself. Who pays when an AI agent gets it wrong, and why does every one of them arrive with a face?
In seven weeks, Meta launched Muse (8 September), SpaceXAI put Grok Bot into beta, OpenAI unveiled dots (29 September) and Instinct raised $1 billion at a $10 billion valuation. Three of those AI agents can spend today, with Stripe's Link wallet behind Muse, Grok Bot and Instinct. The twist is in the contracts: on paper the user pays, the bot never does, and under US rules the bank may end up writing the refund.
Aman Narain and Zubin Vandrevala break down who pays when AI agents make mistakes, who moves the money behind them, and why Meta, OpenAI and Instinct gave their AI agents a face while Google and Anthropic did not.
Key takeaways:
1. On the contract, you pay: Meta and Instinct both put liability for an AI agent's purchases on the user, and in no version does the bot pay.
2. A bot's silence is the new pending: the failure that nearly cost Aman a meeting was a confirmation that never came, not a wrong charge.
3. Muse, Grok Bot and Instinct all pay through Stripe's Link wallet, yet the power sits at the two ends: the bot you trust and the shop with the goods.
4. Amazon blocked Muse twelve days after launch, and Sierra and Meta's answer is a standard that says who is knocking but not who pays.
5. Before an AI agent gets your card, ask three things: does it ask first, what is the cap, and who do you call.
Topics covered:
- Muse, Grok Bot, dots and Instinct: which AI agents can spend, and which cannot yet
- The two clauses that decide liability: Meta's help page and Instinct's section three
- Why silence, not a wrong charge, is the failure mode
- How an agent purchase works: request, approve, issue, pay
- Regulation E and the bank as the unwitting insurer
- Amazon's block of Muse, and the Personal Agent Protocol from Sierra and Meta
- India's NPCI and its Unified Agent Protocol
- Why Meta, OpenAI and Instinct gave agents a face, and why Google and Anthropic did not
- Ask, cap, call: three questions before an AI agent gets your card
Chapters:
Referenced in this episode: Meta Muse launch and help page; Stripe Link wallet for agents; Instinct Series C; Grok Bot beta; OpenAI DevDay and dots; Amazon's block of Muse; Sierra and Meta Personal Agent Protocol; NPCI Unified Agent Protocol; Regulation E; White House executive order on "Super Intelligence"; Microsoft Copilot's Mico.
Related episodes: Google I/O 2026 (Gemini Spark); Stripe and the Money OS [TBD: title and link]; S2E28, Revolut [TBD: link].
Hosted by:
Aman Narain writes at amanwhoblogs.substack.com. Zubin Vandrevala is your payments provocateur.
Enjoying A2Z Fintech? Leave a rating and review on Apple Podcasts. It is the single biggest signal to the Apple algorithm and how new listeners in our world find us.
For information and entertainment only. Not financial advice.
Transcript:
Four personal AI agents landed in seven weeks, and three of them can already spend your money. Meta's help page puts the liability in eleven words, and Instinct's terms bind you as if you had made the purchase yourself. Who pays when an AI agent gets it wrong, and why does every one of them arrive with a face?
In seven weeks, Meta launched Muse (8 September), SpaceXAI put Grok Bot into beta, OpenAI unveiled dots (29 September) and Instinct raised $1 billion at a $10 billion valuation. Three of those AI agents can spend today, with Stripe's Link wallet behind Muse, Grok Bot and Instinct. The twist is in the contracts: on paper the user pays, the bot never does, and under US rules the bank may end up writing the refund.
Aman Narain and Zubin Vandrevala break down who pays when AI agents make mistakes, who moves the money behind them, and why Meta, OpenAI and Instinct gave their AI agents a face while Google and Anthropic did not.
Key takeaways:
1. On the contract, you pay: Meta and Instinct both put liability for an AI agent's purchases on the user, and in no version does the bot pay.
2. A bot's silence is the new pending: the failure that nearly cost Aman a meeting was a confirmation that never came, not a wrong charge.
3. Muse, Grok Bot and Instinct all pay through Stripe's Link wallet, yet the power sits at the two ends: the bot you trust and the shop with the goods.
4. Amazon blocked Muse twelve days after launch, and Sierra and Meta's answer is a standard that says who is knocking but not who pays.
5. Before an AI agent gets your card, ask three things: does it ask first, what is the cap, and who do you call.
Topics covered:
- Muse, Grok Bot, dots and Instinct: which AI agents can spend, and which cannot yet
- The two clauses that decide liability: Meta's help page and Instinct's section three
- Why silence, not a wrong charge, is the failure mode
- How an agent purchase works: request, approve, issue, pay
- Regulation E and the bank as the unwitting insurer
- Amazon's block of Muse, and the Personal Agent Protocol from Sierra and Meta
- India's NPCI and its Unified Agent Protocol
- Why Meta, OpenAI and Instinct gave agents a face, and why Google and Anthropic did not
- Ask, cap, call: three questions before an AI agent gets your card
Chapters:
- (00:00) - Cold open: Maya, Roxy and a $1,200 room, Austin
- (01:38) - Disclaimer
- (02:32) - Bot one: Meta's Muse and its plush face, Jolly
- (04:06) - A creature with a chaperone: Muse and Sentinel
- (04:42) - Bot two: Grok Bot and the staff room
- (06:19) - Bot three: OpenAI's dots and a frog called Todd
- (08:29) - Bot four: Instinct, the agent that lives in your texts
- (09:58) - The fine print: eleven words on a help page
- (10:23) - Instinct's section three: a power of attorney for a bot
- (11:13) - Pending: the most expensive word in finance
- (12:07) - So who pays? The contract, Regulation E and the bank
- (13:06) - Under the bonnet: four steps from request to payment
- (13:47) - The bank cannot tell a bot from a human
- (14:32) - One wallet, three rivals: Stripe's Link
- (14:58) - John Collison and the founder-led lunch
- (16:03) - The Money OS is rented, not owned
- (17:28) - Amazon blocks Muse, twelve days in
- (18:14) - Sierra's Personal Agent Protocol: who is knocking?
- (19:56) - The complaints desk: whoever pays makes the rules
- (20:52) - India builds the road first: NPCI's agent protocol
- (21:31) - Why every bot gets a face (and two don't)
- (23:20) - Is AI going social to grow? Charm is not a control
- (24:35) - When the staff room gets bodies: Optimus
- (25:29) - Ask, cap, call: three questions before a bot gets your card
- (26:30) - Closing wagers: the friend or the contract?
- (27:06) - Mico, Clippy and the costume party
- (27:55) - Casting a soul for your robot: tell me who pays
Referenced in this episode: Meta Muse launch and help page; Stripe Link wallet for agents; Instinct Series C; Grok Bot beta; OpenAI DevDay and dots; Amazon's block of Muse; Sierra and Meta Personal Agent Protocol; NPCI Unified Agent Protocol; Regulation E; White House executive order on "Super Intelligence"; Microsoft Copilot's Mico.
Related episodes: Google I/O 2026 (Gemini Spark); Stripe and the Money OS [TBD: title and link]; S2E28, Revolut [TBD: link].
Hosted by:
Aman Narain writes at amanwhoblogs.substack.com. Zubin Vandrevala is your payments provocateur.
Enjoying A2Z Fintech? Leave a rating and review on Apple Podcasts. It is the single biggest signal to the Apple algorithm and how new listeners in our world find us.
For information and entertainment only. Not financial advice.
Transcript: