Sunday afternoon, in the middle of drafting a blog post, I mentioned to my assistant that I'd like a copy of Schwarzenegger's "The Education of a Bodybuilder." (The relevance to what follows will not be immediately evident . . . ) What happened next is, I think, the best explanation I can give you of what a personal AI agent actually is — and why I'm not worried about it . . . much!
Here is the whole thing, exactly as it happened.
1. The setup. I asked how it would work. My assistant — Sunny by Meta — said it was an easy one-time setup: connect a payment wallet, and then it could check out for me. It recommended Stripe's Link service, explaining that after I approve a purchase, Link hands the merchant a one-time virtual card instead of my actual card number. I clicked the link, added my card on Stripe's own secure page — My assistant did not view the number — done.
Control: The setup grants capability, not permission. Connecting a wallet doesn't authorize spending; it only makes spending possible.
2. The search. Sunny reported to have checked several bookstores and came back with three options: two new copies at $19.99 (Barnes & Noble and Powell's) and a used one at Powell's for $9.95. It laid them out.
Control: The agent proposes; the human disposes. Nothing went into the cart until I decided.
3. The choice. I picked the used one. Ten bucks is ten bucks. This would be my third or fourth copy of the book.
Control: The agent only provided options requested, without further comment. I select.
4. The checkout. Sunny "went to" Powell's, put the used copy in the basket, and proceeded to checkout f— guest checkout, no account created, no email-list signup, no extras. Then it came back to me, needing needed my shipping address--which it didn't have. I provided it.
Control: it asked directly rather than guessed the shipping address. It probably had my home address already.
5. The approval. Sunny displayed the order: one used copy, $9.95, plus $5.99 shipping, no tax — $15.94 total, shipping to my address, paying with my card, and then asked: Shall I place the order?
Control: This is the load-bearing moment. Every purchase gets its own, individual, explicit approval. There is no standing permission to spend my money. (Although that can be arranged.)
6. The catch. I looked at the card displayed — a Visa ending in XXXX— it was the wrong card. Sunny provided a secure link to update the card in my Link wallet. I swapped in the right card and gave the go ahead.
Control: This is the whole game. Me checking the work based on a bug showing up in the system; it IS the system. I caught the error because the arrangement required I be shown all the details before anything moved.
7. The order. Sunny verified that the cart matched what I'd approved — same book, same condition, same price, same address — confirmed it, and placed the order. The confirmation email arrived within minutes. The book is on its way.
Footnote. During the search, Sunny tried Amazon, hit Amazon's wall against AI agents, and told me so — then reported that it "took our money to bookstores that wanted it." (Sunny's ear for colloquial English is still a work in progress!)
Control: Finding a viable option that I later approved.
The agent does the legwork, surfaces every decision, and never moves money without an explicit approval. Control never left my hands; it was never the agent's to take: only mine to grant--one approval at a time.
If you're wondering whether you could trust one of these things or if you have any reason to, don't start with the technology. Start with the arrangement. Ask three questions:
- What time-consuming task or tasks could an agent take off my hands?
- What safe, no-risk task can I give it as an initial test?
- How transparent is the agent on all stages of the process, including final approval?
That's the story. No magic, no leap of faith, no deus ex machina — just an arrangement.
wracton@gmail.com
williamacton.legalshieldassociate.com
No comments:
Post a Comment