Is it safe to let AI shop for you?
Short answer: it can be, if you keep the keys. An AI shopper is safe to the degree that it cannot spend outside limits you set, cannot buy from a store that has not agreed to its terms, and can be stopped in one tap.
What can actually go wrong
An AI shopping agent is software that searches, compares and checks out for you. That is useful, and it also creates four new ways to lose money or privacy.
- It buys the wrong thing, or too much. Agents misread instructions. A loop can repeat the same order.
- It gets tricked. A web page can hide text that tells the agent to do something you never asked for. Security people call this prompt injection.
- Someone pretends to be your agent, or pretends to be a store. Banks have warned that attackers could impersonate agents and merchants, and that scams and payment disputes could rise as agents start paying (Bitdefender summary of the banks' warning).
- Nobody is sure who pays for a mistake. The rules on liability for agent purchases are still unsettled (Bloomberg Law).
Shoppers feel this. Trade press coverage in 2026 keeps finding the same pattern: people are glad to let AI help them choose, and far less willing to let it pay (PYMNTS).
Seven controls to insist on
Before any agent can spend for you, check that you have these. If a tool cannot show you one of them, treat that as a no.
- A hard budget. A daily cap and a per-purchase cap that code enforces, not a suggestion the AI can talk itself past.
- An ask-me line. Above an amount you choose, the agent must stop and wait for your yes.
- A never list. Categories that can never be bought, such as gift cards or crypto, which are common scam targets.
- Proof of who the agent is. The store should be able to check that the agent really carries your permission, and what its limits are.
- The store's consent too. A store should be able to say which agents it accepts and on what terms. If either side says no, nothing is charged.
- A receipt for every decision. Paid, held, blocked and refunded, each with the reason, so you can check later.
- A stop switch. One tap that stops every agent and cancels what has not gone through yet.
The rule that matters most: code decides, not the AI
The safest design keeps the AI away from the final decision. The AI can search, compare and suggest. Whether money moves should be decided by plain code checking fixed rules that you approved. That way a clever prompt, a hidden instruction on a web page or a confused model cannot widen its own permission.
How Lowtides does it
Lowtides is an open source demo of this idea, which we call Consent Shopping. Your AI shopper carries a signed Pass with your limits. The store publishes a signed policy. A Handshake compares the two and returns one of three signals: green (both said yes, it pays), yellow (it waits for a person), red (no money moves). You can pick how much freedom the shopper gets: Off, AI Assist or Full Autopilot. The demo runs on the PayPal sandbox with test money and fictional stores, so you can try it without risk.
Practical habits, whatever tool you use
- Start with the agent in an ask-first mode. Give it more freedom only after you have read a week of receipts.
- Give it the smallest budget that does the job.
- Use a payment method with dispute protection.
- Read what the provider stores about your shopping conversations.
Common questions
Is it safe to let an AI agent buy things for me?
It can be if the agent is held to limits that code enforces: a budget, an approval line, a never list, a receipt for every decision and a stop switch. Without those controls, it is not safe to hand an agent a way to pay.
What is prompt injection in AI shopping?
Prompt injection is when a web page or message hides instructions meant for the AI agent, such as telling it to buy something else or pay a different seller. The defense is to let fixed rules in code, not the AI, decide whether money moves.
What should an AI shopping agent never be allowed to buy?
Most people block gift cards and crypto, because scammers favor them and they are hard to reverse. A good tool lets you write your own never list.