Spending limits are not enough: what makes Lowtides different
A spending limit protects your budget, but it does not check the store, respect the store's rules or cover things like bookings and sign-ups. Lowtides adds a check on both sides, so the AI shopper pays only when you and the store both say yes.
Key takeaways
- Spending limits guard your wallet. They do not look at the store or at what else the AI does.
- Each common control has a real strength and a real gap. Most people end up stacking two or three of them.
- A both-sides check adds what is missing: the store's own rules, a check that the store is real, a job scope, one off switch and a signed receipt for both sides.
- Lowtides is a hackathon demo. Stores are fictional and money is PayPal sandbox test money.
Why a spending limit is only half a control
More people are letting AI tools shop, book and sign up for things. That is handy. It also raises a plain question: how do you keep the AI from spending more than you meant, or from spending in the wrong place?
Most answers today are about the buyer. Put a cap on the card. Set a budget in the app. Make the AI ask first. Those are all good ideas, and we use some of them ourselves. But they all look at one side of the deal. They ask, "Is this okay for me?" They do not ask, "Is this okay for the store?" or "Is this store even real?"
Think of a door with a lock on your side only. You can keep your own things safe. You cannot tell who is on the other side, and the other side cannot tell you what it will and will not accept from a robot. That is the gap this post is about.
We will walk through four common approaches. We are not naming any company or product. For each category, we say what it does well and where it stops.
Four common ways people control an AI that spends
1. A prepaid or virtual card with a cap
What it does well. This is the simplest control, and it works. If you load a card with $100, the AI cannot spend $101. The worst case is known before you start. You do not have to trust the AI to follow a rule, because the card enforces the limit. It also keeps your main bank account out of reach.
The gap. A cap is a blunt tool. It knows an amount, not a purpose. If the AI is fooled into sending the whole $100 to a fake store, the card happily pays. The cap limits the loss, but it does not stop the loss. The card also knows nothing about the store's own rules for AI shoppers, and it cannot tell you the store was a lookalike until the money is gone. It does not help at all with an AI that books a table, signs you up for a newsletter or shares your phone number. Those actions may cost nothing, so no cap ever trips.
2. A budget or spend limit set inside the AI tool
What it does well. This one is easy to set and easy to understand. You type a daily or monthly number, and the tool tries to respect it. Some tools also let you pick categories. It fits right into the place where you already use the AI, so you are more likely to use it.
The gap. The limit lives inside the same software that you are trying to control. If the AI is tricked, confused or simply wrong, the setting may not hold. Security researchers warn about this kind of failure. The group OWASP lists prompt injection as the first entry in its 2025 top ten risks for large language model apps, and describes it as outside text that changes what the model does in ways nobody intended (OWASP, LLM01 Prompt Injection, 2025). A budget that the model itself is asked to follow is only as strong as the model's resistance to being talked out of it. A better design checks the rule in plain code, outside the model.
A tool setting also tends to be one-sided. It reflects what you want, and it ignores what the store allows.
3. "Ask me before every purchase" prompts
What it does well. This gives you the last word. Nothing is bought without a tap from you. It matches the advice from security experts, who say a person should confirm high-impact actions. OWASP makes the same point for AI apps, recommending human approval for privileged operations (OWASP, LLM01 Prompt Injection, 2025).
The gap. Two problems show up fast. First, people get tired. If every $4 coffee filter asks for a tap, you start tapping without reading. The prompt turns into a habit, not a check. Second, the prompt only shows what the AI chooses to show. If the page says "Harbor Hardware" and the real store is a lookalike, you may approve a purchase you would have refused with more facts. And the prompt does not know the store's rules. You can say yes to something the store would have blocked for an AI shopper.
Ask-first is a good safety net. It works best when it fires rarely, only on the purchases that need a human, and when it comes with the facts you need to decide.
4. Trusting the store's own fraud checks
What it does well. Stores and payment companies have years of experience with fraud. Their systems look at patterns across millions of orders, such as odd shipping addresses or sudden bursts of orders. You get this protection without doing anything.
The gap. Those systems were built for a human at a keyboard. A shopper that is an AI agent looks different, and a store may not know how to treat it. Also, the store's fraud tools protect the store, not you. They protect you only when you are the one being impersonated, and not when you are paying a fake store. By then the store has your money and no reason to help.
The scale of the problem is real. The U.S. Federal Trade Commission reported that consumers told it about more than $12.5 billion in fraud losses in 2024, up 25% from 2023, and it ranked online shopping as the second most reported fraud category by number of reports (FTC, 2025). Impersonation scams alone made up $2.95 billion of those losses. Fake stores and fake sellers are part of the world an AI shopper walks into.
What a both-sides consent check adds
Lowtides is built on one idea: both sides say yes, or no money moves. Your AI shopper carries a signed set of your rules, called a Pass. Each store publishes its own signed rules for AI shoppers, called a Store Policy. One check, the Handshake, reads both and gives a single answer.
Here is what that adds on top of the four controls above.
The store's rules count too
A store can say how it treats AI shoppers. For example, it can set the largest order an AI may place without a person, the product categories it accepts, the minimum trust score it wants, or which agents it blocks. A card cap or an app setting can never see those rules. In Lowtides the store signs them and publishes them. The Handshake reads them. If your purchase is fine for you but over the store's limit for AI shoppers, it turns yellow and waits for a person. Both sets of rules matter, and the stricter one wins. The Handshake can only make a payment harder than your own rules, never easier.
The store is checked before any money moves
This is the job of GuardPooch, the watchdog. It looks for signs of a fake store: a name that is a near copy of a well-known one, a very new domain, no signed rules, or prices that are far too good. In the demo, a lookalike hardware store with a misspelled name, a six-day-old domain and prices about 62% too low gets a red light, and nothing is paid. A cap on a card cannot do this. It only counts dollars.
The job scope covers more than purchases
Many things an AI does for you are not purchases. It may book, call, sign up, cancel, send a message or share details. Lowtides calls this check Action Muzzle. You say the job in plain words, such as "book a table for two on Friday at 7." That becomes the scope, and the rule is simple: do only what was asked. If the AI tries to add a mailing-list sign-up, it has to ask. Each business publishes its own signed rules for AI requests, and details like a Social Security number, a password or a bank account number are never shared. You can try it at the Action Muzzle demo.
This matches a point from security guidance. OWASP calls the problem excessive agency, meaning an AI system that can take damaging actions. It names three root causes: too much functionality, too many permissions and too much autonomy (OWASP, LLM06 Excessive Agency, 2025). Its fixes include giving only the permissions a task needs and checking authorization in the downstream system instead of trusting the model to decide. A job scope and a store-side policy are two ways to do that.
One off switch works everywhere
With a card cap, you cancel the card. With app settings, you change the setting. With a prompt, you stop answering. None of these reach every place the AI has access. Lowtides has Stop all, one button that revokes the Pass at every store at once. After you press it, every Handshake turns red. Open holds are voided and unclaimed payouts are canceled. For bookings and sign-ups there is also a stop word you choose. Say it, and everything freezes until you release it.
Both sides get a signed receipt
Every decision, green, yellow or red, leaves a signed receipt. It shows what was asked, which rules were checked and what happened. Anyone with the public keys can verify it. That gives you proof if something goes wrong, and it gives the store proof too. Most other controls leave only a card statement, which shows the amount and nothing about why.
How the approaches compare
This table sums it up. "Yes" means the approach does the job by design. "Partly" means it helps in some cases. "No" means it does not address it. These are general categories, and any one product may do more or less than a category suggests.
| Approach | Protects the shopper's budget | Respects the store's rules | Checks the store is real | Covers bookings and sign-ups | One off switch | Signed proof for both |
|---|---|---|---|---|---|---|
| Prepaid or virtual card with a cap | Yes | No | No | No | Partly (cancel the card) | No |
| Budget set inside the AI tool | Yes, if the AI follows it | No | No | No | Partly (that tool only) | No |
| "Ask me before every purchase" | Partly (you decide each time) | No | No | Partly | No | No |
| The store's own fraud checks | No | Partly (its own rules) | No | No | No | No |
| Both-sides consent (Lowtides demo) | Yes | Yes | Yes (GuardPooch) | Yes (Action Muzzle) | Yes (Stop all) | Yes |
You do not have to pick one. A capped card is still a smart backstop, and ask-first still has a place. The point is that those controls cover one side, and the row at the bottom covers both.
What the numbers say
We checked a few public sources to see how much these gaps matter. Here is what we found, with the links so you can read them yourself.
People want help, but not a blank check
Accenture surveyed 25,590 consumers in 16 countries in January 2026. It found that 32% would let an AI agent decide what to buy while the person makes the payment. Only 9% would let an agent buy on its own within set limits (Accenture, Talk to My AI Agent, 2026). The same page says payments are the highest hurdle to full automation, and it points to configurable permissions, instant override and clear recourse as things that could help. Those are close to the features in this post.
Online fraud is large and growing
The FTC figures above are worth a second look. More than $12.5 billion in reported fraud losses in 2024, up 25% in a year, means that bad actors are finding people. An AI shopper that cannot tell a real store from a fake one will meet them too (FTC, 2025). The FTC also notes that the rise came from a higher share of people who reported fraud losing money, not from more reports. That is why a check made before paying matters more than a check made after.
Paying online has always carried more risk than paying in person
The Federal Reserve studied card fraud using 2015 data. It found a fraud rate by value of 14.23 basis points for card-not-present payments (online and phone), compared with 9.32 basis points for card-present payments, which in that study include ATM withdrawals (Federal Reserve, 2018). A basis point is one hundredth of a percent, so 14.23 basis points is about 14 cents per $100. The data is older, and fraud patterns have changed, so treat it as a sign of direction and not as a current rate. But it shows why the place where a card is typed into a web page is a weak point. An AI shopper does all of its shopping there.
What Lowtides does not do
We want to be plain about the limits. A good safety tool says what it cannot do.
- It is a demo. Lowtides is a hackathon project. It is not a product you can use to shop in real life today.
- The money is test money. A green Handshake pays through PayPal sandbox. Nothing real moves. When the server has no sandbox keys, the payment is simulated and labeled that way on the receipt.
- The stores are fictional. No real store has joined, and nothing real is contacted. The bookings in Action Muzzle record what would have been done and give back a demo confirmation number.
- An agent that holds its own card can go around it. Lowtides only works when the AI shopper goes through the Handshake. If an agent has a payment card of its own, it can pay without asking. That is why we suggest pairing a Pass with a capped card, not replacing one with the other.
- The trust score and the watchdog are heuristics. They are rules of thumb, not proof. GuardPooch can miss a clever fake and can flag a real store by mistake. A new honest store might look risky because its domain is young.
- Stop all does not undo what is done. It cancels what has not gone through. It cannot take back a payment that already settled.
Try it yourself
You do not need a password or a card. Open the shopper panel, search for a drill, and watch the AI shopper list items with a green, yellow or red light. Pick the lookalike store and see GuardPooch turn it red. Then open the store control center and lower the store's no-human limit. The same purchase turns yellow, because the store's rules count as much as yours. For the full picture, the live Handshake shows both sides checked line by line. If you are still deciding how much to trust AI shopping, our guide Is AI shopping safe? is a good place to start.
Sources
- OWASP, LLM01:2025 Prompt Injection. Lists prompt injection first in the 2025 top ten for LLM apps, and recommends least privilege and human approval for privileged operations.
- OWASP, LLM06:2025 Excessive Agency. Names excessive functionality, permissions and autonomy as root causes; recommends least privilege, human approval and checks in downstream systems.
- FTC, new data show big jump in reported losses to fraud, March 2025. More than $12.5 billion reported in 2024, up 25%; impersonation scams $2.95 billion; online shopping second most reported category.
- Federal Reserve, Changes in U.S. Payments Fraud from 2012 to 2016, 2018. Card-not-present fraud rate of 14.23 basis points versus 9.32 for card-present (2015 data).
- Accenture, Talk to My AI Agent: The New Rules of Brand Value, 2026. Survey of 25,590 consumers in 16 countries, January 2026; 32% and 9% figures.
Common questions
Is a prepaid card with a cap not enough to keep an AI shopper safe?
A capped card is a good backstop because it limits the worst case. But it only counts dollars. It does not check whether the store is real, follow the store's rules for AI shoppers, or cover bookings and sign-ups that cost nothing. Lowtides adds those checks and works well alongside a capped card.
What does consent from both sides mean?
Your AI shopper carries your signed rules, called a Pass. The store publishes its own signed rules for AI shoppers, called a Store Policy. One check, the Handshake, reads both. If both say yes, the payment goes ahead. If one needs a person, it waits. If a rule says no, no money moves.
Can an AI agent go around Lowtides?
Yes, if it holds its own payment card or has another way to pay. Lowtides only controls shoppers that go through the Handshake. This is one of its honest limits, and it is why a capped card is still a smart extra layer.
Is Lowtides a real payment service?
No. It is a hackathon demo. The stores are fictional and the money is PayPal sandbox test money. When the server has no sandbox keys, payments are simulated and the receipt says so. It is patent pending and not a finished product.
What does Action Muzzle do?
It applies the same two-sided check to things an AI does besides buying, such as booking, calling and signing up. You state the job in plain words, and the AI may do only that. Anything extra has to ask first. A stop word you choose freezes everything, and every action leaves a signed receipt.