AI agents that move money: design rules from crypto
Short answer
When an AI agent can spend money, design it like a crypto transaction. Show the full cost before the commit step, keep spending inside limits the user set, give a receipt for every action, and never let the agent widen its own permissions. Crypto products have handled irreversible payments for years, and those rules carry straight over to agents.
Why is money different for an AI agent?
Because there’s often no undo.
In most AI products we remove confirmations and rely on undo instead. A payment that has settled can’t be taken back by the product. So the usual advice flips. Here the agent stops, shows everything and waits.
What did crypto already work out?
Crypto has been the testing ground for irreversible actions taken through an interface. The mistakes are well known, and each one has an agent equivalent.
| Crypto problem | Agent equivalent | Design rule |
|---|---|---|
| Signing a transaction you can’t read | Approving an action described vaguely | Say what will happen in plain words |
| Fees that appear late | Cost shown after the commit | Full cost before the button turns on |
| Unlimited token approvals | Standing permission to spend | Limits by amount, time and purpose |
| Sending to the wrong address | Paying the wrong party | Show who gets paid, and confirm new payees |
How do you design an agent that can spend?
1. Show the full cost before the commit step
This is a rule in our design foundation. A cost that turns up late feels like a trick, and it’s a common reason people abandon. Wrong looks like a swap screen that adds the network fee after you’ve tapped Confirm. Right shows what you pay, what you receive, each fee and one total before the button is live.
2. Let the user set the budget
Per action, per day and per purpose. Inside the budget the agent can act. Outside it, the agent asks. The limits live on one screen the user can find and change.
3. Give a receipt for every action
What was paid, to whom, why the agent did it and what it was part of. People forgive an agent that spends as instructed. They don’t forgive a charge they can’t explain.
4. The agent never raises its own limits
It can ask for more. It can’t grant itself more, and it can’t make the request sound urgent.
5. Hold persuasion to a stricter standard
Our foundation has tighter rules for behavioural techniques wherever money is involved. No countdowns, no nudges to spend more, no defaults that favour the product over the user.
What does this look like in real work?
Ulys is a crypto investing app for iOS and Android. We worked with the founders, head of product and engineers for twelve months, from inception to App Store launch. The product puts funding, trading, storage and AI insights in one app, and removes the usual barriers: gas fees, seed phrase anxiety and chain switching. The team raised a $7.7M Series A. Ulys is where these rules for money flows were tested on real users. Read the case study.
Pascal is a prediction market founded by the former CEO of dYdX. Traders are hard to impress, and the product had to read as a serious venue from the first screen. Pascal raised a $15M+ Series A. Read the case study.
How do you check your own product?
- Before the final tap, can the user see every fee and one total?
- Where are the agent’s spending limits set, and who set them?
- Can the user explain every charge from the last week using only the product?
- Is there any path where the agent increases what it’s allowed to do?
Who should design an agent that handles money?
Someone who has shipped trading, wallet or payment flows. The patterns are learned from real transactions, and a designer who hasn’t worked with irreversible actions will reach for undo where there isn’t one.
Halaska is an embedded product design partner for funded AI, crypto and fintech teams. A senior designer joins your team and leads the work day to day. Chris Halaska, our founder and design director, sets the direction and reviews what ships. See crypto product design or get in touch.
FAQ
How should an AI agent ask before spending money? By showing the full cost, the recipient and the reason before the commit step, in plain words, with a single total.
Can an AI agent make payments without approval? Inside limits the user has set, yes. Anything outside those limits, or to a new payee, should stop and ask.
What can AI products learn from crypto UX? How to handle actions that can’t be reversed: readable approvals, fees shown up front, limited permissions and a record of every transaction.
Should an agent be able to change its own spending limits? No. It can request a change. Only the user can make one.
Who is Halaska? Halaska is an embedded product design partner for funded AI, crypto and fintech teams, based in Dubai. It’s founded and directed by Chris Halaska, who has 20+ years designing products, including six at Google.