Start typing to search this publication.
Tankothek
Open menu

Subscribe to Tankothek

Get new posts delivered straight to your inbox.

Would You Give an AI Access to Your Wallet?

When an assistant can spend money, trust comes with a price.

Tankothek
5 min read
Cover image for Would You Give an AI Access to Your Wallet?

Imagine you’re looking for a flight.

Your AI knows your travel dates, departure airport and budget. While you’re working, it finds a suitable option for €420. Good departure times, baggage included, comfortably below your limit.

It could ask you first.

Or it could simply book and pay.

Which would you prefer?

As long as AI only makes suggestions, the final decision remains ours. The moment it is allowed to spend money, its role changes.

A recommendation becomes an action — and that action has real consequences.

The flight example is a thought experiment. The technology behind it is no longer one.

Coinbase, for example, offers Agentic Wallets designed to give AI agents the ability to make payments, trade and operate within defined spending limits.[1]

So perhaps the more interesting question is no longer whether AI will eventually be able to spend money.

It is:

How much control would we actually give it?

Paying a bill: sounds simple enough

Let’s start with an easy task.

A familiar bill is due. The recipient is known, the amount is within an approved range and the AI is supposed to pay it on time.

Amount, recipient and deadline are already defined.

The important decision has essentially been made.

But even here, exceptions appear.

What happens if the bill is suddenly twice as high?

What if the payment address changes?

Or what if the same invoice appears twice?

“Pay my bills” sounds simple.

In reality, a useful assistant would need to recognise when its instructions are no longer clear enough — and when it needs to ask.

Choosing a flight: now it gets personal

A flight is more complicated.

“Book the best flight under €500” sounds fairly specific at first.

The price has a clear limit.

The word “best” does not.

Maybe the cheapest flight leaves at five in the morning.

Maybe a long connection saves €80 but costs you half a day of your holiday.

Maybe flexible rebooking matters more to you than getting the lowest possible price.

An AI could stay perfectly within every financial limit and still book something you would never have chosen yourself.

A spending limit therefore answers only one question:

How much is it allowed to spend?

It does not answer:

What would I have chosen?

Perhaps you would want to see three options first.

Or perhaps the AI should be allowed to book automatically whenever price, departure time, baggage and total travel time all stay within limits you defined beforehand.

Both approaches could work.

The important part is knowing exactly how much freedom the agent has before it acts.

Investing: what exactly is “a good opportunity”?

The problem becomes even clearer with an instruction such as:

“If you find a good opportunity, invest for me.”

What qualifies as a good opportunity?

What is the objective?

How long is the investment supposed to last?

How much risk is acceptable?

And at what loss should the AI exit?

A bill has a known amount.

A flight has characteristics that can be checked before buying it.

An investment has an outcome that remains uncertain.

Even a perfect-looking analysis cannot turn that uncertainty into certainty.

The AI would therefore be making decisions about risks whose financial consequences are ultimately carried by the user.

Someone who would happily give an AI €20 for a clearly defined task may still be nowhere near comfortable letting it manage a substantial part of their savings.

How much access does an AI actually need?

“Giving AI access to my wallet” quickly sounds like an all-or-nothing decision.

But permissions could be divided much more precisely.

Can the AI only see the balance?

Can it prepare a transaction?

Can it spend up to €50 on its own?

Or can it move larger amounts without additional approval?

Coinbase, for example, documents spending limits per transaction call and per session. According to its documentation, the agent must respect those limits and cannot change them itself.[2]

For systems like this, I would want five things to be immediately clear:

  • How much money is available to the AI?

  • Which recipients or actions are allowed?

  • When does it need my confirmation?

  • How long does its permission remain active?

  • How can I stop further access immediately?

Giving an AI a small, separate amount for a specific task is fundamentally different from giving it unrestricted access to everything you own.

There is a very large space between:

“You can do nothing.”

and

“Do whatever you think is best.”

That space may ultimately determine whether people trust these systems.

What happens after the first mistake?

Back to the flight.

The AI booked it.

Only afterwards do you notice that it chose the wrong travel date.

At that point, the explanation “It was under €500” is not particularly useful.

You want to know why that flight was selected, which instructions were considered and whether anything can still be corrected.

A kill switch might stop the next action.

It does not automatically reverse a payment or purchase that has already happened.

That is why the discussion around autonomous payments needs to go beyond convenience.

Users need decisions they can understand, visible permissions, clear limits and a readable history of what the agent actually did.

Because trust does not come from believing that an AI will never make a mistake.

It also comes from being able to understand what happened when something goes wrong.

Maybe autonomy starts small

Paying a bill, booking a flight or buying a token all require different levels of permission.

That is why the question “Would you give an AI access to your wallet?” probably cannot be answered with a simple yes or no.

Maybe real financial autonomy starts much smaller.

With €20.

With one task.

With clear rules.

And with a question at exactly the moment the AI would have to step outside those rules.

If that works, perhaps we give it a little more responsibility next time.

And eventually, a question that feels unusual today might become completely normal:

What is my AI allowed to pay for today?


What would you let an AI handle with your money — and which decision should it always ask you about first?

Sources

[1] Coinbase — Introducing Agentic Wallets: Give Your Agents the Power of Autonomy

[2] Coinbase Developer Documentation — Show Wallet App

Sources last checked on October 3, 2026. The everyday scenarios in this article are thought experiments and do not constitute financial or investment advice.

Subscribe to Tankothek