A few days ago, I found myself thinking about a pretty simple question.
If AI agents are starting to get their own wallets and can spend money, what exactly stops them from making financial decisions for us?
Apparently, the answer is coming faster than I expected.
This week, Public announced a partnership with Kalshi that brings AI agents into prediction markets.
And we're no longer talking about an AI that simply analyzes some data and tells us:
"If I were you, I might keep an eye on this market."
We're starting to talk about AI agents that can monitor markets and, based on rules defined by their users, execute trades.
And that immediately led me to another question.
What happens when the AI gets it wrong?
From Assistant to Actor
Until now, the line was relatively simple.
I can ask an AI to analyze Bitcoin.
It can look at data, compare different scenarios, summarize information and maybe point me toward something interesting.
But at the end of the process, I'm still the one clicking Buy or Sell.
My decision.
My money.
My responsibility.
With autonomous agents, that line starts to become much less obvious.
Imagine I give an agent this instruction:
"Watch BTC. If conditions A, B and C are met, open a $100 position."
Then I go to sleep.
During the night, those conditions appear to be met.
The agent buys.
Bitcoin drops 15%.
Who made the bad decision?
Me, because I wrote the rules?
The agent, because it interpreted the data?
The developer who built the model?
Or the platform that allowed the agent to execute the trade?
And this is where the subject becomes much more interesting than simply asking:
Can AI trade?
Maybe AI Isn't Actually the Problem
There's an important distinction here.
An agent doesn't necessarily have to "decide" the way a human does.
It can simply turn instructions into executable rules.
That's essentially the approach Public is taking: users define and approve the rules before the agent is allowed to act.
In that situation, responsibility seems relatively straightforward.
I created the strategy.
I authorized its execution.
The agent is basically a much more sophisticated form of automation.
But now imagine something slightly different.
Instead of saying:
"Buy BTC if the price crosses X."
I say:
"Analyze the market and find the best opportunities for my portfolio."
Now I'm no longer defining the decision.
I'm defining the objective.
And the AI decides how to achieve it.
That's a very different situation.
What If the Agent Misunderstands What I Want?
This might be one of the most interesting problems in this emerging agent economy.
We're already used to giving AI vague instructions.
And that's usually fine.
If ChatGPT misunderstands my question, nothing dramatic happens.
I just rephrase it.
But when an agent controls money, misunderstanding an instruction can have an immediate financial cost.
Today's payment systems were largely designed around humans explicitly authorizing transactions. Once software begins acting through delegated authority, things get more complicated.
What did the user actually intend?
What was the agent authorized to do?
And did the agent stay within those boundaries?
That completely changes the consequences of a simple AI mistake.
A hallucination inside a conversation can produce a bad answer.
A hallucination with access to a wallet can produce a real transaction.
This Goes Far Beyond Trading
What's interesting is that the same debate is already appearing elsewhere.
AI agents are beginning to select products, interact with merchants and make purchases on behalf of users.
Visa, for example, has been working on agentic commerce systems designed to let AI agents transact within parameters defined by cardholders.
And financial institutions are already asking uncomfortable questions about this new model.
Fraud.
Data protection.
Unauthorized transactions.
And perhaps the most important question for consumers:
Who do you call when something goes wrong?
Sooner or later, we'll have to answer the same question for financial agents.
You Can't Send the AI a Bill
There's also one fairly obvious answer we can eliminate.
We're not going to send the AI the bill.
A model doesn't have its own personal bank account.
It doesn't legally own your portfolio.
And it can't simply become responsible for your losses in the same way a person or a company can.
So responsibility inevitably comes back to the humans and organizations surrounding the agent.
The user.
The developer.
The platform.
The financial provider.
And the more autonomy we give an agent, the harder those boundaries become to define.
Who gave it authority?
What limits were placed on that authority?
Did it stay within those limits?
And what happens if it didn't?
Maybe the Real Solution Is Limits
The more I think about it, the less I believe the future will look like this:
"Here's my wallet. Go make me money."
That would obviously be a terrible idea.
Instead, it will probably look more like:
"You can use a maximum of 2% of this portfolio."
"You can only trade these assets."
"You can only open one position at a time."
"Anything above $100 requires my approval."
"If daily losses reach X, stop everything."
In other words, we may eventually have to manage AI agents the same way we already manage permissions in computing.
A program doesn't need access to everything.
Neither does a financial agent.
And perhaps the safest AI agent won't necessarily be the smartest one.
It might simply be the one with the best-defined boundaries.
But There Could Be a Bigger Problem
Now imagine we're not talking about one agent.
Imagine millions of them.
All analyzing similar information.
Some using the same models.
Some following similar strategies.
And potentially reacting to the same signals at roughly the same time.
This is where things become really interesting.
If enough autonomous systems respond to the same market event in similar ways, they could potentially amplify price movements rather than simply react to them.
Suddenly, the question isn't just whether my agent can lose my money.
It's what happens when a meaningful part of the market is made up of autonomous software capable of analyzing, deciding and executing within seconds.
That's an entirely different scale of problem.
We May Be Creating a New Kind of Economic Actor
This is what fascinates me most about AI agents.
At first, they were assistants.
Then we gave them tools.
Then APIs.
Then wallets.
Now we're starting to let them monitor markets and execute financial strategies.
At every step, we're giving software a little more ability to act.
And that creates a strange new boundary.
At what point does software stop simply advising us and start acting economically on our behalf?
I don't think the answer is to keep AI agents away from money.
Quite the opposite.
I think this could become one of the most interesting applications of the technology.
But if we're going to give machines financial autonomy, we'll need to build something alongside it:
limits, permissions and, above all, a clear chain of responsibility.
Because the moment an AI can click Buy without asking us first, a bad answer is no longer just a bad answer.
It becomes a financial decision.
If you enjoyed this article and want to support my work, you can help by leaving a like, sending a tip, or using one of my affiliate links. It costs you nothing extra and helps me keep researching, experimenting, and publishing new articles.
Thanks for reading — see you in the next one!
🚀 Interested in trying OKX? You can support my work by using my referral link to sign up and complete the eligible tasks to unlock available rewards.
