Can AI Decide on Our Behalf?

Letting an AI agent handle research can make life much easier. The real question is where that should stop: how much decision-making authority are we ready to hand over?

In the previous post, we talked about how our use of the internet might change once we start describing our needs instead of searching for products. It's quite appealing for an agent to review hundreds of products, compare prices, weigh reviews, and present us with a few meaningful options. It cuts down the time we spend on research and makes deciding easier.

Still, making research easier and deciding on our behalf are not the same thing.

Let's go back to the headphones example. I find it useful for an agent to review hundreds of models and narrow them down to three options based on my usage habits, budget, and expectations. But if the same system also picks one of those three and places the order, something else is going on. We're no longer talking about a system that merely lightens the research load; we're talking about an actor that makes a choice on my behalf and turns that choice into an actual transaction.

Where does an agent helping us end, and deciding for us begin?

Is Giving a Recommendation the Same as Making a Decision?

We're already used to digital systems making recommendations to us. E-commerce sites show products based on our past choices, streaming platforms surface content that might interest us, and banks offer deals based on our usage habits. The system may be influencing our decisions, but the final choice still belongs to us.

With AI agents, this boundary can shift. An agent doesn't just make recommendations; by weighing our needs, our budget, our past preferences, and the rules we've set all together, it can act on our behalf.

Consider an instruction like this: “Every month, buy the household's basic needs within the budget I've set, keep using the products I already use whenever possible, but also consider alternatives if there's a meaningful price difference.” In this case, we don't choose every product one by one. We set the expectations and the boundaries, and leave the choice to the system.

In a model like this, who actually made the decision becomes increasingly unclear. Is it the user who set the rules, the agent interpreting those rules, or the platform that built the agent's evaluation system?

Can We Delegate Every Decision Equally Easily?

An agent reordering the coffee I use every month is not the same level of decision as choosing a credit card. Narrowing down hotel options within my budget carries a different weight than completing a non-refundable reservation outright.

As the amount involved grows, as reversing the decision gets harder, or as the financial consequences get heavier, we may want to keep control in our own hands. So the real question may not be whether we grant AI decision-making authority at all.

The more accurate question is:

In which decisions, within which boundaries, and how much authority are we ready to grant?

We already use similar control mechanisms today. We put spending limits on cards, define different approval levels in corporate structures, and require extra verification for certain transactions. It seems natural for a similar structure to emerge for AI agents as well.

Low-value, frequently repeated transactions could happen automatically. For larger amounts, the agent could make a choice and then ask the user for approval. For decisions with major financial impact or that are hard to reverse, the agent's role could stay limited to research and recommendation.

That way, instead of drawing a single line between control and convenience, we could build tiers of authority that scale with risk.

Who Is the Agent Really Deciding For?

An agent acting on my behalf doesn't necessarily mean it's only looking out for my interests.

Will the agent really choose the product that suits me best, or will it surface the product that has a commercial relationship with the platform? Will commission rates influence the choice? Will the platform's own products be rated more favorably than competitors'?

Today we can spot a sponsored product or an ad to some extent. But if an agent evaluates the options and eliminates some of them before we ever see them, understanding how the decision was shaped could be far harder.

For this reason, transparency in Agentic Commerce shouldn't be limited to the question “What did the agent recommend?” Users should also be able to see why a particular option was recommended, which alternatives were considered, and whether any commercial influence played a role in the decision.

Because if we want to delegate decision-making authority, we also want to know how the decision was made.

Who Will Be Responsible for a Wrong Decision?

If an AI system recommends the wrong product, I can simply pick another one. But if the system carried out the purchase on my behalf, the same mistake is no longer just bad advice. The product might be wrong, the budget might be blown, or a condition I would never have accepted might have been approved in my name.

If the user granted the agent a general authority, is it right to treat every transaction under that authority as a conscious approval? If the agent stayed within the defined boundaries but misread the user's intent, who bears the responsibility? When a user disputes a transaction, will it be possible to later examine why the agent made that decision?

From a payment-systems perspective, I find this distinction especially important. In financial transactions, it isn't enough for the transaction to simply happen; who initiated it, what authority it relied on, and which controls it passed through also matter.

AI agents will need a similar kind of traceability.

Just as important as the agent making the right decision will be its ability to leave a trace of the decision it made.

Will the Human Decision-Maker Turn Into an Authority-Granter?

I think this might be one of the most significant changes Agentic Commerce could bring to user behavior.

Today we research, compare, and choose. In the future, instead of dealing with the details of every single decision, we might define what we want, within which boundaries the system can act, and in which situations we want to give approval.

We're already familiar with this model in business life. A manager doesn't make every decision personally; they set the goal, the budget, and the limits of authority, and expect execution to proceed within that framework. The difference with Agentic Commerce is this: the party we grant authority to isn't a person or a team, it's software.

In that case, one of the most important screens in the future user experience may not be the product catalog at all. It could be the screen where we manage which authorities we've granted to our agent.

I believe what will define the future of Agentic Commerce is not unlimited autonomy, but controllable and auditable autonomy. If the user has to give approval at every step, most of the convenience the agent provides disappears. If they hand over all control, trust and accountability problems emerge instead.

What needs to be designed is the balance between these two extremes.

What Changes When We Move From Deciding to Paying?

In the first part, we talked about how the next customer on the internet might not be human; in the second, about how the search economy could be affected once people start describing their needs instead of searching for products. Now we've reached the boundary between delegating research and delegating decision-making authority.

But the process doesn't end once an agent chooses a product on our behalf. The chosen product still has to be bought, and that step takes us directly into payment systems.

It's one thing for an agent to decide which headphones we should buy; it's a very different level of authority for it to be able to initiate a payment from our card or bank account.

Which payment instrument will it use? Who will set the spending limit? Should the user be notified of every transaction? Who will be responsible if a wrong or unauthorized transaction happens?

Delegating research can bring convenience. Delegating decision-making authority requires trust.

The authority to touch money is a different boundary altogether from all of this.

I'll take up this question in the next part of the series:

Are we ready to give an AI agent the authority to make payments?