Growing pains: The long road to agentic commerce

The emergence of agentic commerce has sparked big dreams about millions of agents fanned out across the digital economy on missions to perform every imaginable task.

Those dreams will come true but not before we clear some major hurdles.

How will we certify agents at scale? How will merchants and other businesses react to the new technology? How will people identify agents they know they can trust?

We don’t have clear answers yet. But one thing is certain: It’s going to be messy in the beginning.

How will we certify agents at scale?

If we get to the point of needing to certify millions of new agents every day, there’s no way humans can do it.

We’ll turn the job over to AI and find ourselves needing to trust two sides of the technology. That’s a big leap and it will be hard to develop confidence that the AI checker is doing its job.

The problems arise from the things we can’t anticipate. Without a human involved in certification, we have to program the checker to do everything we do when deciding if an agent is suspicious. But there are always surprises humans react to in the moment.

That leaves us with questions about how permissive or restrictive we are when we program the checkers. Do they do only what we say they can do? Or do we program them so they can’t do certain things?

If we program them not to do anything outside of boundaries we set, we’re limiting the agents’ power. So we’ll have to strike a balance between how much freedom we give them and how restricted we make the space where they operate.

In the beginning, we’ll have supervised checkers so we can sample the activity to make sure they’re doing the right thing. But eventually, we’ll have to trust them and set up automated guardrails to make sure things don’t get out of hand.

How will people and businesses react?

Big retailers will almost certainly take the lead with early adoption. But we’ll live for a long time in an unbalanced world where some merchants accept agents and some don’t.

Merchants have spent years developing defenses against automated bots. Now we’re asking them to trust a new kind of automation in the form of AI agents that are orders of magnitude more powerful than bots.

Merchants are going to have questions about whether they can trust the technology and if they need investment and infrastructure on their end. Some might rely on proxy browser shoppers, which are intermediaries on the agentic side that allow the agent to communicate with the merchant company.

The agent would have an interface that essentially lets it push buttons on the merchant’s website to make a purchase. It would be clunky, but it would work and some organizations I’ve talked to have said that’s their fallback plan to include merchants in agentic commerce.

Consumers will also have questions about whether they’re getting the best deal using an AI agent. If one major airline, for example, immediately adopts the technology while another chooses to wait years, consumers using an AI agent to book travel won’t know for sure if they could have found a lower price.

Agentic commerce is coming, but it will be an unevenly supported world for a while.

How will people find trusted, certified agents?

Once agentic commerce hits its stride, people will need to find the right agents for certain tasks.

I predict we’ll see agent directories that look like app stores. There could be different stores based on certain technologies, large language models and platforms, such as Google, OpenAI and Apple.

There also could be a unifying force that serves as a directory of directories. That federated directory would provide a hierarchical structure people could use to search for agents by name, function, certificate or globally unique identifier.

It would operate like a distributed domain name system. The best approach would be an open environment without a single authority or any walled-off areas, which is how the internet has succeeded.

The growing pains of new technology

Agentic commerce is going to work, but it will start slowly and in restrictive ways that are human-approved at the decision gates.

Right now, people are having conversations with agents, asking them to gather information and then narrowing it down to the point of purchase. In the early days of agentic, the agent will do all the work, but the person will make the decision to let the agent make the purchase.

The use cases likely will be simple, such as assigning an agent in the morning to order lunch in the afternoon with a spending limit. The stakes are low in those situations, which is why they’ll be prevalent in the early stages.

Eventually, agents will gain more autonomy to make purchases and the use cases will grow in complexity. But an underlying theme for years to come will be figuring out how to deal with the messiness of a technology that holds the potential to reshape the digital landscape.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?