Agentic Commerce Won't Scale Until Trust is Established

Agents that browse catalogs, compare prices and complete purchases without a person clicking a button were a novelty two years ago. Today, retailers are scrambling to keep up. The technology arrived faster than most road maps anticipated — and with a trust problem the industry hasn't solved.

For the first time in Cloudflare's history, machines accounted for more than half the traffic on its network, according to the company's recent earnings call, a milestone it hadn't expected to hit until 2027.

This influx of agent traffic comes with changing shopping behaviors. Comparing prices across merchants, applying loyalty rules, and checking out without humans reviewing the final decision is just the start. Investment in agent-driven transactions is accelerating because faster transactions and a more efficient path from intent to purchase offer clear benefits to consumers, merchants, and payment providers alike.

Payment providers and retailers are aggressively betting on the payoff. Anthropic recently released a blueprint for building shopping and merchant agents on Claude, giving developers reference code to launch an agent that browses catalogs and completes checkout in days. Meanwhile, the FIDO Alliance, Visa, Mastercard, and EMVCo are still working out the standards that would let merchants trust those agents, including how to confirm one accurately represents what a shopper wants and what it's authorized to spend. So, while the payments industry is working to embed trust into this new era of commerce, the commerce layer is still arriving quicker than trust.

Human Purchasing vs. Agents

When a human completes a purchase, the retailer can generally assume the customer authorized it. An agent might accurately represent a consumer and still exceed the budget it was given, or purchase counterfeit product from a fraudulent website while trying to fulfill simple instructions. From the outside, a properly authorized agent and a compromised or malfunctioning one can look identical until something goes wrong.

That ambiguity around authorization and responsibility is one of the central problems agentic commerce needs to solve as adoption scales. Most tools available to merchants today answer whether traffic is automated or human. That’s useful for filtering bots and scrapers, but it doesn't answer the questions that matter once an agent is authorized to spend money: Who does this agent represent? What authority has it been given? And did this transaction stay within those boundaries? Without answers, merchants can't reliably tell an authorized purchase from an agent that has gone off script, and neither can the platforms or financial institutions between them.

The stakes also rise with scale: an agent can repeat the same mistake across thousands of transactions before anyone notices, turning what might once have been an isolated error into a systemic risk.

Online retail has faced versions of this problem before. Identity verification providers, for example, are measured against established error rates such as APCER and BPCER, which track how often a system accepts an impostor or rejects a legitimate customer. Agents can be held to the same standard, with clear success and failure rates for the tasks they perform. That gives platforms and payment providers a practical basis for deciding who is responsible when something goes wrong.

Agentic commerce adds a wrinkle since agents can be asked to do almost anything. Therefore, task-specific metrics only go so far. The industry needs to build toward a broader notion of trust, in both the agents themselves and the entities they represent. Getting there also requires strong monitoring of how agents behave, with a human in the loop overseeing the system so responsibility can be defined in concrete terms.

Closing the Trust Gap

Merchants need a way to confirm which consumer an agent is acting for before honoring a transaction, rather than assuming access to a payment method is enough. They need clear terms on record for what that agent is authorized to do so a purchase can be checked against those permissions, along with a way to trace disputed transactions back to the authorization meant to cover them. Without that visibility, responsibility defaults to guesswork every time an agent oversteps.

The industry has already built substantial infrastructure for verifying human identity in commerce. Extending those same principles to the agents acting on humans' behalf is the logical next step, and arguably the more urgent one given how quickly agent activity on the web is scaling.

Retailers, merchants and platforms don't need every open question about agentic commerce answered at once, but the sequencing of controls and governance around agents needs to be right.

Given the diversity of tasks agents can already perform on today's stack, trying to manage agentic activity through strict, task-specific, rule-based controls won't scale in any practical or cost-effective way. That challenge will only grow as agent capabilities advance in the near future. Retailers and vendors should invest in strong, data-driven observability within their systems and rely on a continuous, data-driven notion of trust, rather than evaluating each interaction against a rigid, point-in-time, rule-based check.

Trust will come from being able to confidently answer who an agent represents and what it's allowed to do. Establish that foundation first and agentic commerce has a path to scale securely. Skip it and the technology risks moving faster than the trust required to sustain it.

Keivan Bahmani, Ph.D., is director of machine learning and AI integrity at Mitek Systems, a global leader in digital identity verification.