The conversation becomes the interface

Agents and enterprise systems · Working draft

Based on notes from June 2026

For most of the web’s history, a business reached a customer through a page the customer opened. More of that contact is starting to happen inside conversations: a person asks an assistant in a chat or message thread, and the assistant finds, compares and sometimes acts.

For that to spread, assistants need distribution where people already spend time, and businesses need to expose their catalogs, prices and actions in a form an agent can use.

Search splits by who is asking

Three search paths. Humans: the Google search box. Chatbots: web search inside ChatGPT or Claude. Agents: AI search built for machines, not people.
People at a search box, chatbots searching for them, and agents calling search built for machines. Open full size

Three behaviors often get lumped together here. Retrieval is an assistant answering a question with information from the web. A referral is a person clicking from that answer through to a business. An autonomous purchase is the agent completing the transaction itself.

Each asks something different of a business, and the evidence for each is very different.

Referrals are measured; purchases are early

The clearest number so far is about referrals. Adobe measured traffic to US retail sites from generative AI tools over the 2025 holidays.[1]

+693%

Traffic to US retail sites from generative AI tools in the 2025 holiday season (Nov. 1 to Dec. 31), compared with a year earlier. Source: Adobe Analytics, January 7, 2026.

That is people starting in an AI tool and then visiting a store. It says nothing about agents buying on their own.

The action side is earlier. The Agentic Commerce Protocol is an open standard for checkout between buyers, AI agents and businesses, in which the business stays in control of what it sells.[2] Standards like this show the plumbing being built. Whether people trust agents with their money is a separate question.

Three things AI commerce can mean. 1. AI referral: a person asks an AI tool, then visits the store; Adobe measured +693% traffic. 2. Information retrieval: an agent reads catalogs, prices and availability. 3. Agent-completed purchase: the agent checks out on the person's behalf; checkout standards such as ACP are early. Each step adds autonomy and needs more trust.
Three different things an “AI commerce” figure can mean. Only the first is measured here (Adobe, above). Open full size

Why messaging

Messaging needs no new app or login, and many tasks fit a thread: a quick request, a confirmation, a reminder, a simple choice. Detailed comparisons and anything visual fit less well.

I built a small event finder for iMessage that suggests events that fit your calendar. It’s a prototype that shows the shape of the idea, not evidence that people want it.

Businesses that depend on people seeing their interface, through ads, upsells or onboarding, would feel this first. Those that make their products easy for agents to understand and buy could gain, as could whoever runs identity and payments for agent-initiated purchases.

It’s also possible that people keep using assistants for research but still choose and pay on familiar sites, so referrals grow and autonomous purchases stay rare. I’d change my mind if agent-completed purchases stay negligible after checkout standards are widely available, or if messaging platforms close off third-party agents.

Notes:

  1. Adobe, “2025 holiday shopping season results,” January 7, 2026. The figure covers Nov. 1 to Dec. 31, compared with a year earlier. ↩︎

  2. Agentic Commerce Protocol. ↩︎