When the Buyer Is an Agent

Every argument in this series has assumed a human reads the AI answer. A person asks ChatGPT for a vendor shortlist, sees three names, and decides. That assumption has maybe two years left.

The short answer: agentic buying is real but early, with autonomous agent traffic still under 1% of sessions. Anyone telling you to rebuild your marketing for machine buyers today is ahead of the evidence. What is not early is the infrastructure. Amazon's Rufus already purchases autonomously, Mastercard has shipped cryptographically signed mandates so agents can transact on a user's behalf, and the protocol layer is being standardized right now. When that settles, the thing evaluating your company will be software that cannot be charmed, and the deciding factors shift from persuasion to verifiable identity and structured data.

That shift lands squarely in territory I have spent my career in, so let me be careful to separate what is happening from what is forecast. (For the wider picture across consumer and business buying, including how Visa and Mastercard built the payment rails, see my agentic commerce overview.)

What is actually shipping

Amazon is furthest along, and its numbers are disclosed rather than projected. (Amazon renamed Rufus to Alexa for Shopping in May 2026. The figures below were reported under the Rufus name.)

More than 300 million customers used Rufus during 2025. By the third quarter, monthly users were up 140% year over year and interactions up 210%. Amazon reported it delivered close to $12 billion in incremental annualized sales, above the $10 billion pace flagged a quarter earlier. Shoppers who use Rufus are roughly 60% more likely to complete a purchase. On the earnings call, Andy Jassy described Rufus as "our agentic AI shopping assistant," which is deliberate language.

The genuinely new capability arrived in November 2025. Rufus auto-buy lets a Prime member set a target price on an item. Rufus then checks the price every 30 minutes and completes the purchase autonomously when it hits, with a 24-hour cancellation window. If the customer does nothing, the transaction closes with no human in the loop.

Around it, the rails are being built. Mastercard's Agent Pay gives AI agents a verifiable way to transact using cryptographically signed mandates. Google shipped agentic shopping and checkout. OpenAI, Google, and Microsoft are each moving on agentic commerce protocols, and Perplexity and Shopify are building agent-facing storefront layers.

The honest scope, because the forecasts are doing heavy lifting

Now the part the agentic commerce evangelists skip.

Autonomous agent traffic is still under 1% of sessions by most estimates. The much larger category today is the AI-assisted human, where a person is still deciding but the interface is conversational, and that share is growing. Those are different problems and conflating them inflates the story.

The B2B numbers everyone quotes are forecasts, not measurements. Gartner projects that 90% of B2B buying will be AI-agent intermediated by 2028, representing more than $15 trillion. It also projects that 33% of enterprise applications will feature agentic AI by the same year, up from under 1% in 2024. Forrester expects 20% of B2B sellers to engage in agent-led quote negotiations during 2026. I have watched enough analyst forecasts miss to treat a 90%-by-2028 figure as a direction rather than a number, and I would say the same about any projection in this series.

What makes me take the direction seriously is not the forecast. It is that the payment rails and identity mandates are being built now, by card networks and platforms that do not ship infrastructure speculatively.

What changes when your reader is software

Here is the part that matters strategically, and it is a genuine break from everything else in this series.

Persuasion stops working. A human reading a comparison page responds to framing, social proof, and narrative. An agent evaluating options against a specification does not. It parses attributes, checks constraints, and ranks. Your positioning paragraph is not an input.

Data accuracy becomes a ranking factor. In agentic commerce, inventory accuracy, pricing currency, and specification completeness determine whether you are eligible to be selected at all. Microsoft's own AEO guidance already frames AI product recommendations around data completeness, currency, structured markup, and contextual richness rather than keyword tactics. A stale spec sheet is no longer an embarrassment, it is a disqualification.

The evaluation becomes machine-readable or it does not happen. If your pricing requires a sales conversation, your technical specifications live in a gated PDF, and your integration list is rendered in JavaScript, an agent cannot evaluate you. In B2B software, that describes a large share of vendor websites, including most security vendors.

Why this is an identity problem

When a buyer is an agent, three questions arrive that marketing has never had to answer, and all three are identity questions.

Who is this agent acting for? A purchase request arriving from an autonomous agent needs to be traceable to a human or organization with the authority to make it. This is delegation, and it is why Mastercard's approach uses cryptographically signed mandates rather than a shared credential.

What is it authorized to do? Spending limits, merchant whitelists, and transaction rules have to be enforceable at the rail level rather than trusted to the agent. The card networks are embedding hard caps and merchant restrictions precisely because an agent with unbounded authority is an unacceptable risk.

How do you attribute the transaction afterward? If an agent researched, selected, and purchased, your analytics see a machine session. Attribution models built on human behavioral signals degrade.

Anyone who has built authorization systems will recognize all three as the machine identity problem, arriving at the commercial edge instead of inside the data center. We spent years learning that service accounts with static credentials and unbounded scope were a liability. Agentic commerce is that lesson replayed with a buyer's wallet attached, which is why the signed-mandate approach matters more than any individual shopping feature.

What I would actually do in 2026

Not rebuild your marketing. The evidence does not support that, and anyone selling agent-first positioning today is selling a 2028 problem at 2026 prices.

I would do three unglamorous things.

Make your facts machine-readable and current. Structured data for products, pricing, integrations, and specifications. Not because a schema trick wins AI citations, which it does not, but because factual completeness is the eligibility criterion in agent-mediated selection. This work also pays off immediately in conventional AI answers, so it is not a speculative bet.

Get your entity represented consistently. Agents disambiguate entities before evaluating them. If your company is described inconsistently across your own properties and third-party sources, you are harder to resolve and easier to skip.

Watch the shortlist layer, which is already agent-mediated even though the purchase is not. This is the near-term version of the whole argument. Long before an agent buys your product, an AI assistant is compiling the list of vendors a human will consider. My data at GrackerAI (disclosure: my company, a GEO platform for B2B SaaS) shows 4 in 10 B2B security buyers already start vendor research inside AI assistants. Shortlist formation has moved. Purchase execution has not yet.

That sequencing is the practical takeaway. The buying is still human. The consideration set is increasingly not.

Here is my question, particularly for anyone selling complex B2B software. Could an agent evaluate your product today from public, structured information alone, without a demo, a sales call, or a gated PDF? If not, you have a fair amount of time to fix it, and roughly zero excuse for being surprised when the deadline arrives.

Frequently Asked Questions

Is agentic commerce actually happening yet?

Partly. Amazon's Rufus was used by more than 300 million customers in 2025 and delivered close to $12 billion in incremental annualized sales. Since November 2025, its auto-buy feature purchases items for Prime members when a set price is reached, checking every 30 minutes with a 24-hour cancellation window. However, fully autonomous agent traffic remains under 1% of sessions, while AI-assisted human shopping is far more common.

What do analysts forecast for B2B agentic buying?

Gartner projects that 90% of B2B buying will be AI-agent intermediated by 2028, representing more than $15 trillion. It also projects that 33% of enterprise applications will feature agentic AI by 2028, up from under 1% in 2024. Forrester expects 20% of B2B sellers to engage in agent-led quote negotiations during 2026. These are forecasts rather than measurements and should be treated as directional.

How does marketing change when an AI agent is the buyer?

Persuasion stops being the mechanism. An agent parses attributes, checks constraints, and ranks options against a specification rather than responding to framing or social proof. Data accuracy becomes a selection criterion, since inventory currency, pricing, and specification completeness determine eligibility. Information locked in gated PDFs, sales calls, or JavaScript-rendered pages effectively removes a vendor from agent evaluation.

Why is agentic commerce an identity problem?

Three questions arrive at once: who the agent is acting for, what it is authorized to do, and how the transaction is attributed afterward. These are delegation, authorization, and accountability questions. Mastercard's Agent Pay addresses them with cryptographically signed mandates rather than shared credentials. Card networks are embedding spending caps and merchant restrictions at the rail level, mirroring how enterprises learned to govern machine identities internally.