A wave of coverage in mid-September 2026 described a genuinely new pattern in agentic commerce: AI agents shopping not just on behalf of a human end customer, but for other AI agents, one system sourcing inputs, comparing options, and transacting with another system, with no human directly in that specific loop. If your funnel model assumes every conversion traces back to a person who saw an ad, clicked, and decided, agent-to-agent commerce is the first scenario that breaks that assumption at the structural level, not just at the margins.
Here's the distinction worth being precise about before going further, because "agentic commerce" gets used loosely. An agent helping a human shop, summarizing options, completing a checkout the human approved, is still fundamentally a human-driven purchase with an AI assistant in the loop. Agent-to-agent commerce, as newly reported, is a different thing: one agent making a purchasing decision that serves another agent's task, where the eventual human beneficiary may be several steps removed from the actual transaction moment.
- Mid-September 2026 coverage (widely syndicated via Stacker/PYMNTS) described AI agents transacting directly with other AI agents, not just assisting a human shopper.
- This is structurally different from agent-assisted human shopping: the purchasing decision itself is being made agent-to-agent, without a human decision moment at that specific step.
- Standard funnel models (impression, click, decision, purchase) assume a human decision-maker at the point of conversion; agent-to-agent transactions don't have one at that step.
- This connects directly to the ad-buying side of the same shift: protocols like AdCP and IAB Tech Lab's new AAMP 3.0 are the infrastructure for agents transacting on the ad-buying side specifically.
- No standard measurement methodology yet exists for attributing marketing influence when the transacting party at the point of purchase is an agent rather than a person.
What does "AI agents shopping for other AI agents" actually mean in practice?
The clearest way to think about it: a chain of agents, each handling a piece of a task, where a purchasing decision happens at some link in that chain that isn't the human end-user. A simple example of the shape, without claiming this exact scenario is documented: a personal AI assistant tasked with "restock my kitchen supplies" delegates a sub-task to a specialized shopping agent, which compares vendors and completes a purchase. The human set the goal, but wasn't present for, and didn't review, the specific purchasing decision.
That's meaningfully different from an agent that shows a human three options and waits for a click. In the agent-to-agent model, the "decision-maker" your marketing needs to reach and persuade may not be a person browsing anything. It may be another system evaluating options against criteria a human set once, upstream, and then stopped actively supervising.
Why does this break standard funnel measurement?
Every standard funnel model, impression, click, add-to-cart, purchase, assumes a human at each stage, making a legible decision that ad creative, targeting, and messaging can influence. Take the human decision-maker out of the final step and several assumptions stop holding.
Ad creative may not be "seen" the way your measurement assumes. If a shopping agent is comparing structured product data, price, and vendor terms rather than viewing a display ad or a video, your creative's actual influence on the transaction is a different question than "did the eventual human end-user see and respond to this ad," because they may not have seen it at all.
Click-based attribution has no clean analog. A click is a legible, individually attributable human action. An agent evaluating and selecting among options isn't clicking in the way attribution models are built to log and credit.
The funnel's timing compresses or disappears. Awareness, consideration, and decision as sequential human mental states don't map cleanly onto an agent executing a pre-defined evaluation criterion in milliseconds. The funnel as a metaphor for a human journey over time may simply not describe what's happening.
None of this means measurement becomes impossible. It means the existing toolkit, built around human attention and clicks, needs a genuinely new layer for this specific transaction type, not a patch on the old one.
Is this the same problem AdCP and AAMP 3.0 are trying to solve?
Related, but not identical. AdCP and AAMP 3.0 are protocols for the ad-buying side of agent-to-agent interaction: how an agent representing a media buyer transacts with an agent representing a seller of ad inventory. Agent-to-agent commerce, as covered here, is about the purchase side, an agent transacting with another agent (or a merchant's agent) to buy a product or service, downstream of whatever advertising influenced that decision, if any advertising was involved in a legible way at all.
The two are connected in a real sense: as more of both ad-buying and end-purchase transactions move to agent-to-agent interaction, the entire chain from ad impression to final sale increasingly involves machine-to-machine steps that today's measurement infrastructure, built around human clicks and human purchase decisions, wasn't designed to see through.
What can a marketing team actually do about this today?
Realistically, not much at the level of solving it. This is early enough that no standard methodology exists yet, and claiming otherwise would be overselling a toolkit that doesn't exist. What's genuinely actionable now is awareness and a specific kind of measurement hygiene.
Watch for unexplained conversion patterns that don't map to a legible human journey. If a growing share of conversions in a given channel show minimal or no engagement signal before purchase, no clicks, no time-on-site pattern consistent with human browsing, that's worth investigating as a possible agent-mediated transaction rather than assuming it's a measurement or tracking failure.
Treat first-party data and structured product information as a new kind of "creative" investment. If purchasing agents are evaluating structured data, price, specs, availability, terms, rather than persuasive creative, the quality and completeness of that structured data becomes a genuine competitive input, parallel to how ad creative quality matters for human-facing channels. This is a real, low-regret investment regardless of how large agent-to-agent commerce ultimately becomes.
Don't retrofit old attribution logic onto a fundamentally different transaction type. Forcing an agent-to-agent purchase into a last-click or multi-touch model built for human journeys will produce a number, but not a meaningful one. It's more honest to flag these as a distinct, currently-unattributed category than to force a false precision onto them.
Is this actually happening at scale yet, or is it early speculation?
Genuinely early. The September 2026 coverage describes an emerging pattern and real examples, not a mature, high-volume transaction category with established measurement practice. The honest read: this is a "watch closely, start building the right instincts now" moment, not a "your current funnel measurement is already obsolete" one. Most conversions today still trace back to a legible human decision.
What makes it worth a dedicated look now rather than waiting: the infrastructure enabling it, agentic checkout protocols, agent-to-agent ad-buying standards, AI shopping assistants with real purchasing authority, is being built out rapidly and in parallel across multiple platforms in 2026. The measurement problem tends to arrive faster than the industry expects once the enabling infrastructure is in place, which is the same pattern that played out with mobile and with programmatic before it.
Frequently asked questions
Related reading
- Cross-Platform Ad Reporting: Why Nothing Adds Up (And What To Trust): the cluster pillar this new measurement gap sits inside
- Incrementality Testing: The Only Measurement That Answers the Question You're Actually Asking: a measurement approach less dependent on tracking an individual human's click path, worth revisiting as agent-mediated transactions grow
