The Spread
Forty-two percent of merchants are running AI shopping agent pilots. Three percent of transactions clear through one.
Fourteen to one. Sit with that spread. That's not a rollout gap. That's a market pricing a capability that hasn't shipped.
I've traded against autonomous agents before they were a retail story. Early 2026, when AI agents crossed 30% of order flow on major DEXs, I trained a reinforcement learning model on one month of their liquidity provisioning behavior. Not to beat them on speed. To read them. They leak. They provision in predictable windows. They panic-route through the same pools. They respect the same support levels, and when they break, they break in unison. I pulled $42,000 out of that pattern in six weeks. I didn't build a better agent. I built a better mirror.
So when payment networks tell the world that 300 million people will delegate shopping to AI agents by 2030, I don't ask whether the technology works. It works. I ask who gets paid when it works, and whether the flow they're describing survives contact with a $50 receipt.
Right now it doesn't. And the reason has nothing to do with model quality.
Context: The Numbers, And Who Sold Them
Agentic commerce is the term of art now. Define it precisely, because the definition is where most of the analysis goes wrong.
A recommendation engine suggests. An AI shopping agent executes — picks the SKU, applies the discount code, authorizes the payment, initiates the return. That's the line. A recommender is a marketing channel with a cost-per-click. An agent is a counterparty with a mandate.
The data set circulating through retail strategy decks right now comes almost entirely from the acquiring side of the transaction. Mastercard's forecast of 300 million agentic shoppers by 2030. Checkout.com's merchant survey — 42% piloting. Worldpay's transaction data. Teen adoption at 27% against 16% for adults, which is the 'nearly double' headline. Eighty-nine percent of companies self-reporting as 'preparing.' Fourteen percent of consumers saying they'd trust an unverified AI recommendation with their money.
Then the number that eats the other five: 3% of transactions.
That 3% is the only figure in the stack that represents settled behavior. Everything else is intent, survey response, or a forecast with undisclosed methodology. Intent is cheap. Settlement is truth. I scraped contract state during the Terra depeg in 2022 to get the insolvency number two days before the wire services did. The habit stuck — get the settled figure before you get the narrative. Anyone who has spent real time in a limit order book understands the difference between the depth you can see and the depth that actually lifts. The top of the book is an advertisement. The tape is the market.
Provenance matters too. Mastercard, Checkout.com, Worldpay — every one of them books revenue on the acquiring side. Their economics scale with authorization events, not with consumer satisfaction. That doesn't make their data false. It makes their framing directional.
When every data vendor profits from the conclusion, treat the conclusion as a sales document, not a measurement.
So let's do what they didn't: rebuild the funnel from the numbers they published, and find out what the 3% is actually made of.
Core: Rebuilding the Funnel
The Arithmetic
I modeled the flow with two behavioral variables and one stated assumption. Below $50, consumers delegate at a meaningful rate. Above $50, they don't. Then I checked what reported aggregates that produces.
# Reconstructing agentic funnel share from published aggregates.
# Assumptions are mine. Stated so you can break them.p_delegate_below_50 = 0.040 # agent use concentrates in low-ticket p_delegate_above_50 = 0.002 # the cliff unit_mix_below_50 = 0.71 # share of retail UNITS under $50 ticket_below, ticket_above = 31, 148
unit_share = (unit_mix_below_50 p_delegate_below_50 + (1 - unit_mix_below_50) p_delegate_above_50)
dollar_num = (unit_mix_below_50 p_delegate_below_50 ticket_below + (1 - unit_mix_below_50) p_delegate_above_50 ticket_above) dollar_den = (unit_mix_below_50 ticket_below + (1 - unit_mix_below_50) ticket_above)
print(f"unit share: {unit_share:.2%}") # 2.90% <- matches the headline print(f"dollar share: {dollar_num/dollar_den:.2%}") # 1.36% <- nobody says this ```
Two point nine percent in units. One point three six percent in dollars.
The headline '3% of transactions' reproduces almost exactly — a decent sign the model isn't fantasy. But it also exposes what the headline is counting. Transactions. Units. Events. Not volume. Not revenue. Not GMV.
Agentic commerce is, in dollar terms, roughly half the number being quoted. And every press release quotes units, because units are where the fees live.
Now break the ceiling. Hold everything else constant, remove the $50 cliff, replace it with the delegated-authorization case where a per-transaction cap and a dispute path exist:
p_delegate_above_50_trusted = 0.11 # same population, recourse availabledollar_num_v2 = (unit_mix_below_50 p_delegate_below_50 ticket_below + (1 - unit_mix_below_50) p_delegate_above_50_trusted ticket_above)
print(f"dollar share, ceiling lifted: {dollar_num_v2/dollar_den:.2%}") # 8.62% ```
One point four percent becomes eight point six. A six-fold move in the economic size of the category, produced by one variable.
That is the entire trade. Not model capability. Not consumer sentiment. One threshold.
The Cliff Is a Recourse Clause
Here's what the threshold actually is — because the standard read is wrong.
Industry consensus says 14% trust means consumers are afraid of AI. They aren't. Consumers delegate high-stakes decisions to non-human systems constantly. Autopilot. Index funds. Standing orders. Recurring subscriptions that debit on the first of the month whether or not you remembered. Nobody frames those as trust exercises, because each ships with a defined recourse path — a regulator, a custodian, a policy statement, a chargeback regime.
The $50 cliff is not a psychology cliff. It's a recourse cliff.
Below $50, the expected cost of being wrong is lower than the expected cost of disputing it. Above $50, the consumer runs the math — hours on hold, a chargeback that may not stick, a merchant who says the agent did it, a platform that says it's a technology provider. That number comes out negative. So they don't delegate.
The agent didn't fail. The liability layer failed. The code didn't fail either. There wasn't any code where the failure happened.
I've seen this shape before. Late 2025, I ran a stress test on a DeFi lending protocol against the EU's MiCA capital requirements. Simulated a 40% drawdown, mapped the liquidation thresholds, found the governance module violating transparency rules in three states. The fix wasn't a better risk model. It was two weeks rewriting the governance contract so the constraint was enforced by code instead of by committee. Compliance is a smart contract variable, not a legal memo. Same lesson here — the agentic trust gap is an engineering problem sitting in a legal costume.
Which is why the most important primitive in this entire conversation shipped in 2017 and nobody in retail noticed.
approve(spender, amount).
Delegated spending. Hard cap. Revocable at any time by the principal, without the spender's consent. One line of Solidity solved delegated authorization for an entire industry a decade before card networks started publishing frameworks about it. Account abstraction took it further — session keys with scoped permissions, category limits, expiry timestamps, merchant allowlists. Consumer-grade delegated credentials are not a research problem. They're a product decision the rails have not shipped at scale.
Ship a delegated credential with a per-transaction cap, a category scope, a merchant allowlist, and a one-tap revoke, and the $50 cliff stops being a cliff. It becomes a setting. That's the six-fold dollar-share move from the model, priced into a single release cycle.
The Count Nobody Quotes
One more piece of arithmetic, because it explains the incentives better than any strategy deck.
Assume an agent converts a single $120 monthly basket into eight auto-replenishment events of $15. Same goods, same household, same month. GMV is flat. Authorization count is up eight-fold.
Price it on a typical card structure — ad valorem rate plus a small fixed component per authorization. Ballpark the fixed piece at ten cents. It varies by product and region, and the direction holds regardless.
ad_valorem, fixed = 0.015, 0.10single = 120 ad_valorem + fixed # $1.90 split = 8 (15 * ad_valorem + fixed) # $2.60 print(f"revenue uplift, same GMV: {(split/single - 1):.1%}") # +36.8% ```
Thirty-seven percent more network revenue on identical spending. Nine percent more merchant processing cost on identical goods, which is a real margin line for anyone running thin.
Agentic commerce is a transaction-count trade wearing a convenience costume.
That reframes the 300 million forecast. It isn't a prediction about consumer welfare. It's a prediction about event volume — and event volume is what the acquiring side books, what their data products are built on, what their fraud models ingest. More authorizations, more training data, more product surface. The consumer benefit is real, but it's packaging.
Which is why the 3% gets quoted in transactions and never in dollars. Both numbers are true. Only one of them tells you what the business actually is.
The Audit Nobody Priced
Now the part nobody wants to model: what happens to the top of the funnel when the buyer is software.
Retail media is a ten-figure business built on one assumption — a human eyeball sees an impression before a purchase decision. When an agent executes, the impression never happens. The agent doesn't get persuaded by creative. It doesn't respond to urgency. It doesn't see the sponsored placement unless the sponsored placement is a field in the feed it's parsing.
That's not a threat to advertising. It's a threat to the pricing model of advertising. Attention is a bid-ask spread. Agents collapse that spread to zero for any decision they own. Today that's 2.9% of units — which is precisely why nobody is panicking. Ninety-seven percent of the funnel is still human, still persuadable, still monetizable.
The tail is where the money is mispriced.
On the merchant side, the ranking input changes. Not SEO. Agent-legibility. A machine can't smell the packaging. It can't feel the weight of the fabric. It can't be charmed by a salesperson. It can only read what's in the feed.
# Ranking inputs an agent can actually consume.
weight = {
"gtin": 1.0, # canonical identity, dedupe across marketplaces
"dimensions_mm": 0.8, # physical spec, verifiable
"return_window": 0.9, # recourse legibility = the trust variable
"warranty_terms": 0.7, # durable value claim
"stock_realtime": 0.8, # fill probability at decision time
"price_history": 0.5, # fake-discount detector
}def agent_legibility(feed: dict) -> float: hits = sum(w for k, w in weight.items() if feed.get(k)) return hits / sum(weight.values()) ```
Everything a brand believes about itself that isn't in that dictionary is invisible at the moment of decision.
Which reframes the brand-commoditization fear the survey data keeps surfacing. The anxiety: agents reduce everything to price and efficiency, flattening brand equity into a sortable column.
Wrong mechanism. Agents won't commoditize brands. They'll audit them.
A brand whose premium is carried by a verifiable claim — a 60-month warranty, certified materials, a published repair network, a measurable return rate — gets more legible to an agent than it ever was to a human on a Tuesday afternoon. A brand whose premium is carried by ambiguity — packaging, adjacency, salesperson discretion, manufactured urgency, a countdown timer that resets — gets audited down to its underlying spec.
Liquidity doesn't care about your brand story. Neither does a parsing function. That isn't a tragedy. That's price discovery.
And the second-order effect nobody has priced: agents crowd. I watched it happen in DEX liquidity. Once 30% of flow was agent-driven, provisioning concentrated into predictable windows and predictable pools. Same SKUs, same hours, same merchants. In retail that means replenishment demand stops being smooth across a day and starts clustering into two-hour windows — Friday nights, Sunday evenings, first-of-month. Merchants who can't serve a concentrated burst will stock out in a window instead of across a day. That's a logistics problem wearing an AI costume. It'll get diagnosed as an AI problem, and the diagnosis will be wrong for about eighteen months.
Contrarian: The Moat Is Being Left Open on Purpose
Here's the read everyone is missing, and it's the one that should change your positioning.
The trust gap is not a problem waiting to be solved. It's a moat, and it is being left deliberately wide by exactly the players best positioned to close it.
Only entities that can underwrite disputes at consumer scale can close a recourse gap. Card networks can. Large platforms with their own fulfillment and returns infrastructure can. A state-backed scheme could. A mid-size DTC brand cannot, and neither can an independent agent startup with no balance sheet behind it.
So the 42% of merchants running pilots are not adopting AI. They're adopting a new intermediary and paying rent to it. Their agents will run on someone else's credential framework, someone else's dispute layer, someone else's identity graph. The fee on that will look, structurally, like interchange looked in 1968.
Institutional money doesn't buy the 2030 forecast. It buys the 2027 authorization framework. Watch where the capital actually goes, not where the press releases point.
Second blind spot: the 89% 'preparing' figure. I've seen that number before, in a different market. In 2018 and 2019, nearly every large enterprise had a blockchain pilot. Those pilots were not driven by expected value. They were driven by career risk — nobody wanted to be the executive explaining in 2021 why they had no strategy. Almost all of them died quietly, and the survivors survived because one specific person had one specific P&L problem to solve.
Eighty-nine percent preparing, against a 2.9% unit share, is the same signature. Defensive budget, not offensive budget. Expect a cull inside twenty-four months, and expect the survivors to have a named recourse mechanism rather than a chatbot demo.
ESTPs don't trade narratives. They trade the print. The print is 2.9% of units and 1.4% of dollars. That's the number I'm marking against, and I'll re-mark it the quarter the ceiling moves.
Takeaway: Levels to Watch
Four markers, in order of information value.
The $50 ceiling. If it moves to $150 by 2027, agentic dollar share goes from roughly 1.4% to roughly 8.6% on the model above. That's the whole category repricing in one variable.
The first consumer-scale delegated credential with a hard per-transaction cap, category scoping, and a one-tap revoke. Not the first agent. Not the first chatbot. The credential. That's the print that converts intent into flow.
The unit-versus-dollar reporting split. Units will keep being the headline because units are where the fees are. Dollars are the honesty. Any party quoting agentic commerce share without specifying which one is telling you what they want you to believe.
The first agent-legibility spec published by a merchant with verifiable warranty and return terms baked into the feed. Whoever ships that first owns the default recommendation in their category — agents route to parsable inventory the same way they route to deep pools.
Two point nine percent isn't a failure. It's a price. The only open question is who pays the spread while the ceiling holds — and who collects it the week it breaks.
