The Agentic Consumer

What happens when AI starts making purchase decisions for us — the infrastructure that arrived, the adoption data, the trust ceiling, and the reliability gap nobody advertises.

Key findings

• The checkout rails for AI agents were built in roughly eighteen months: OpenAI and Stripe shipped Instant Checkout and an open Agentic Commerce Protocol, Google announced its Agent Payments Protocol with 60+ partners, Perplexity partnered with PayPal, Amazon deployed agentic buying — and Visa and Mastercard both launched agent-payment stacks within a single day of each other. ^{[1][2][3][5][6]}
• Demand followed fast: AI-referred traffic to US retail sites grew 693% year over year in holiday 2025 and converted 31% better than other channels, per Adobe's trillion-visit panel. ^[9]
• But there is a hard trust ceiling: willingness to let AI actually make the purchase decision peaked at 11% even for low-stakes categories (Gartner), only 8% would pay for an auto-purchasing assistant (Capgemini), and roughly 3% of transactions today involve an AI agent at all (Checkout.com). ^[11][12][13]
• Agents are measurably manipulable: in peer-reviewed marketplace experiments, moving a product to the top row raised an agent's selection rate ~5x, platform endorsements swayed choices, and agents concentrated demand on a handful of 'modal' products. ^[16][17]
• Nobody agrees on the size of the prize: credible 2030 forecasts for US agentic commerce span $190 billion to $1 trillion — a 5x spread driven almost entirely by incompatible definitions of the word 'agentic.' ^[14][15][20]

01The infrastructure arrived before the behavior

Most technology shifts start with behavior and wait for infrastructure. Agentic commerce inverted the order: between April 2025 and September 2025, virtually every layer of the payments and platform stack shipped machinery for AI-initiated purchases — before more than a sliver of consumers had asked for it.

The pattern matters more than any single launch: when Stripe, Google, PayPal, Visa, Mastercard, Amazon and OpenAI all build rails for the same behavior within eighteen months, they are not reacting to demand — they are positioning for where they believe demand is going. ^{[1][2][3][4][5][6][7]}

02The demand data: explosive, from a small base

Adobe Analytics — whose panel covers over a trillion visits to US retail sites — provides the cleanest behavioral series. Traffic from generative-AI sources to US retail was up 1,200% by February 2025 versus July 2024 ^[8]; by the 2025 holiday season it was growing 693% year over year, with November alone up 769% — the largest gain of any industry Adobe tracks. ^[9]

Salesforce's platform data pushes the claim further: it estimates AI and agents influenced 20% of global online holiday sales in 2025 — roughly $262 billion — and that retailers deploying shopper agents saw materially faster sales growth. ^[10] That figure deserves its caveat in the same breath: 'AI-influenced' is Salesforce's own broad, undisclosed attribution model, not a count of agent-executed transactions.

03The trust ceiling

The gap between 'AI helps me shop' and 'AI shops for me' is the defining number of this field, and three independent studies locate it consistently low.

Capgemini's twelve-country study adds the texture: 76% of consumers want to set explicit boundaries for AI assistants, and 69% would accept AI product suggestions — if the reasons are given. ^[12] Checkout.com found 24% say they will never delegate purchases to AI, and 27% trust no organization at all to operate a shopping agent on their behalf. ^[13] Delegation of discovery is happening now; delegation of decision is being rationed by trust.

04How agents actually choose — and how they can be steered

Inside an agent's purchase decision — what the experiments show

1. Scan
   The agent reads listings as structured text — position and format matter enormously.
2. Filter
   Price, ratings, review counts and machine-readable specs cut the field.
3. Weigh
   Platform endorsements and corroborated signals sway the model — 'Sponsored' tags can hurt.
4. Shortlist
   Demand concentrates on a few 'modal' products; everything else is invisible.
5. Human approves
   The person picks from the shortlist — recognition and reviews close the sale.

If agents make decisions, the trillion-dollar question becomes: what moves an agent? Two 2025 research programs give uncomfortable answers. Researchers from Columbia and Yale built a simulated marketplace ('What Is Your AI Agent Buying?', accepted to the ACM Web Conference 2026) and found that simply moving a product from the bottom-right of results to the top row raised one agent's selection rate roughly fivefold; platform endorsements like 'Overall Pick' increased selection while 'Sponsored' tags decreased it; and agents concentrated demand on a few 'modal' products far more than human shoppers do. ^[16]

Microsoft Research's Magentic Marketplace — an open-source simulation of 100 customer agents and 300 business agents — found frontier-model agents susceptible to manipulation tactics including fake credentials and prompt injection, subject to a 'paradox of choice' where welfare falls as options grow, and biased toward the fastest responder rather than the best offer. ^[17]

On the human side of the loop, a four-country survey of 4,250 adults found price is the top factor when choosing among AI recommendations (32%) — but 89% still say seller brand recognition matters and 92% prioritize reviews. ^[19] Even in agent-mediated commerce, the machine shortlists; reputation still closes.

05Can agents even shop yet?

The reliability record is the least-advertised dataset in agentic commerce. ShoppingBench, a benchmark of realistic, intent-grounded shopping tasks, found even the best-performing language agent succeeded on fewer than half of tasks — and in a related benchmark, success fell from 51.8% to 27.2% as implicit requirements accumulated. ^[18]

The fair counterpoint: production flows like Instant Checkout and Buy for Me are deliberately narrow — single items, scoped payment tokens, human confirmation — precisely because their designers know open-ended agent shopping is not reliable yet. ^[1] The constraint is the product. Benchmarks measure the frontier; checkouts ship the safe interior of it.

06The forecast fog

How big does this get? The honest answer is that the most credible institutions in the world disagree by a factor of five about the same country in the same year — and the disagreement is definitional, not analytical.

Gartner's forecast trio frames the enterprise side: 40% of enterprise apps featuring task-specific agents by end-2026 — and, in the same breath, over 40% of agentic-AI projects canceled by end-2027 for cost or unclear value. ^[11] Explosive adoption and mass abandonment are both officially forecast. That is not a contradiction; it is what an immature market looks like.

Limitations

• Several primary pages (Gartner, Salesforce, OpenAI, Morgan Stanley, Adobe holiday post) blocked automated verification; their figures were cross-checked through at least one reputable secondary outlet, but exact wording rests on that coverage.
• Salesforce's 'AI-influenced 20% / $262B' uses a proprietary, undisclosed attribution model — treat it as a vendor-modelled estimate, not a measurement.
• Gartner's 11% delegation ceiling comes from a small US sample (n = 322, January 2026).
• Adobe's growth percentages start from a very small base; AI referrals remain around 1% of total retail traffic in absolute terms.
• Benchmark failure rates (ShoppingBench <50%) test research agents on open tasks, not the deliberately constrained production checkout flows — the two should not be equated.
• Every figure here dates from April 2025–June 2026 in a field that moves monthly; date-stamps are part of the data.

Conclusion

What happens when AI starts making decisions for us? On the evidence: it already shortlists for us, increasingly checks out for us in narrow lanes, and is not yet allowed to decide for us. The traffic curves say the front end of the funnel is being rewired now ^[8][9]; the trust surveys say the last step remains human by a wide margin ^[11][12][13]; the agent-behavior research says whoever understands how machines choose will quietly own the shortlist. ^[16][17]

The window is defined by that asymmetry. When the trust ceiling lifts — and every generational cohort study suggests it lifts with Gen Z — the brands already built into the agents' consideration set will find the transition invisible. The brands that waited will find it unaffordable.