The short answer
Consumers are ready to let AI shop for them, but not to let it pay unsupervised. Visa's Trust Index for agentic commerce, announced September 9, 2026, found that just 23% of US adults would trust generative AI to execute a payment on their behalf, while 61% said they would trust Visa to handle an agent-initiated transaction — the highest of any brand tested. Among 18-to-34-year-olds that rose to 68%, and among frequent AI users to 71%. The survey was run by The Harris Poll among 2,065 US adults, fielded May 26–28, 2026.
The practical reading for builders: the discovery and selection half of shopping can be delegated to an agent, but the money-movement half needs a trusted rail and a human checkpoint. Design agentic checkout so the agent orchestrates the experience while a tokenized, revocable payment credential and an explicit approval step handle the transaction. That split — smart agent, trusted settlement — is what closes the gap.
What Visa's research actually found
Agentic commerce is the emerging pattern where a software agent does more than answer questions — it searches, compares, selects, and ultimately pays for goods or services with limited human input. Visa's new Trust Index set out to measure whether consumers are actually willing to hand that last step to an AI, and the answer is a clear "not yet, not alone."
The standout figures: 72% of US adults have already used an AI assistant, but only 23% said they would trust generative AI to execute a payment for them. When Visa asked which brand people would trust to handle an agent-initiated transaction, 61% named Visa — ahead of the technology and social-media names in the same test — rising to 68% among adults aged 18 to 34 and 71% among frequent AI users. Group President Oliver Jenkyn framed the takeaway directly: "Trust will be foundational to driving agentic commerce adoption." The underlying survey was conducted by The Harris Poll on its Omnibus platform between May 26 and 28, 2026, among a nationally representative sample of 2,065 US adults.
Read as an engineering signal rather than a marketing stat, the numbers describe where users draw the line. They are comfortable with AI as a shopper and researcher. They are not comfortable with AI as an unsupervised spender. That distinction is exactly where product and payments architecture decisions get made, and it maps cleanly onto the fintech and payments stack any commerce agent has to plug into.
Why the trust gap exists
The gap is rational, not irrational. Paying is the one action in a shopping session that is hard to reverse and easy to abuse. A wrong product recommendation costs a few seconds; a wrong or manipulated payment costs money and creates a dispute. Consumers intuitively apply a higher bar to the step that touches their funds, which is why usage of AI assistants can be high while willingness to let them pay stays low.
Two technical realities reinforce that instinct. First, large language models are probabilistic and susceptible to prompt injection and manipulated content — a shopping agent reading a hostile product page or review can be steered toward the wrong action. Handing such a system unrestricted authority to move money is a risk consumers sense even if they cannot name the attack. Second, people already have a mental model for who is accountable when a payment goes wrong: their card network and bank, with chargebacks and fraud protection. That is why trust concentrates on the payments brand rather than the AI — the network is the party they expect to make them whole.
The design implication is that trust is not won by making the model sound more confident. It is won by keeping the agent's autonomy scoped, making its actions visible and reversible, and routing the actual money movement through the rail users already trust. The agent earns the right to transact by proving it operates inside guardrails, not by asking the user to believe it won't make a mistake.
What it means for US & EU software teams
First, treat trust as a product surface, not a slogan. The teams that win agentic commerce will be the ones that make the agent's reasoning, state, and limits visible, and that put an explicit human-approval checkpoint before any charge above a user-set threshold. Silent, fully autonomous payment is precisely the behavior the research shows consumers reject. Approval gates, clear spending caps, and a plain-language summary of what the agent is about to buy are conversion features here, not friction.
Second, keep the agent away from raw payment credentials. The durable architecture is a split: the agent orchestrates discovery and selection, while payment executes through tokenized, delegated-authority mechanisms on an established network. The agent presents a scoped, revocable token carrying explicit merchant and spending limits — never a card number. That both matches the trust data and shrinks your payment API integration and PCI DSS scope, because sensitive credentials never sit inside the agent's context.
Third, the compliance surface moves with the money. In the US, any flow that authorizes payments pulls in PCI DSS and strong-authentication expectations. In the EU, an agent that decides and pays on a user's behalf touches GDPR duties around automated decision-making, plus the data-minimization question of what it collects to shop. Auditable, timestamped logs of every agent decision — what it saw, what it chose, and what the user approved — are the difference between a defensible system and an unexplainable one when a disputed transaction lands. Building those logs in from the start is far cheaper than reconstructing intent after the fact.
How to build trust into agentic checkout
- Split orchestration from settlement. Let the agent handle search, comparison, and cart building; route the actual authorization through an established payment network with tokenized credentials rather than letting the model touch card data.
- Add human-in-the-loop approval with limits. Require explicit user confirmation for purchases, and let users set per-transaction and per-period spending caps and allowed-merchant rules the agent cannot override. Make autonomous spend opt-in and bounded.
- Use delegated, revocable authority. Issue the agent a scoped token with a clear mandate and expiry, revocable in one action. If the agent misbehaves or is compromised, the blast radius is a capped, cancelable credential — not the user's account.
- Make reasoning and state visible. Show what the agent is doing, why it selected an item, and where it is in the flow. Visible reasoning and clear trust cues are what move users from curiosity to delegation.
- Log everything for audit and dispute. Record each decision, the inputs behind it, and the user's approval, with timestamps. That record is your defense for PCI DSS, GDPR automated-decision duties, and any chargeback that follows.
Frequently asked questions
What is the Visa Trust Index for agentic commerce?
It is research Visa announced on September 9, 2026, measuring how much consumers trust AI to shop and pay for them. The survey was run by The Harris Poll on its Omnibus platform, fielded May 26–28, 2026, among a nationally representative sample of 2,065 US adults, with brand-specific payment-trust questions asked of roughly 1,030 respondents each. Agentic commerce means software agents that search, select, and pay with limited human input.
How many consumers trust AI to make a payment?
Only 23% of US consumers said they would trust generative AI itself to execute a payment, even though 72% have already used an AI assistant. Trust rose to 61% when a familiar payments brand was involved — the share who named Visa as the company they would trust for an agent-initiated transaction — reaching 68% among adults aged 18 to 34 and 71% among frequent AI users.
What does the agentic-commerce trust gap mean for developers?
Trust is now a design constraint. Users will delegate discovery and selection to an agent long before they let it move money unsupervised, so agentic checkout needs explicit human-approval checkpoints, spending limits, visible reasoning and state, and payment routed through established, tokenized rails rather than the agent handling raw card data. The trusted payment network becomes the settlement layer while the agent runs the shopping experience.
How should agentic checkout handle payments securely?
Keep the agent away from raw credentials. Use network tokenization and delegated authority so the agent presents a scoped, revocable token with explicit spending and merchant limits instead of a card number. Require a human confirmation for the authorization, log every agent decision for auditability, and settle through an established payment network. This mirrors the research: consumers accept an AI doing the shopping but want a trusted brand behind the money movement.
Why does this matter for FinTech and e-commerce teams?
Agentic commerce reshapes both the checkout funnel and the compliance surface. In the US and EU, an agent that authorizes payments touches PCI DSS scope, strong-authentication expectations, and, in the EU, GDPR duties around automated decision-making and the data the agent collects. Teams that build approval controls, tokenized payments, and auditable decision logs in early can launch agentic checkout without inheriting a trust and compliance problem later.
Sources
Visa — New Visa Research Finds Consumer Trust is Accelerating the Path to Agentic Commerce
PYMNTS — Visa CEO Says AI Shopping Has Arrived but Agentic Payments Haven’t
Finextra — Consumer trust in agentic payments continues to lag