Would You Let an AI Spend Your Money?

<p>I would judge an AI shopping assistant by how clearly it lets a person define and correct a purchase. Finding a product is useful. Accepting responsibility for a payment is a different step, and the experience should make that transition understandable.</p> <p>My view is that the strongest AI commerce products will earn broader permission through bounded tasks people can inspect. A vague promise that an assistant will “take care of everything” is less persuasive than a clear account of what it can buy, under which conditions, and what happens when something goes wrong.</p> <h2>The news behind the question</h2> <p>On September 9, 2026, <a href="https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22736.html">Visa published its Trust Index for agentic commerce</a>. It reported that 23% of U.S. consumers trusted generative AI to handle payment transactions on their behalf, while 61% said they would trust Visa with agentic transactions. The Harris Poll survey was commissioned by Visa and fielded May 26–28 among 2,065 U.S. consumers; the payment question samples were smaller.</p> <p>This is reported sentiment from a U.S. survey. It does not establish actual purchasing behaviour, demonstrate product safety or measure Canadian consumers’ preferences. The release date and fieldwork date also answer different questions: the announcement is current, while the responses were collected several months earlier.</p> <h2>Make permission specific enough to use</h2> <p>Consider a hypothetical request to reorder household supplies. A useful authorization could specify approved items, acceptable substitutions, a total including delivery and taxes, and a period in which the order may be placed. A product should make those constraints visible before it acts.</p> <p>A single maximum amount is helpful but incomplete. Spending within that ceiling on the wrong item would still fail the task. So could buying from an unacceptable merchant or creating an ongoing subscription when the customer expected a one-time purchase.</p> <p>I would want the assistant to distinguish those decisions explicitly. If a requested item becomes unavailable, it should know whether substitution was already permitted. If the final cost changes, it should know whether the change remains within the agreed scope. The point is to translate the buyer’s intention into conditions the product can enforce.</p> <h2>A receipt should explain the delegation</h2> <p>The normal purchase record tells a customer what was bought and what it cost. An agent-initiated purchase should also make it possible to understand the authorization behind it: which instruction applied, which conditions were checked and where the assistant exercised discretion.</p> <p>This does not require overwhelming the customer with internal logs. A short explanation could state that the assistant reordered an approved item, used an allowed merchant and stayed within the agreed total. More detail should be available when the outcome is disputed or unclear.</p> <p>The same principle applies to a cancellation. A product should clearly distinguish cancelling an unsubmitted request, stopping a future recurring action and asking a merchant to reverse an already completed purchase. Presenting all three as one simple “undo” would hide materially different outcomes.</p> <h2>Trust is tested after the mistake</h2> <p>Much of the appeal of delegated shopping is convenience. But the buyer’s experience also includes delayed deliveries, unsuitable substitutions and requests that were interpreted incorrectly. I would evaluate the product by how well it helps resolve those situations.</p> <p>Who can the customer contact? Can they inspect the original instruction? Can they stop future actions without losing access to their records? Does the interface explain which party handles the next step? These are concrete product questions; the exact rights and responsibilities depend on the service and purchase terms.</p> <p>There is room for different preferences. Some people may want an assistant to prepare a basket for final review. Others may permit routine repeat purchases within established limits. A good product can support both while making the current permission level unmistakable.</p> <h2>What would persuade me to widen the scope?</h2> <p>I would look for evidence from actual use: purchase accuracy, disputed-action rates, time to resolve problems and whether users understand the authority they granted. A rising willingness-to-trust number would be encouraging, but I would not treat it as a substitute for those outcomes.</p> <p>Evidence that users consistently understand broad delegation and resolve mistakes easily would strengthen the case for more autonomous purchasing. Evidence of confusion about recurring charges, substitutions or responsibility would support a narrower approach.</p> <p>This is one of the practical questions surrounding <a href="https://jasonansell.ca/books/the-ai-money-revolution">The AI Money Revolution</a>. The ability to transact becomes useful when people can understand the arrangement and remain in control of their commitments. Readers can explore related work in the <a href="https://jasonansell.ca/topics/artificial-intelligence">AI topic guide</a>.</p> <p><strong>Disclosure:</strong> Jason Ansell is Co-Founder of Vector Smart Chain and author of the linked book. This article presents his editorial perspective; the shopping scenarios are hypothetical.</p> <h2>Source</h2><p><a href="https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22736.html">Visa: New Visa Research Finds Consumer Trust is Accelerating the Path to Agentic Commerce</a>, September 9, 2026; reviewed September 17, 2026.</p>