JPMorgan puts up gates against AI agents. What happens when customers want a better deal?

JPMorgan Chase is setting boundaries around outside AI agents, according to reporting on its October 13 earnings call. The immediate issue is protecting customer accounts. For AEO practitioners, the wider question is different: what happens when an assistant advises an existing customer that they would be better off elsewhere, even if it cannot execute the move? Access controls and a defensible reason to stay address different parts of that decision.

What the bank remarks establish

Business Insider reports that CEO Jamie Dimon raised heightened security risks from outside agents, possible data misuse and liability concerns. CFO Jeremy Barnum specifically objected to credential sharing and screen scraping, and said the bank was blocking access to the extent it could. Dimon also said JPMorgan was developing its own agentic tools for clients. The reported position concerns how outside agents gain access, rather than rejecting the technology altogether. [1]

Those concerns deserve attention. An assistant authorised to inspect information is different from one empowered to move money. Authentication, permissions and accountability matter. The reporting does not establish that every outside agent is blocked, provide a complete technical policy or settle the legal liability of a particular customer arrangement.

Wells Fargo CEO Charlie Scharf offered a more sceptical view of potential deposit flight. MarketWatch reports that he pointed to longstanding higher-rate alternatives and customers' existing ability to move money online. Barron's account adds important context: he questioned how soon consumers would give agents unilateral authority to move money to institutions they did not know or understand. These are reported management views, not measured estimates of AI-induced deposit loss. [2][3]

Advice does not require authority to act

The disagreement raises a useful research question: does making a decision easier change how often people act on it? The existence of a better alternative does not guarantee that customers will find it, understand it or pursue it. They may value their current provider. They may also postpone reviewing the relationship because the effort feels disproportionate to the potential benefit.

AI could reduce some of that work by comparing public terms, identifying missing information or helping a customer prepare questions using information they lawfully provide. That is a plausible mechanism, not proof of an impending wave of departures. Its magnitude would depend on adoption, trust, advice quality and the practical barriers that remain. Advice, customer approval and execution should be observed separately; an answer recommending a switch does not show that an agent can make it happen.

This is a new bank-sector proof point for the question explored in What would AI advise your existing customer to do?, not evidence that its retention hypothesis has been confirmed. Through the System 0 lens of delegated decision-making, a relationship could be evaluated again after purchase. System 0 is used here as a conceptual lens, not an empirical result from these bank calls.

Five answers an agency should deliver

A customer may stay with a bank because of reliable service, useful branches, integrated products or confidence that a problem will be resolved. Another may stay because reviewing alternatives never reaches the top of their list. Similar questions can be investigated for telecom providers such as Bell and Rogers, insurers and subscription businesses. These are illustrative applications, not findings about those companies.

The agency brief should move beyond whether the brand appears in an AI answer. It should investigate which action the assistant advises, whether that advice is right and which business decision follows. The following five outputs are AEO Updates recommendations, not validated interventions or features guaranteed by a visibility dashboard.

1. Which customer situations make AI recommend leaving?

Build realistic scenarios from customer research, service enquiries and current product information. For a bank, these could include reviewing fees, comparing returns on idle cash or deciding whether a package still fits. For a telecom provider, they could include an expiring promotion, unused allowances, a household bundle or a service complaint. Include customers whose arrangement represents genuinely good value; otherwise, the study will be designed to discover reasons to leave.

Use neutral questions and record staying, changing products with the same provider, negotiating, switching and seeking more information as distinct actions. Preserve the alternatives, conditions and decisive reasons attached to each response. Specify models, interface, market, date, scenario and repetition; label provider API tests separately from consumer-interface observations. If reporting a switching-recommendation rate, divide primary-switch responses by all valid responses, including those seeking more information, and report failed or excluded observations separately. That rate describes the tested answers, not the percentage of customers who will leave.

2. Is the advice accurate once the whole relationship is considered?

A superficially attractive alternative may look different once eligibility, full costs, temporary rates, bundled benefits and other obligations are included. Verify the material facts supporting each recommendation against current applicable terms. Does the offer apply to this customer? Would changing one service remove another discount? Has the assistant compared equivalent products and acknowledged facts it cannot establish?

Require a distinction between accurate advice that exposes a weak offer and inaccurate advice caused by a mistaken comparison. The first may require a commercial or service decision. The second calls for investigating the information and how it is being interpreted. A favourable answer is not automatically accurate, and a citation does not by itself establish that the cited terms were understood correctly.

3. What makes AI recognise that staying is the better choice?

Identify which advantages the assistant recognises, which it overlooks and whether they matter to the customer in the scenario. Reliable service, appropriate coverage, useful support and lower overall household cost can justify staying. Each needs evidence and context rather than a generic brand claim.

The deliverable should identify defensible reasons to remain a customer, not simply recommend publishing more content about the brand. It should also reveal where the business lacks a convincing answer. Clearer wording cannot fix an offer that does not meet the customer's needs, and retaining a customer is not inherently the correct outcome of every comparison.

4. What information or offer changes the recommendation?

A customer conversation develops. Someone may initially ask for the cheapest option, then explain that reliability matters, a bundle discount applies or a competing offer has restrictions. Follow the conversation and identify the turning point without treating a single response as stable evidence across models or time.

An illustrative finding might be that an assistant changed its advice after a lost bundle discount was included. That would identify a particular explanation worth improving. It would not establish that changing a public page will cause the same result in naturally encountered AI answers. Distinguish controlled tests, where information is supplied directly, from observations of what customers encounter without that intervention. Label hypothetical offers clearly and do not imply that unavailable concessions have already been approved.

5. What should change, and how will the agency check whether it helped?

Require an action brief containing the customer situation, verified issue, proposed response, accountable team and retest plan. An outdated explanation belongs with the information owner. An uncompetitive eligible offer belongs with pricing and retention. A recurring service failure belongs with operations. Repeating the assistant test can check the answer, but it cannot establish the customer outcome.

Then investigate whether customers actually seek this advice and act on it. Interviews and consented decision journeys can establish relevance. Controlled customer tests or suitable longitudinal work can assess sustained retention and contribution after discounts, service and implementation costs. A simulated switching-recommendation rate is a research finding, not a churn or revenue forecast.

Access policy and customer value need separate decisions

JPMorgan's reported position highlights the need to decide how an outside agent can interact safely with a business. Providers should define what information can be shared, which permissions are required and which actions need explicit customer approval. This brief does not recommend credential sharing, bypassing a provider's restrictions or unauthorised account changes.

Alongside that work, providers should examine the advice customers receive outside their systems. A secure access policy answers who can act and under what conditions. Customer research asks why someone would want to act. Neither substitutes for the other, and preventing an unauthorised transfer does not establish that a customer will choose to stay.

AEO Updates Takeaway

The practical brief for an agency is to show when AI advises existing customers to leave, whether the advice is right, what changes it and which response the business should test. JPMorgan's gates address an access concern; the customer-loyalty question requires evidence about value and behaviour. The advantage lies in knowing whether to improve information, change the offer, fix the service or protect a strength, rather than turning favourable AI answers into an unearned retention claim.

References

[1] Business Insider: Jamie Dimon says JPMorgan Chase is putting up gates to protect against 'rogue agents', October 13, 2026. Original reporting on the bank's earnings-call remarks; not an independently audited access policy.

[2] MarketWatch: Wells Fargo CEO does not believe AI agents will lead to deposit flight, October 13, 2026 call coverage.

[3] Barron's: Wells Fargo CEO dismisses AI deposit flight concerns as 'overdone', October 13, 2026 call coverage. MarketWatch and Barron's are both Dow Jones publications, not two independent empirical studies.

[4] JPMorganChase: Third-quarter 2026 earnings-call announcement and Wells Fargo: Quarterly earnings confirm the event chronology. Management remarks above are attributed to the original reporting; complete call transcripts or audio were not independently reviewed.