How much do bank customers trust AI, and where is that trust heading?

Executive summary

Bank customers accept AI that informs, flags and handles routine servicing, and resist AI that decides or moves money without asking. Stated trust in AI for money matters sits at roughly one in four to one in three adults, has been flat since the 2023 step-change around ChatGPT, and is falling among adults under 30. Banks hold a small and unstable lead over fintechs and big tech on trust with money and data. The practical risk is not losing the customer to Google or Meta; it is losing tasks one at a time around an account that stays put. None of the three stakeholder beliefs survived intact. Each holds in a narrower form.

The pattern that fits all four topics: customers grant AI an assistive role readily, an approved-action role grudgingly, and an autonomous role rarely, with the line drawn more by how much the AI does without asking than by which task it does. Trust is earned per interaction through performance and recourse, not inherited from the bank's brand or from general AI familiarity. Migration pressure is real but arrives task by task, led by money questions going to AI chatbots and payments going to apps, while the primary account stays through inertia and deposit insurance.

Task-level trust: what customers will and will not let AI do

Stated comfort runs in a clear order, from AI that watches and warns, to AI that advises, to AI that acts, to AI that decides on credit. The order looks like a task ranking, but the same surveys show it is mostly an autonomy and stakes ranking.

Four things complicate the simple reading. First, the autonomy drop appears inside every task concept the FCA tested, and at the autonomous level the spread between tasks is about the same size as the drop from autonomy itself (T1-04). Second, no instrument found holds stakes constant and varies only the task; fraud flagging is both low-stakes and a flag-not-decide role (T1-01). Third, behaviour contradicts stated preference. In incentivised investment experiments, 38% delegated to an algorithm against 17% to a fixed-fee professional, while the same people's self-reported real-life delegation ran the other way (T1-11). People also resist delegating to human agents about as much as to AI, which points to a general preference for control rather than distrust of AI (T1-12). Fourth, what customers say would make them comfortable is control and protection, not the provider: human override, per-action approval, reimbursement, undo and proven accuracy top both the US and UK lists, and only 13% of UK respondents picked "a provider I already trust" (T1-03).

The academic base is thin on this exact question. The one meta-analysis found excludes task-type comparisons by design, and the main review calls task factors the least studied (T1-14). Every experiment read uses crowdworker or student samples, not bank customers.

Drivers: what actually moves trust

The drivers that measurably change behaviour are experienced performance, real recourse and familiarity. The drivers stakeholders expect to matter most, brand trust and privacy concern, move behaviour least.

No source measures all of these drivers in one instrument on bank customers, so their relative weights cannot be stated from external evidence. The hints point to product experience and familiarity sitting above institution, and to an asymmetry in which a first error costs more than good performance earns.

Durability: does trust stay with banks?

In the only direct comparisons found, consumers trust banks more than fintechs and far more than big tech with their money and data. That lead is small, was zero for two years, and is thinner among younger and minority customers. Meanwhile discrete tasks have already moved.

What has moved. Half of US households use a nonbank payment app with a stored balance (49.7% in 2023), and 68.5% of households aged 25 to 34 do (T3-03). Buy now, pay later rose from 12% to 16% of adults between 2022 and 2025 (T3-17). After the Wells Fargo scandal, mortgage borrowers in exposed counties shifted toward fintech lenders (T3-05). For the information task, 18% of guidance-seekers have used AI tools, and ChatGPT use has more than doubled since 2023 (T3-09, T3-10). One neobank reports two thirds of its 8.6 million active members treat it as their primary relationship, by its own definition (T3-08, company disclosure).

What has not. Bank account ownership held at 94% through the same years (T3-17). Only a third of banked households use payment apps for core account functions (T3-03). Fully insured retail deposits did not run in the 2023 bank failures, though uninsured consumer dollars left within days (T3-06). Big tech's attempts at accounts and credit were wound down or stayed bank-issued: Google Plex, Diem, Apple's own BNPL, and Apple Card's issuer move to Chase (T3-07). AI labs are used but not trusted for money: no more than three in ten have even some confidence in AI's financial expertise against 79% for advisors (T3-10), an AI label lowered trust in identical financial information (T3-11), and only 13% of UK consumers would give AI real-time access to their bank data (T3-12).

What is untested. No source compares trust in a bank's AI agent with a fintech's, big tech's or AI lab's agent for the same task, and no source follows customers moving an AI-delegated task between providers. There is also no second wave of the bank-versus-fintech-versus-big-tech data-trust ranking, so whether the gap is narrowing cannot be stated. Stickiness is behavioural and may reflect inertia and deposit insurance rather than trust; the UK competition authority treated 3% annual switching as a problem, not as loyalty (T3-13).

Projection: 2026 to 2030

Every repeated measure points the same way: one step down in sentiment around ChatGPT, then a plateau. Use keeps rising while trust does not. The cohort that was expected to carry trust upward is moving the other way.

Cohort replacement will not do the work on its own. Younger adults are more confident in AI for money guidance than older ones, 36% of millennials against 15% of boomers, so replacing cohorts would lift the average, but only toward a ceiling of about one in three, and the peak is among people in their late 20s and 30s rather than the youngest (T4-06). Over time, under-30s are the group moving fastest against AI (T4-01, T4-04).

The banking analogues say adoption precedes stated trust rather than following it. Mobile banking use roughly doubled between 2011 and 2015 while 73% of non-users still cited security concern and 40% said they did not trust the technology; the main barrier was no need, not distrust (T4-10). Mobile payments in stores reached 11% of in-person payments after a decade (T4-11). For AI agents specifically, the FCA's 2030 review names trust, control and access as preconditions and offers no numerical forecast (T4-12). The honest projection is a plateau through 2030 with task-by-task gains for low-autonomy, reversible uses, and no basis for a numerical forecast of bank-specific AI trust, because no bank-specific repeated item exists.

Implications for the bank

1. Design around autonomy level, not task list. The strongest evidence says trust falls in steps from recommend, to act-with-approval, to act-alone, inside every task (T1-04, T1-02). Build each AI capability so it can start at the assistive level and earn its way to approved action, with autonomous action reserved for low-value, reversible moves such as transfers between a customer's own accounts.
2. Lead with watching and warning. Fraud flagging is the only AI encounter a majority of customers felt positive about, at 70% (T1-01). Alerts, summaries, lookups and routine servicing are accepted. Credit decisions and unapproved money movement are not, and should keep a visible human or customer decision in the loop.
3. Build recourse, not reassurance. An appeal route moved perceived fairness; a "human reviews this" label did not (T2-04). Reimbursement, undo and per-action approval are what customers themselves name as the conditions for letting AI handle money (T1-03). Explanations should be designed to help customers catch a wrong decision, since richer explanations made people more confident and less accurate (T2-01).
4. Plan for the first error. One visible mistake cut adoption by a quarter in experimental data (T2-02). Error-recovery experience, fast human escalation and a stated reimbursement position are part of the launch, not the backlog.
5. Do not count on the brand or on Gen Z. Institutional trust in banks is low and barely transfers to AI (T2-10, T2-11). Younger adults use AI more and trust it less each year (T4-04). Measure trust in your own AI, in your own customers, by autonomy level, rather than assuming either reservoir.
6. Model the threat as task leakage, not account loss. Money questions are going to AI chatbots and payments to apps while the account stays (T3-03, T3-10, T3-17). The defensive priority is to make the bank's own assistive AI the first place customers ask, before the question leaves for a provider they trust less but use more.
7. Run the two studies nobody has run. Nothing in the external evidence varies provider against the same task, or links what customers say to what they later do with a real bank AI. A conjoint test of bank, fintech, big tech and AI lab agents on identical tasks, and a survey-to-usage linkage on your own AI features, would answer the two questions this research leaves open.

Gaps and limits of the evidence

These are documented negatives: what was searched for and not found, with the source that would answer it.

• No source compares AI agents across provider types for the same task. Belief 3's central claim is untested. A conjoint or split-sample survey randomising provider against task would answer it.
• No experiment varies only the banking task (pay a bill, recommend, dispute) with undo, on bank customers. Every delegation experiment uses crowdworkers, students or general populations.
• No bank-specific repeated AI-trust item exists. All trend series ask about AI in general or business in general. A Fed SHED or Pew module repeated across waves, or J.D. Power's items (blocked at 403), would provide one.
• No second wave of the bank-versus-fintech-versus-big-tech data-trust ranking. The NY Fed module ran once in 2020. Whether the gap is narrowing is unknown.
• No single instrument measures all drivers together on bank customers, so relative weights are unstated.
• No same-person link between stated trust and later use of a real bank AI. Only within-study experimental contrasts exist.
• Open-banking consent research from UK and US regulators (FCA, CMA, Open Banking Ltd, CFPB Section 1033) was not reached. Only a 2020 US survey and a 2024 Canadian survey compare providers for data sharing.
• Edelman Trust Barometer sector and AI series could not be fetched from a primary source (403 on every route); no Edelman figure is used.
• Atlanta Fed payment diaries 2015 to 2022 returned 404, so the payments migration series covers only 2023 to 2024.
• Non-US regulator surveys outside the UK (EU, Canada beyond one bank poll, APAC) carried no consumer AI-by-task items that could be retrieved.
• Agentic payments uptake has no disclosed consumer count from any network or issuer; launch press and pilot counts were excluded.
• Several full texts were paywalled or blocked (Castelo 2019, Jussupow 2020, Dietvorst 2018, Athey 2017, Germann and Merkle 2023, Lee and See 2004, Glikson and Woolley 2020), so rows resting on them are rated Low or omitted.

Method and sources

Every statistic in the findings table was re-checked against the page it is cited to. The research ran on 2026-10-08 as a parallel desk-research sweep: four topics, each split into three or four non-overlapping angles, one research agent per angle (15 in all) plus one targeted recovery pass, followed by a separate compile pass per topic that merged duplicates, pruned weak sources, re-fetched every surviving URL and confirmed that each cited number appears in the fetched text.

Evidence was ranked: regulator and administrative data first, then research with a named method, then regulatory text, then company disclosures and self-reported surveys, which are flagged as such and rated Low where the sponsor has an interest in the result. Vendor surveys of their own users, marketing, Wikipedia and secondhand trade press were excluded. Each of the three stakeholder beliefs was treated as a claim to test; researchers were required to report counter-evidence at equal weight, and each row records the direction of its evidence. Across the table, 10 rows support the belief they bear on, 14 contradict it and 36 are mixed.

All 95 unique source URLs resolved. About 20 individual statistics were corrected or removed because they did not appear on their cited page as stated, including figures read off charts and filed as text, wrong populations, and one reversed direction; a further dozen labels, dates and attributions were corrected. Sixteen proposed findings were dropped after verification or to hold row caps. Those corrections are logged per topic in the project's compile notes.