Shouting at the Machine

Since 2024, virtually every large US bank has put an AI layer between its customers and its people — chatbots in the app, voice bots on the phone line, automated verification everywhere. This study asks a simple question of the customers themselves: did complaints about digital interactions rise as AI assistants became the front door? And a harder one behind it: what exactly are people angry about — the bot, or what the bot fails to do?

To answer it we analysed the public review record of the 100 US banks in Backbase’s AI Visibility Tracker universe: a current snapshot of 2,672 review snippets across the 57 institutions with Trustpilot profiles, plus a reconstructed time series of 3,735 dated, full-text reviews (January 2023 – March 2026) recovered from archived snapshots of the same pages. Revolut and Monzo — two UK digital challengers with 458,000 combined reviews — run through the identical pipeline as a contrast baseline.

Along the way, the corpus surfaced patterns we did not go looking for: an escalation economy in which one in six angry customers now invokes a regulator; decades-long relationships ending in a single unresolved incident; estates, bereavements and expat lockouts that automated service models handle worst of all; and a small-business segment whose reviews average 1.03★. This report covers all of it.

The current snapshot is bleak in a way that is easy to misread. Of the 57 banks with Trustpilot profiles, the large retail incumbents cluster between 1.1★ and 1.9★. Capital One holds 1.2★ across 3,644 reviews; Bank of America 1.3★ across 3,018; U.S. Bank and PNC 1.3★ across more than a thousand each. The outliers at the top are credit unions and digital-first players — Navy Federal (4.5★, 47,081 reviews), SoFi (4.0★) — institutions that treat the review channel as part of the product.

Reading the fee complaints closely, most are not about the existence of fees but about their discovery: maintenance fees appearing on accounts sold as free, late fees on payments the bank itself delayed, $30 wire fees quoted as the only alternative to a blocked Zelle transfer, annual fees billed from application date rather than card delivery. The same is true of the giant “loans, mortgages & cards” category — the anger concentrates on servicing events the customer cannot see coming: escrow recalculations that move a mortgage payment by $400 a month, credit limits slashed the day after a large balance payment, loans sold to servicers the customer never chose. Opacity, not price, is the accelerant.

The dated corpus lets us move from a snapshot to a trajectory. Between 2023 and 2025–26, the average rating of new reviews at US incumbents slid from 2.39★ to 2.11★, and the one-star share rose from 60% to 68%. Over the same period the challenger baseline climbed from 3.34★ to 4.49★ — the two curves are not converging, they are crossing in opposite directions.

Reclassifying every negative (≤2★) dated review against a consistent taxonomy shows the complaint mix rotating, not just growing. Three things went up: digital-channel failures, payment/transfer delays and loan-and-card servicing. One thing clearly went down: complaints about hold times and unreachable call centres.

The single largest complaint surface in 2025–26 — bigger than fees, bigger than fraud — is the digital channel itself: apps that lock users out, verification codes that never arrive, logins that break after a phone-number change, automated systems that “do everything possible not to connect you to a human.” Taken together (app/online failures + AI/bot complaints + no-human-available), digital interaction failures appear in 38.2% of all negative reviews in 2025–26, up from 33.9% in 2023 — and that understates it, because 16.7% of negative reviews additionally describe being forced offline entirely: told to visit a branch, mail a form, or send a fax to resolve something the digital channel started.

Verdict: supported in direction, smaller in magnitude than the raw tagging suggests, and more interesting in mechanism.Explicit blame of AI, bots and automation roughly quadrupled between 2023 (0.8% of negative reviews) and 2024 (3.4%), then plateaued around 3% — it did not keep climbing as deployment widened. An earlier keyword pass on this corpus put the figure at ~6%; our stricter reclassification, which excludes incidental mentions of “automatic payments” and the like, lands lower. Customers who are angry at a bank in 2026 are still mostly angry about money mechanics: fees, holds, frozen funds, denied disputes.

But three secondary signals show AI-era service design reshaping how those failures are experienced:

• Hold-time complaints fell (15.3% → 11.3%) exactly as bots absorbed first contact — the queue got shorter because fewer people reach it.
• Digital dead-ends roseto 38% of negative reviews, and “forced offline” resolution (branch, mail, fax) rose to 1 in 6 — the human escape hatch is being removed faster than the automation can carry the load.
• The tone changed. 2023 complaints read as impatience; 2025–26 complaints read as abandonment: 40-year customers describing verification loops with no exit, disabled customers locked out abroad, executors faxing death certificates into the void.

The most under-reported number in the corpus: 15.8% of all negative incumbent reviews — 310 reviews — name a regulator, a lawsuit, or a watchdog.CFPB complaints filed, state attorneys general contacted, BBB cases opened, class actions invoked, “I will pursue legal remedies.” The review is no longer the escalation; it is the press release for an escalation already underway.

Two readings matter. First, the regulatory number is stable across all three periods — this is not a moral panic but a settled behaviour: for a meaningful minority of customers, the CFPB is now a standard step in the service journey, listed in the review the way an order number would be. Second, the one escalation category that fell — acquisition and loan-sale fallout, 5.8% → 1.4% — shows the record responding to real-world events: the Union Bank/U.S. Bank, Bank of the West/BMO, TCF/Huntington and Flagstar integrations of 2023 flooded the channel with conversion complaints that have since worked through. The record is a functioning sensor. What it is sensing in 2025–26 is service design, not merger indigestion.

6.5% of negative reviews — 127 of them — open with a tenure statement: “customer for 25 years”, “banked here since 1983”, “40 years, multiple homes and vehicles”. These are the reviews banks should read twice, because the authors are not serial complainers; they are narrating the end of the most profitable relationships the bank has. The pattern is remarkably uniform: decades of silence, one incident — a frozen account, a denied fraud claim, an unreachable department — and the discovery that tenure buys nothing at the moment it is tested.

A further 5.1% of negative reviews explicitly announce a switch in progress — accounts being closed, money being moved, “taking my business to a credit union.” Combined with the tenure citations, roughly one negative review in nine is written by a customer describing measurable, in-flight attrition rather than venting. On Trustpilot’s self-selected sample that is not a churn rate — but it is a churn narrative, publicly indexed, feeding both prospective customers and, increasingly, the AI assistants that summarise “what is the best bank?” for them. The reputational surface this study measures is the same one Backbase’s AI Visibility work shows LLMs are reading.

The smallest categories in the corpus are the most damning, because they mark the situations automated service handles worst: the moments that are rare for the customer but routine for the bank.

Forty reviews describe estates and bereavement, and they read like the same story told forty times: death certificates mailed, faxed and “lost”; beneficiary departments that cannot be called, only awaited; funds released only after the CFPB or a lawyer is engaged; one widow describing a two-year, seventeen-call effort to close a $5 account. Estate handling is a pure process problem — no credit risk, no fraud ambiguity — and it is where the gap between automated front doors and back-office reality is widest: the bot cannot process a death, and the humans behind it have been made unreachable.

The expat lockout is the same architecture failing a different way: security models that assume a US phone number and a domestic IP treat every American abroad as an attacker. Eight reviews describe multi-month lockouts while overseas — small in number, total in impact, and structurally identical to the domestic verification loop of Section 03. In every variant, the system’s designers assumed an average customer; the reviews are written by everyone else.

128 negative reviews — 6.5% of the record — are written by business owners, and they average 1.03★, the lowest of any segment we can isolate. The profile is distinct: 35% involve an account frozen, closed or funds held (versus ~6% of consumer negatives), and the sums are operational — payroll runs, contractor payments, $30,000 wire fraud on a business account denied reimbursement. De-risking behaviour that reads as inconvenience to a consumer reads as an existential event to a business: several reviews describe companies unable to pay staff for weeks while a compliance review proceeds without explanation, timeline, or a named human.

This is also the segment where the challenger contrast is least flattering to the challengers: Revolut Business draws the same frozen-funds complaints when its own risk operations fire. The difference is again resolution architecture — in-app case threads with status, versus a risk department that “does not accept calls.”

The challengers are not loved because their support is warmer — support barely features in their praise. They are loved because the product resolves things before support is needed. Positive reviews cluster around four drivers, all of them self-service outcomes: the app just works, money moves instantly and visibly, travel and FX are frictionless, and spending control (pots, budgets, instant notifications, one-tap card freeze) sits in the customer’s hand.

Three structural differences explain the gap better than any single feature:

• Failure is visible and self-diagnosable.A Monzo customer watching a transfer sees its state in real time; an incumbent customer discovers a deposit hold when a payment bounces. Half of incumbent “delay” complaints are really opacity complaints — nobody could say where the money was.
• Control sits with the customer. Card freezing, limits, notifications and disputes are self-serve. The incumbent equivalents — stop payments, travel notices, limit changes — routinely require the phone queue the customer is reviewing angrily.
• The organisation answers.Nearly every negative challenger review carries a personalised public reply routing into an in-app case thread. The reply does not fix the problem, but it closes the abandonment loop that generates the incumbents’ harshest language.

Their complaint profile is the mirror image of the incumbents’. When challenger customers are unhappy it is about risk operations— fraud outcomes, KYC freezes, compliance holds — the genuinely hard problems every regulated institution shares. What is almost absent is the reachability complaint that defines the incumbent record: “long waits / unreachable” appears in 2.1% of challenger negative reviews versus 11.3% at incumbents.

The review page is itself a service channel, and the corpus records who treats it as one. Overall, only 4.7% of incumbent review-page content is bank responses. Where responses exist, they split into two regimes with opposite outcomes:

The mechanism is the same abandonment loop as Section 04. A templated “we’re sorry, please call us at…” reply routes the complainant back into the queue they are complaining about — the response isthe dead-end, restated politely. CIT Bank posted 67 such replies against a 1.6★ rating. Navy Federal’s named-human replies, Fifth Third’s branch-level engagement and the challengers’ case-thread routing all do the one thing boilerplate cannot: move the resolution somewhere it can actually happen. The lesson is not “reply more”; it is that a reply without a resolution path is measurably worthless.

Read as an operating brief rather than a scorecard, the record points to seven moves:

• Instrument the loop, not the bot.The metric that predicts a one-star review is not bot containment rate — it is the number of customers who exit the digital channel unresolved. Every “visit a branch / mail a form / fax it” instruction is a defection event in waiting.
• Keep a human escape hatch, and make it findable. Hold-time complaints fell because fewer people reach the queue; abandonment complaints rose for the same reason. An agentic front door only works if escalation is one utterance away — the reviews punish concealment, not automation.
• Give security models an exit for legitimate users. The modal 2025–26 complaint is a lockout with no recovery path. Every fraud control needs a designed, human-reachable appeal lane — especially for the phone-number-changed, travelling-abroad and bereaved cases the models treat as anomalies.
• Build the estate journey like an onboarding journey. Bereavement is the highest-severity, lowest-volume failure in the record and a pure process problem. It is also perfectly automatable — document intake, status visibility, a named case owner — and nobody has done it.
• Treat business-account freezes as incidents, not cases. The 1.03★ segment needs timelines, named owners and interim access provisions. Silence during a compliance review is the single most litigation-threatening behaviour in the corpus.
• Fix the money mechanics AI now fronts for.Fees, deposit holds, frozen accounts and fraud outcomes still drive most anger — and most of that anger is about opacity. Show the state of the money the way challengers do, and a large share of the “delay” category dissolves.
• Work the channel — with resolution, not templates.Fifth Third’s 1.06★ → 4.2★ arc shows the record is not destiny; CIT’s 67 boilerplate replies at 1.6★ show what does not work. Respond publicly, resolve in-thread, invite the silent majority to speak — and remember the same text is now read by the AI assistants that recommend banks.