Bernstein Insights: Consumer-grade Agents Have Become a Trend—Who Is Most at Risk in the Financial Industry?

Summary: Insurance and banks feel the pressure first, while payment networks actually benefit

TL;DR:

· The first thing AI Agents disrupt is not financial products themselves, but the "consumer inertia" the financial industry has long relied on. Bernstein insurance renewals, low-interest deposits, idle cash in brokerage accounts, and top-of-wallet credit card status could all come under pressure from automated price comparison and cash optimization.

· Bernstein's real bottleneck is not technology, but permissions, trust, and liability. Bernstein banks, brokerages, and insurers still control account and data access, and consumers are more willing to let AI "assist decisions" rather than operate fully autonomously.

· The Bernstein shock will not hit evenly. Bernstein businesses that rely on user stickiness and switching costs face greater pressure, but payment networks like Visa and Mastercard could benefit from rising demand for identity verification, tokenization, risk control, and dispute resolution.

· The most important thing to watch for Bernstein going forward is whether Agents truly gain "execution rights." Bernstein factors include the degree of consumer authorization, whether financial institutions open data interfaces, and whether operating metrics such as insurance renewals, deposit stickiness, and cash sweep begin to change.

Editor's note: Consumer-grade AI Agents are moving from "helping users answer questions" to "completing tasks for users." Bernstein noted in its latest report that after launch, Muse quickly rose to the top of the U.S. App Store, reaching about 2.8 million downloads, and users have already begun using it to book services, cancel subscriptions, compare insurance, fill out forms, and contact customer service. At the same time, a group of financial stocks that rely on consumer stickiness and operational friction saw notable declines.

The market's most intuitive concern is whether AI Agents will directly bypass banks, insurers, brokerages, and credit cards. But what Bernstein is really discussing is not simple "technological replacement," but a more fundamental question: if AI can continuously compare prices, switch products, and move money for consumers, will the profits the financial industry has built on consumer inertia, information friction, and switching costs begin to be compressed?

The report argues that insurance renewals, bank deposits, idle cash in brokerage accounts, and top-of-wallet credit card status could all be affected. But this does not mean Agents will soon be able to fully take over financial decisions. Financial institutions still control accounts, data, and transaction permissions, consumers may not be willing to hand their money entirely to AI, and liability and regulatory frameworks have not kept pace with technological development.

Therefore, the impact of this wave of Agents on the financial industry may not be an all-out disruption, but a redistribution of value: the more a segment relies on consumers "not acting" to make money, the greater the potential pressure; the more it can provide infrastructure for identity verification, payment security, risk control, and data interfaces, the more important it may become instead.

The following is a compilation of the original text:

After Muse launched, financial markets quickly began trading a new question: if consumers have an AI Agent that can continuously compare prices, switch products, cancel subscriptions, and even move funds on their behalf, what will traditional financial institutions rely on to retain customers?

Bernstein statistics show that since Muse was launched, insurance, large banks and credit card issuers, regional banks, brokerages, and mortgage lenders have all experienced varying degrees of decline, with mortgage-related companies falling the most; the payments sector has been relatively limited in the impact it has suffered.

The logic behind this market reaction is not complicated.

A considerable portion of the financial industry's profits does not come because consumers make the wrong decision every time, but because consumers simply will not continuously optimize their choices.

And what AI Agents are most likely to change is precisely this.

The first thing AI Agents disrupt is "consumer inertia" in the financial industry

Insurance is the most typical example.

Many users will directly renew their policies after they expire, rather than re-comparing prices, coverage, and products from different companies every year. As long as this renewal inertia exists, insurance companies have a certain degree of pricing power.

Banks and brokerages also have similar logic.

Consumers will not compare deposit rates at different banks every day, nor will they continuously deal with idle cash in brokerage accounts. As a result, low-yield deposits can remain in banks for a long time, and brokerages can also earn revenue from businesses such as cash sweep (idle cash aggregation).

The credit card industry relies on another habit.

Consumers often use the same credit card for a long time, and this position of "using a certain card by default first" is usually called top-of-wallet. Through points, cashback, and long-term usage habits, banks turn a card into the default choice when consumers pay.

AI Agents have the potential to weaken these advantages at the same time.

If an Agent can automatically compare insurance quotes, find higher deposit yields in real time, transfer idle cash from brokerage accounts to higher-yield products, or compare cashback, points, and interest rates across different credit cards before every purchase, then the "search—compare—switch" process that consumers previously had to complete on their own will be greatly compressed.

Bernstein therefore argues that automated cash management could make deposit migration even easier, thereby driving up bank funding costs and compressing net interest margins; brokerages' cash sweep revenue could come under pressure; and credit card issuers could lose some of the advantages tied to top-of-wallet status and long-term usage habits.

What is truly changing here may not necessarily be the financial products themselves, but rather the declining cost for consumers to optimize their financial products.

In the past, switching insurance annually, comparing rates across five banks, or researching which credit card offers the highest cash back required consumers to invest time and effort; if these tasks can be continuously handled by an Agent in the background, then the economic value of "users being too lazy to switch" itself could decline.

[Image suggestion: Place Exhibit 1 from Page 1 of the report here — stock price performance across different financial sectors following Muse's launch.]

But just because an Agent can do it doesn't mean it has the right to

If one continues to extrapolate along this logic, it is easy to arrive at an extreme conclusion: AI Agents will ultimately bypass banks, insurers, and brokerages to directly complete all financial decisions on behalf of consumers.

Bernstein believes the reality is not so simple.

Third-party Agents face a fundamental contradiction: without cooperation from merchants and financial institutions, they can hardly complete complex transactions; but if opening access means losing customer relationships, transaction entry points, or a portion of revenue, financial institutions have no reason to cooperate unconditionally.

Banks, brokerages, and insurers still hold several key control points: account login, identity verification, data permissions, formal quotes, and whether a user qualifies for a particular product.

This is also why some platforms have already begun restricting Agent access.

Bernstein notes that Amazon chose to block Muse, with disputes between the two sides centering on Agent identity recognition and the use of user login credentials; insurance comparison platform Insurify also restricted Muse from scraping quotes, on the grounds that insurance is not simply a price comparison — beyond premiums, there is a wealth of information including coverage limits, deductibles, discount conditions, eligibility, and regulatory disclosures. If an Agent ultimately presents only a "lowest price," what consumers get may not be a truly comparable product.

This means the core bottleneck for financial Agents is shifting from "whether the model can do it" to "who will allow it to."

Data is one of the most important barriers. Financial institutions control account authentication, account data, and data-sharing permissions, so Bernstein raised a possibility opposite to "AI platforms charging banks": could banks, in turn, charge Agents data access fees in the future?

The report noted that earlier U.S. open banking rules around Section 1033 originally sought to require banks to provide account data free of charge to consumers and their authorized third parties through secure APIs; the relevant rules were subsequently blocked, and JPMorgan has since begun charging data aggregators for customer data access. Bernstein believes that in the future, around Agent data access, there will likely be more blocking, paid agreements, and situations in which financial institutions actively control which information is shown to Agents.

Even if the technology and data interfaces are already in place, consumers themselves are another constraint.

A 2026 TD Bank survey of more than 2,500 U.S. consumers showed that 55% had already used AI to help manage personal finances, but only 18% were willing to let AI make important financial decisions independently. Consumers are clearly more receptive to a model in which "AI provides advice and humans retain the final decision."

Other surveys show similar results. A survey by ACI Worldwide and YouGov showed that only 7% of U.S. and U.K. consumers were willing to let an AI assistant make purchases directly without approval; an Accenture survey showed that 32% were willing to let an Agent make purchasing decisions within set limits, but when it came to the actual payment stage, only 12% were willing to let the Agent decide completely autonomously.

[Image suggestion: place Exhibit 2 on page 3 of the report here—consumer acceptance of "AI-assisted" and "fully autonomous AI" across different scenarios, with financial investment decisions in the lowest tier of acceptance for autonomy.]

Therefore, the earliest large-scale form of financial Agents may not be full autonomy, but rather: AI completes search, comparison, screening, and execution preparation, while humans retain confirmation rights at key financial decisions and payment stages.

What is even more troublesome is that once Agents truly begin executing transactions, the question of liability will also arise.

Can a loan application submitted by an Agent represent valid authorization from the consumer? If an Agent uses outdated data and selects for the user a product that is technically eligible but not suitable for them, should the loss be borne by the model company, the financial institution, or the consumer? If an Agent begins actively comparing and recommending insurance or investment products, under what circumstances would that constitute regulated financial advice?

Bernstein categorizes these unresolved issues into several areas, including authorization, identity verification, financial advice, licensing, data access, and liability attribution.

Therefore, the biggest difference between financial services and ordinary e-commerce may lie in this: getting an Agent to order a meal for a user is easy, but having it allocate assets, apply for loans, or purchase insurance on a user's behalf requires solving an entire set of permission, trust, and liability mechanisms.

The Impact Won't Be Evenly Distributed: Visa and Mastercard May Actually Benefit

It is precisely for this reason that the impact of AI Agents on the financial industry will not be evenly distributed. If a company primarily profits from consumer inertia, product switching costs, or long-term usage habits, the emergence of Agents may erode some of its advantages.

But for infrastructure companies responsible for solving identity verification, payment security, and transaction liability issues, the logic may be completely reversed.

Visa and Mastercard are among Bernstein's most favored examples.

Intuitively, if AI Agents can make payments autonomously, card networks might seem likely to be bypassed. But Bernstein believes that Agentic Commerce actually constitutes a positive factor for Visa and Mastercard.

The reason is that when payments shift from "humans operating personally" to "machines operating on behalf of humans," the entire payment system needs to solve even more trust issues: Who is initiating the transaction? Who does it represent? Is it authorized? How much is it allowed to pay? And once fraud or disputes occur, how are they handled?

These all require more sophisticated identity verification, risk control, Tokenization, and dispute resolution mechanisms.

Tokenization here refers to using specially generated digital credentials to replace real bank card numbers for completing payments. In Agent scenarios, such Tokens can also carry more information about the transaction entity, usage context, and authorization scope.

Visa has launched Visa Intelligent Commerce, and Mastercard has also launched Mastercard Agent Pay. Bernstein believes that as Agent transactions increase, such Agentic Tokens and the identity and risk management capabilities behind them may become even more important.

This also explains why the impact on card-issuing banks and card networks may not be the same.

Issuers' traditional top-of-wallet advantage could be eroded as Agents re-compare rewards and interest rates every time; but the transaction networks, identity verification, risk control and dispute resolution provided by Visa and Mastercard may instead become more important as machine transactions increase.

The same logic applies to PSPs such as Adyen and Stripe, i.e., Payment Service Providers.

The more Agents there are, the more complex the platforms, protocols and payment interfaces merchants need to be compatible with. Bernstein believes that modern PSPs can help merchants uniformly handle different Agent channels, which may instead create new room for differentiation. Adyen has already launched Adyen Agentic, allowing enterprises to avoid rebuilding a complete commerce system for every AI platform.

PayPal's position is more subtle.

If Agents already handle payments for users, the value that digital wallets previously provided in reducing Guest Checkout friction may decline; but on the other hand, digital wallets can also expand their role into trust, fraud protection, dispute resolution, offer discovery and payment method optimization.

Therefore, Agents will not necessarily eliminate the payment middle layer. They are more likely to change the rationale for the existence of different middle layers.

What really matters next is: who holds the Agent's "execution authority"

Muse's rapid growth has already proven that consumer-grade Agents can gain a large number of users in a very short time. But for the financial industry, how smart the model itself is may actually not be the most important variable in the next stage.

What truly needs to be observed are three questions.

First, will consumers gradually move from "letting AI give advice" to "letting AI directly execute."

Second, how will banks, brokerages and insurance companies handle Agents' requests for data and account access—will they restrict, charge, cooperate, or launch their own Agents.

Third, and most critically, whether real operating metrics such as insurance renewal rates, bank deposit stickiness, brokerage cash sweep balances and credit card top-of-wallet are beginning to show sustained changes because of the spread of Agents.

Only when these metrics truly change will the impact of AI Agents on financial business models have moved from market expectations into the operational realm.

So the real question raised by Bernstein's report is not "Will AI disrupt the financial industry?" Rather, it is: once consumers have, for the first time, an agent that can work around the clock to find them lower prices, higher yields, and better terms, which profits in the financial industry are essentially built on consumers' past reluctance to act?

At the same time, which companies truly control identity, data, payments, and liability systems will also become increasingly important. This may be what truly needs to be repriced in finance in the Agent era.

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps