For the better part of two years, agentic AI in banking has been discussed mostly as an internal efficiency story, in which autonomous software helps employees underwrite loans faster, triage fraud alerts, or write code. That framing is about to feel dated. The more consequential shift is happening on the other side of the glass, where consumer-facing agents are starting to shop, compare, and move money with minimal human input.

Accenture’s Top Banking Trends 2026 report made the point bluntly. Customers increasingly trust AI to act on their behalf, and agentic systems can optimize deposits and loans with essentially zero friction, accelerating the shift of funds across institutions. Accenture pegs the incremental AI benefit for the top 200 global banks at roughly $289 billion, but warns that capturing it depends on who owns the decision layer sitting between the customer and the balance sheet.

Deloitte’s 2026 Banking and Capital Markets Outlook strikes a similar note, arguing that banks must move beyond department-level pilots to enterprise-wide agentic strategies because the competitive pressure on deposits is shifting from product pricing to who controls the agent making the choice. Only 11% of organizations currently have agentic AI in production, according to Deloitte’s 2026 Tech Trends report, which means the window to set the rules is closing fast.

Why Deposits Are the First Domino

Deposits look like the softest target for consumer agents, and the reasons are structural. Rate transparency has never been better, account opening has been digitized at almost every large institution, and rails like FedNow and RTP can move funds in seconds. Hand a well-instructed agent a customer’s goals and permissions, and there is very little left between that customer and a higher-yield account at another bank.

The exposure is not evenly distributed. Accenture’s analysis groups the at-risk pool into three buckets that should be familiar to any retail banking chief. The first is corporate working capital sitting in low-yield transaction accounts. The second is retail transaction balances clearing through debit cards and peer-to-peer apps. The third is cross-border settlement float stuck in nostro and vostro networks. Each is precisely the kind of balance an agent can recognize as inefficient and reroute on a schedule.

Deloitte’s outlook adds a sobering overlay, noting that even relatively small disruptions to deposit and loan rates could put a meaningful share of pre-tax income at risk for U.S. banks. When the agent doing the optimizing never gets tired, never forgets to move a balance after a rate change, and never feels guilty about closing an account it opened six months ago, that risk compounds quickly.

What “Agent-Ready” Actually Means for a Bank

Being agent-ready is not the same as having a chatbot on the homepage, and executives who conflate the two will spend a lot of money on the wrong problem. The real work sits one layer deeper, at the intersection of identity, APIs, and controls.

Identity is the hardest piece. An agent acting for a customer needs a verifiable, revocable credential that a bank can trust, and banks need the ability to distinguish a legitimate customer-authorized agent from a scripted attack or a coerced transfer. Capgemini’s World Cloud Report for Financial Services 2026 found that nearly half of banks and insurers are creating roles specifically to supervise AI agents, and Accenture’s research indicates most CIOs expect agent activity to run under a central governance model with real-time telemetry. That is the baseline, not the ceiling.

APIs are the next layer. If a bank’s rate sheet, account-opening flow, and funding rails are only reachable through a branded mobile app, the bank effectively hides from the agent economy. The institutions quietly winning here are building structured endpoints for product terms, eligibility, and funding instructions with the same rigor they once applied to open-banking mandates. Without that plumbing, a bank can still be chosen by an agent, but only if a third-party aggregator decides to list it.

Controls close the loop. Agentic transactions look a lot like authorized push payments from a fraud perspective, which means the same real-time signals banks have been building for scam defense will need to extend to legitimate agent-driven movement. The goal is to let good agents move funds frictionlessly while still catching the coerced or spoofed ones, and that is as much a policy exercise as a technology one.

Dynamic Pricing Is the Defensive Half of the Playbook

Building for the agent economy is only half the response. The other half is making sure customers have a reason to stay, because an agent that scans the market every morning will find a better rate somewhere eventually. That shifts deposit pricing from an annual exercise into a continuous one.

Advanced institutions are already moving in this direction. Consulting firm Simon-Kucher has described banks using AI and behavioral analytics to understand how money moves between accounts and what patterns precede attrition, so they can intervene before balances walk out the door. Rather than reacting after outflows, these banks predict which customers are likely to move funds and proactively adjust rates or offers, which is exactly the muscle needed when a customer’s agent starts shopping at 2 a.m.

There is a catch, and it deserves board-level attention. Hyper-personalized pricing creates fair-lending-style questions on the deposit side of the house, particularly if models end up offering materially different rates to otherwise similar customers. The institutions that handle this well are treating dynamic pricing as a governed product, with documented inputs, explainable outputs, and customer-disclosure practices that can survive a regulator’s read.

Primary Relationships Are Still the Best Defense

All of this reinforces a lesson McKinsey has been making for a while, which is that customers with primary relationships keep most of their deposits in place and stay open to additional products and services. That pattern holds up even as channels shift, because a primary relationship is not just a checking balance, but a bundle of direct deposit, bill pay, card usage, and advice that an agent cannot easily replicate on day one.

For community banks and credit unions with smaller technology budgets, this is actually good news. The institutions most at risk from agent-driven deposit flight are the ones holding large, low-engagement balances on thin relationships. A small institution with genuinely deep relationships, local-business deposits, and relevant advice has defenses that a direct bank with no human touchpoints simply does not. That is why firms like Cornerstone Advisors have argued that the real question for most banks is not whether to chase every new technology, but whether the customer base would have any reason to consider leaving in the first place.

What Executives Should Do Before the End of the Year

The tempo of the response matters more than the sophistication. Given how quickly consumer agents are likely to scale, a handful of moves through the remainder of 2026 will separate banks that set the terms from banks that spend next year reacting to them.

First, run a deposit-at-risk analysis that explicitly assumes a motivated agent acting for the customer. That analysis should segment balances by rate sensitivity, stickiness, and channel, and it should produce a specific dollar figure the executive team owns. Without that number, the rest of the agenda is abstract.

Second, publish structured product and rate information in a machine-readable form, even before any formal agent-interoperability standard arrives. Doing so puts the institution on the menu when agents go shopping and, just as importantly, builds the internal muscle to keep those feeds accurate.

Third, stand up an agent-governance function that spans risk, technology, compliance, and retail. The Capgemini research suggesting banks are already hiring explicitly for AI-agent oversight is a useful benchmark, and it is a clear signal that this cannot sit inside a single department.

Fourth, move deposit pricing and retention into a near-real-time operating model, with clear guardrails on disclosure and fair treatment. The goal is not to win every rate comparison, but to ensure the bank gets a fair hearing before a customer’s agent pulls the trigger.

None of this requires waiting for regulators to finish writing agent-specific rules. Institutions that act while the rules are still forming tend to be the ones whose practices become the template, and in a market where an AI agent can move a deposit in the time it takes to read this sentence, waiting is the most expensive option on the table.