Skip to content
All insights

Artificial Intelligence

Redesign the Process Before You Add the Agent

By Alex Jimenez

Function managers in banks and credit unions are used to hearing from their executive team that they need to improve efficiency. CFOs look at operating ratios closely to ensure that they are keeping up with the industry, and most importantly, with their competition. With the advent of LLMs and Agentic AI, the calls for efficiency are now being paired with requests to roll out AI agents.

Banks and credit unions can find efficiency gains in many functions. In marketing, a campaign waits four to six weeks at best for compliance review, according to a Cornerstone Advisors report. Roy Zur, CEO of Charm Security, which sells AI agents to fraud teams, told a Money20/20 session in October that fraud losses in North America grow about 20% a year, faster than teams can hire. Executives expect agents to do the same work with far fewer people.

Executives may not get what they expect from agents in old processes

Ankura, an advisory firm, explains where current processes come from. Institutions built workflows around human limits: batches because staff worked in shifts, queues because attention was scarce, and approval layers because supervision cost a lot. The queue is still there in customer data. A fall 2026 Digital Banking Report survey of banking and credit union executives, sponsored by NGDATA, found that at roughly six in ten institutions, retail and lending leaders cannot build a list of their own customers without help from IT or an analyst. Cornerstone describes a $3 billion institution with one compliance resource covering marketing reviews on top of other duties. An agent that produces campaigns faster adds to that person’s queue.

In other functions the pattern repeats. In campaign review, Cornerstone says an agent alone does not fix marketing and compliance working against each other. Drafts and redlines still go back and forth, only a little faster. In online account opening, applicants abandon 3.36 checking applications for every one they complete, according to Cornerstone’s 2026 benchmark. Cornerstone traces much of that to identity checks and to applicants who switch devices and must start over, in a process designed for compliance first and conversion second. The data comes from 89 institutions that reported their own numbers.

The same survey asked executives what most often stops customer insight from turning into action, and they could pick up to three answers. Ownership and data tied at the top: 46% said no one clearly owns acting on it, and 46% said data is trapped in the core or in separate systems. Budget drew 16% and compliance review 14%.

Theo Lau, founder of Unconventional Ventures, warned on our podcast, Next Best Action, about institutions that buy AI from several vendors and add it to broken processes: “Perhaps you’re just doing the wrong thing faster.”

“You are automating things that are broken. You are automating mistakes. You’re automating the wrong things.” Theo Lau

Managers can ask what each step is for

Managers can look at each step before they add an agent. Masha Shunko of the University of Washington and Serguei Netessine of Wharton wrote in Harvard Business Review in September that managers should ask what each step was built to work around and whether that reason still holds. In account opening, that means asking what risk each screen and each upload covers, and whether every applicant presents it.

Frida Lebowitz, CEO of Debbie, a vendor that works with credit unions, described two credit unions with nearly identical technology on the Banking Transformed podcast, as Jim Marous of The Financial Brand reported in August. According to Lebowitz, one completed 5% of applications and the other more than 60%. She estimates that roughly 60% of applicants leave at the identity document upload, and the better credit union asked for a document only when a risk signal called for one.

In campaign review, Cornerstone reports that institutions with 90% shorter legal review brought compliance into the writing of campaigns and stopped treating it as a final approval. Persado, which commissioned the report, claims reductions of up to 90% for its platform.

Not every step needs an agent

Once managers know what each step is for, they can ask whether an agent is the right tool for it. Institutions can use three kinds of AI. The first, deterministic AI, follows fixed rules such as “send any application with a mismatched address to review” and gives the same answer every time. People use the second, generative AI, to draft and summarize. The third, an agent, takes a task such as clearing a campaign through compliance review and decides the steps on its own.

The data a step needs varies. The Digital Banking Report notes that a welcome offer for a new checking customer needs only a small set of well-understood data. Predicting which households will leave needs matched identities, household links, at least a year of history and categorized transactions.

Shunko and Netessine describe four ways to split work between AI and people. With Assist, the AI suggests and a person acts. With Approve, the AI prepares an action, and a person approves it before it runs. With Audit, the AI acts and people review samples afterward. With Automate, the AI acts within rules people set, and people step in only for exceptions. They say to keep more human control when stakes are high or an action is hard to undo, and to give agents more freedom when errors are visible and easy to fix. Lau advises institutions to decide early which decisions AI handles and which people keep.

Account opening has three points where an institution can apply these choices. Before the application, a prospect discovers products and picks one, sometimes with an AI assistant’s help. During the application, some applicants leave. After it, the institution onboards the new account holder.

We work at each point. We personalize what a prospect sees while choosing a product, and we offer an MCP server, a standard way for AI assistants to ask an institution for its products, rates and locations. A person at the institution approves everything before it is published. We bring back applicants who leave. At one of our clients, 1st United Credit Union, completed applications per session were 0.38% with personalization on and 0.29% in a held-out control group, a 29% lift for the whole program. We can build personalization around onboarding.

Starting over may work better than improving step by step

Improving one step at a time keeps the process that already exists. Michael Hammer argued in Harvard Business Review in 1990 that companies with outdated processes should “obliterate them and start over.” Shunko and Netessine wrote in September that his warning is “as urgent as ever,” and they cite McKinsey QuantumBlack’s finding that 60% of organizations still see no enterprise-wide earnings impact from AI.

Accenture, a consulting firm, says that leaving agents to continuous improvement teams, without a mandate to reinvent the work, risks missing their full value.

Startups built around agents from the beginning now enter markets with a fraction of the time and capital that established companies need, according to Shunko and Netessine. Zur says institutions need to redesign the whole organization around teams of people and agents. Organizations that improve one step at a time can stay with a slightly better version of the process they started with. A manager who starts from a clean slate can ask what the process should look like if agents did the work from the start.

Gartner expects many agent projects to be cancelled by end of 2027

Gartner made its forecast in June 2025. Halfway to the deadline, adoption has picked up. In its 2026 Hype Cycle for Agentic AI, Gartner reported that 17% of organizations had deployed AI agents and that more than 60% expected to within two years. It placed agentic AI at the Peak of Inflated Expectations, its term for the point where expectations run ahead of results. By August, 37% of respondents in Gartner’s CIO planning survey said they had deployed agents, and another 34% planned to within 12 months.

Banks and credit unions are earlier on that curve. In a Cornerstone survey, 9% of marketing leaders said their institutions use agents that can execute actions. Another 63% said their institutions would be very or somewhat comfortable letting agents act on their own within approved limits in two years. A fall 2026 Digital Banking Report survey shows the same early stage: 55% of institutions use no models or scores today.

The forecast stands: more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Executives will keep asking for agents, and managers will say yes more often. Managers can answer all three of Gartner’s reasons before an agent goes live: what it will cost, what it is worth and who will watch it. They also decide whether the agent runs the old process or a better one.

Insights

More from the Finalytics team

  • By Alex Jimenez

    Digital is Where the Relationship Lives Now

    A member walks into her local branch to sort out a mixed-up direct deposit. The teller pulls up her account, fixes it in four minutes, and asks about her daughter’s college applications. She leaves…

  • By Alex Jimenez

    They Leave Well Before They Close the Account

    Most retention reports tell a reassuring story. Members stay, deposits hold, accounts stay open, and churn comes in low again this quarter. Underneath those steady numbers, the relationships they…

All insights