Deloitte: Bank AI applications should be task-oriented, not customer-tier-based

Lauren Littlefield, a managing director at Deloitte Consulting, noted that customer service centers have become one of the most suitable areas within banks for AI-driven transformation. She also emphasized that such deployments must be carried out prudently, with the goal of enhancing customer experience rather than simply pursuing cost reduction—the latter often being the starting point of conversations between banks and consultants.

Littlefield is one of the authors of a related report released by Deloitte late last month. In a recent interview, she said that without this mindset, banks risk customer dissatisfaction and damaged relationships. The report, which focuses on AI adoption considerations, surveyed bank executives and customers between November 2025 and January 2026.

On the positive side, AI can provide human agents with "unprecedented capabilities," enabling them to engage in personalized conversations with customers using tools such as real-time prompts and contextual information. Littlefield noted that this capability is exactly what institutions like Bank of America are promoting among their customer service staff.

Editor's note: This interview has been edited for clarity and brevity.

Bank Dive asks: What stood out to you about the survey findings on bank customer service centers?

Lauren Littlefield:Even before discussing AI, there was already a significant gap between how bank executives perceived customer experience and what customers actually experienced. My executive clients often say: "We're doing well; the focus should be on reducing costs." The survey gave me a new framework. When executives begin considering AI deployment—almost always first for efficiency gains—I can remind them: "Let's make sure we don't widen the existing experience gap." AI implementation can act as an amplifier: if deployed poorly, it can worsen already existing problems.

What customer service experience issues could lead to broken customer relationships?

For many bank customers, a fundamental pain point is the sense of fragmentation across channels. Because AI relies heavily on core data flowing between systems, in multi-channel interaction scenarios, using different models for chat, voice response, and human agent assistance can actually exacerbate this fragmentation. AI models draw on knowledge bases in ways similar to humans. For example, if the chat AI and voice AI are trained on different knowledge bases, customers may receive contradictory information across channels. Banks have historically had hygiene issues in knowledge management. If two models reference different knowledge bases, banks face enormous risks—not only providing inappropriate information, but more seriously, severely damaging customer trust. In some deployment tests, conflicts between knowledge bases were not uncommon.

What are common misconceptions about AI applications in bank customer service?

A common misconception involves how to strategically allocate tasks between AI and human agents. There are two extremes, and the right answer lies in between. For example, one large financial institution client believed that its high-net-worth customers should not interact with AI and should always receive 100% human service. But based on the data I've seen, even high-net-worth customers, in the simplest use cases, often prefer fast, efficient automated channels over waiting in line for a human. The key is understanding where AI can add value, rather than simply asserting that "certain customer types should only receive human service."

So, should AI adoption be driven by customer needs rather than customer tier?

Correct. If a customer calls about a mortgage or commercial loan, human involvement is likely needed; but if a customer calls to report a lost debit card and request a replacement, handling it manually is more time-consuming and less efficient. Because many banks tier their agents, with the most expensive (often nearshore) agents assigned to serve high-value customers, this creates a double inefficiency: it makes high-value customers wait for matters that don't require human handling, while also using the most expensive human resources. Of course, customer tiering will still exist, but the difference in who interacts with AI should depend on intent, task, and use case. All customers have basic human needs—to be heard, reassured, and guided during high-pressure moments. So if banks want to retain customers, especially lower-value ones, they must provide equally good experiences, or a single poor service encounter could lose them.

Which banks have stumbled in terms of AI-human agent collaboration?

This is a huge change management challenge: getting employees accustomed to working with eight screens to instead face a single interface where a model provides real-time prompts, context, or next-best-action suggestions. I've observed adoption issues, with employee acceptance falling short of bank expectations. One client didn't roll out a fully featured new desktop experience to agents. Instead, they took an incremental approach: starting with one feature, beginning with summarization, sparking interest, then gradually adding other features, and introducing gamification and rewards. Like all of us, customer service agents don't want to wake up the next day and completely change how they work.

Do you think the "human in the loop" will disappear soon?

Banks' attitudes toward removing people from processes are shifting, with more generative AI capabilities with appropriate guardrails being deployed directly to customers. But will human agents disappear within the next five years? Absolutely not. However, the agent's job profile will fundamentally change—they will focus more on empathy, problem-solving, and advice in high-pressure situations, rather than performing mechanical, routine tasks. As AI capabilities grow, when customers finally reach a human, the expectation will be that this agent has full knowledge of the customer and can provide personalized decisions and recommendations.