How to Use AI to Improve Customer Communications: A Practical Guide for Financial Services

Key Takeaways

• The ROI gap is real, but closeable. Only one in five AI initiatives are delivering measurable ROI, according to Gartner. The differentiator is not budget or access to technology, it is deploying AI that is purpose-built for your domain and connected to real business outcomes.
• Agentic AI is moving from concept to production. The industry has evolved from scripted chatbots to large language model assistants and is now entering the era of true agentic AI: multi-agent systems that initiate conversations, negotiate outcomes, and fully close the loop with customers without human intervention.
• The business case for conversational agents is compelling. AI-powered voice and text agents are already demonstrating a 20–30% improvement in containment rates over rules-based systems, and a 90–95% reduction in operational cost per contained call. One in ten calls is estimated to be fully automatable by voice agents by end of year, and that figure may be conservative. Conversational agents are also unlocking complex use cases that traditional virtual agents couldn’t contain.
• Generic models are not fit for financial services. A model trained on internet data is not equipped for payment negotiation, hardship identification, or fraud verification. Domain-specific, focused language models outperform general-purpose alternatives on specialized tasks, are massively more compute-efficient, and keep customer data sovereign and auditable.

There’s no shortage of headlines about artificial intelligence reshaping the business world. Boards are demanding AI strategies. CEOs are name-dropping it in every earnings call. Analysts are calling it the most transformative technology shift of our lifetime. And yet, according to Gartner, only one in five AI initiatives are actually delivering measurable ROI. An MIT study puts the failure rate even higher, estimating that 95% of AI projects fail to return any meaningful value.

This raises the stakes for businesses looking to use AI to transform customer communications and the customer experience.

For financial services organizations in particular, the cost to serve customers is a constant battleground. Margins are under pressure, and every institution is looking for ways to do more with less, but not at the expense of compliance or customer experience. Ungoverned AI in financial services is not a cost saver. It is a liability.

At the same time, competition is fragmenting the value chain. Fintechs, neobanks, and embedded finance players are all competing for pieces of your customer relationship. The moat is no longer the speed of deployment. The moat is the expertise embedded in the model you are using. And customer loyalty, once lost, is hard to win back.

The decision window with customers is also shrinking. Fraud happens in milliseconds. A customer in financial hardship decides within the first few seconds of a call whether they are going to engage or hang up. That means AI in customer communications has to operate at speed, within governed infrastructures, and without creating risk.

A New Paradigm: From Reactive to Agentic

To understand where AI in customer communications is headed, it helps to trace the journey from where it started.

The first generation of automated customer interaction was the chatbot: rules-based, scripted, and effective only until a customer said something outside the decision tree. Anyone who has navigated a “press one for billing” phone menu knows the frustration. The moment a customer says something like “I want to pay, but I need to talk about my situation first,” the whole process breaks down and the call gets transferred to a human agent. That transfer is essentially the system admitting defeat.

The second generation — where much of the industry sits today — is large language model-powered assistants. These understand natural language and can hold a real conversation. But most of what is marketed as an “AI agent” today is, in reality, still a reactive system. It waits for a human to ask it something. It does not go and do things on its own.

The third generation — and this is where leading organizations are now investing — is true agentic AI. Multi-agent systems can initiate conversations, execute multi-step interactions across multiple channels, understand tone, detect financial hardship, negotiate payment terms, and fully close the loop with a customer. They do not just answer questions. They achieve outcomes.

 This shift is fundamental, and it is happening now. Here are three approaches that represent best practice in deploying AI for customer communications.

Approach 1: Conversational Agents

Conversational agents are the front line of AI-powered customer interaction. Built on multi-agent architectures, these systems can handle voice and text interactions end-to-end, from authenticating a customer to guiding them through a collections process or transaction verification, without a human agent ever being involved.

The business case is compelling. Research and real-world deployments are showing a 20 to 30% increase in containment rates compared to traditional rules-based systems. That means fewer escalations and fewer calls consuming your human agents’ time for tasks that AI can handle. Industry estimates suggest that one in ten calls could be fully automated using voice agents by the end of this year, and many practitioners believe that figure is conservative. The operational cost reduction per contained call is estimated at 90 to 95%. 

But the value of conversational agents goes beyond cost reduction. Every touchpoint with a customer is an opportunity to capture information about their financial situation, their preferences, their sentiment toward your institution. A well-designed conversational agent does not just help the customer self-resolve in the moment. It builds a richer customer profile that can be used to personalize future communications and improve decision-making for everything from cross-selling to fraud detection.

Consider how this works in a collections context:

• A customer who is 21 days past due receives an outbound call. 
• The agent authenticates them, explains the situation, and offers a range of payment plan options. 
• The customer selects a plan that works for their circumstances, say, 12 monthly installments.
• The agent records the agreement, sends a confirmation email, and closes the loop. 

The entire interaction is governed, auditable, and compliant. And the information captured — that this customer prefers installment payments, that they are experiencing short-term financial difficulty — becomes a valuable input for future decisioning. Since the customer provided their preference, you can use that in future interactions to provide a customized plan up front.

The same principle applies in fraud prevention. Rather than sending a one-way informational alert about a suspicious transaction, a conversational agent can actually talk to the customer. It can ask: do you know who you are transferring this money to? When did you last speak with them? By getting that information, the agent both helps the customer self-resolve and captures data that can inform future fraud decisions for related accounts.

For organizations implementing conversational agents, particularly in voice, the architecture matters enormously. A production-grade voice agent requires a reliable speech-to-text service, a large language model with enough contextual intelligence to handle the inevitable imperfections in speech recognition, integration with offer and account data, and a text-to-speech layer that sounds natural enough to keep customers engaged. Vendors need to understand the full end-to-end telephony and voice path, not just the AI component. And crucially, the system must be safe, compliant, and designed for the speed that voice conversations demand. The goal is sub-second latency, so that the interaction feels as natural as talking to a human.

FICO® Platform Communications Capability delivers conversational agents across both voice and text channels, with out-of-the-box agents for collections, fraud, and transaction verification. It also has a conversational agent builder for organizations that need custom workflows tailored to their specific processes.

Approach 2: Focused Language Models

If conversational agents are the front line, the language model powering them is the foundation. And this is where many AI deployments go wrong.

General-purpose models (the large language models behind tools like ChatGPT and Gemini) are trained on vast amounts of internet data. They are impressively capable across a huge range of tasks. But in specialized domains like financial services, “one size fits all” is actually “one size fits none.” A model trained on celebrity gossip, Reddit threads, and news articles is not equipped to handle the nuanced, high-stakes conversations that define financial services customer interactions.

The alternative is a focused language model trained not on the internet at large, but on decades of domain-specific data. In financial services, that means training on payment negotiations, hardship identification, fraud verification, right-party communication, and the full spectrum of collections and origination interactions. The difference between a general model and a focused model is not a minor technical distinction. It is the whole ballgame.

There are four meaningful advantages to focused language models in this context.

1. They outperform general models on domain-specific tasks. When the task involves understanding financial hardship or negotiating a payment, a model trained specifically on those scenarios will consistently outperform a model that has to generalize from unrelated training data.
2. They are dramatically more compute-efficient. A focused model is built from the ground up to do less work and get better answers, because it is not wading through billions of parameters of irrelevant information. The performance improvement can be massively more efficient. That efficiency directly enables the sub-second latency that voice interactions require.
3. Data sovereignty is preserved. With a general-purpose model hosted by a third party, you are effectively sending your customer conversations to someone else’s infrastructure and hoping for the best. A focused language model that operates within your own or your vendor’s controlled environment means that data is sovereign, auditable, and secure. In financial services, that is not a nice-to-have. It is a requirement.
4. Focused models are genuinely defensible as a competitive asset. They cannot be replicated by a competitor simply purchasing access to a public model. The training data, the domain expertise embedded in the architecture, and the years of investment in building the foundational model represent a durable competitive advantage.

Gartner has identified financial services and banking as one of the top three domains that most need focused language models. FICO has been building its foundational model for years, training it on financial conversations and decisioning data. The resulting FICO® Focused Language Models (FLMs) are purpose-built for financial services, not fine-tuned from a general base. That distinction matters.

Approach 3: AI-Powered Internal Assistants

Conversational agents and focused language models address the customer-facing side of communications. But there is an equally important application of AI on the operational side: empowering the people who configure, manage, and oversee these systems to work smarter and faster.

This is where AI-powered internal assistants come in. They represent a third dimension of best practice that is often overlooked in discussions about AI in customer communications.

An embedded AI assistant within a customer communications capability can democratize expertise across an organization. New team members can ask questions and get answers instantly, without having to track down the one person who has been in the role for years and “just knows” how everything works. Experienced practitioners can use it to accelerate their own workflows: analyzing complex solution logic, identifying unused or misconfigured elements, and generating plain-language summaries of intricate multi-channel strategies.

Consider the practical implications. A team member wants to understand a complex collection strategy with multiple subflows and communication channels. Rather than spending hours reading through configuration menus and documentation, they can simply ask the assistant to summarize the logic, identify areas for improvement, or flag anything that looks broken or inactive. The assistant, which is context-aware and understands what solution the user is working in, can analyze everything and provide a clear, actionable answer in seconds.

The same assistant can help with case analysis. Reviewing a long customer interaction log (a multi-turn voice conversation, for example) to extract the key facts is time-consuming and error-prone when done manually. An AI assistant can synthesize all of that into a concise summary, highlighting the resolution, the customer’s chosen payment arrangement, and any notable points in the conversation. That is not just a productivity gain; it is a quality and compliance benefit.

Perhaps most valuably, an AI assistant can be loaded with your organization’s own documentation, policies, and best practices. That means the knowledge base reflects your standards, your processes, and your institutional expertise. The assistant becomes a living repository of organizational knowledge, accessible to everyone.

FICO® Assistant, embedded within the FICO® Platform - Communications Capability, is designed to do exactly this. Users can interact with the product through a natural language interface, ask questions, optimize configurations, and gain insights from case reporting without manually reviewing logs.

Building the Value Loop

What makes AI in customer communications genuinely transformative is the feedback loop it creates between communication and decision-making.

Every customer interaction generates information. That information, when captured systematically and fed into a decisioning platform, improves the quality of future decisions. Better decisions lead to more relevant, more personalized communications. More personalized communications generate better customer responses and richer data. And so the cycle continues.

This is what separates AI-enabled customer communications from a simple automation exercise. It is not about replacing a human with a bot. It is about building a system that continuously learns, adapts, and improves — one that gets smarter about each customer with every interaction, and uses that intelligence to deliver better outcomes for both the customer and the institution.

The organizations that will win with AI in customer communications are those who understand this flywheel effect and invest accordingly. They will choose models built for their domain, not fine-tuned from generic ones. They will design systems that capture and use information, not just process transactions. And they will govern their AI rigorously, because in financial services, ungoverned AI is a major risk.

• How FICO Can Help You Improve Customer Engagement Read 3 Ways Conversational AI Aids Collections & Customer Support
• See how Customer Communications Services for Fraud engages customers in real time to disrupt a variety of fraud typologies
• Learn how real-time customer communication can be a competitive differentiator across the entire customer lifecycle