Customer service is one of the first areas where banks, insurers, and healthcare organizations have deployed AI directly in front of customers, according to the U.S. Government Accountability Office.

In financial services, all ten of the country’s largest commercial banks now use chatbots to engage customers, and more than 98 million U.S. consumers interacted with a bank chatbot in 2022, according to the Consumer Financial Protection Bureau.

The CFPB has warned that poorly designed chatbots can provide incorrect information, fail to recognize when consumers are exercising federal rights, and leave customers unable to reach a human representative.

Healthcare shows a similar split between adoption and readiness. 71% of U.S. hospitals now use predictive AI, and the share applying it to scheduling rose from 51% to 67% in a single year, according to the Office of the National Coordinator for Health IT. Yet health system leaders cite immature AI tools as their top barrier to adoption at 77%, with regulatory uncertainty close behind at 40%, according to a national survey published in the Journal of the American Medical Informatics Association.

Regulatory oversight itself is struggling to keep pace with that adoption curve. The GAO has concluded that the federal agency responsible for supervising credit unions lacks some of the tools it needs to oversee how they use AI — leaving a gap between how fast the technology is deployed and how closely it’s being governed. ​

Emerj’s Yolandi de Weerdt recently hosted a conversation with Shri Nandan, VP of AI Products and Experiences at Comcast, to examine how AI scales in regulated industries by grounding CX in governance, clean data, and clear human–AI boundaries. ​

This article examines three core insights from that conversation that matter most for CX, digital, and AI leaders in banking, insurance, and healthcare:​

• Bounded AI scope to secure high-stakes interactions: Define what AI can resolve autonomously and where human escalation is required before deploying agents into clinically, financially, or legally sensitive customer journeys.
• Unified customer data for reliable AI context: Establish enterprise ownership, freshness standards, and a shared customer record before expecting AI to deliver consistent personalization across business units.
• Centralized AI governance to ensure operating scale: Pair governance and data strategy with controlled experimentation and clear decision authority so successful use cases can scale without multiplying organizational risk.

Listen to the full episode below:​

Episode: Precision CX in Regulated Industries – with Shri Nandan of Comcast​

Guest: Shri Nandan, VP of AI Products and Experiences at Comcast​

Expertise: Artificial Intelligence, Customer Experience, Product Strategy, Digital Products​

Brief Recognition: Shri Nandan is a technology and product executive with more than 20 years of experience leading digital and AI initiatives across telecommunications, healthcare, financial services, and insurance, including prior roles at Momentum Financial Services Group, Main Line Health, and MetLife. She holds a master’s degree in computer science from Mississippi State University.

Bounded AI Scope to Secure High-Stakes Interactions

Much of the optimism around AI in customer service assumes a generic enterprise setting. Nandan’s starting point is that BFSI and healthcare differ in kind, beginning with the emotional register of the conversation. A contact center agent helping someone buy an insurance policy is handling a transaction; an agent working out why a patient needs an appointment may be handling something far more sensitive, and the design of any AI system has to reflect that difference before a line of code is written.​

She argued that the decisive step is an honest, explicit conversation about what the organization wants AI to do for the customer, and where it should stop:​

“When you’re designing your agentic system, there has to be an honest discussion about what it is that you want your AI to do to help the customer. Is it just scheduling and rescheduling appointments, or is it something fairly simple, like looking at your lab work results? If it’s a little bit more complicated, especially in things like oncology or something more serious than that, how would you expect AI to help the customer? I think it’s important for the organization to understand that there isn’t a lot that AI can do in certain situations, and you need human intervention.”​

— Shri Nandan, VP of AI Products and Experiences at Comcast

​Financial services raises a parallel problem on the risk side. Nandan described the appeal of an AI financial advisor and then the question that follows it: how does the institution know the advice is sound, and how does it know the agent has considered every option that could earn more for the customer? Building an agent that far-reaching, she said, is harder than it looks, and regulation compounds the difficulty. An agent built to comply with U.S. rules may not satisfy the rules in the UAE, a lesson she said came directly from her time at MetLife, where regulations varied from one country to the next.

Shri Nandan argues that the practical consequence for CX leaders is the need to set an AI agent’s scope deliberately and to treat the boundary between automated and human handling as part of the design rather than an afterthought:

Classify interaction weight  

Sort interactions by emotional and legal stakes before choosing use cases. Routine scheduling and information retrieval sit at one end. Oncology conversations, financial advice, and fraud disputes sit at the other. Nandan’s test is whether the organization can state plainly what AI should do for the customer at that moment.

Treat human escalation as a feature  

Define the handoff upfront. AI handles bounded, repeatable interactions. Humans handle situations with higher emotional, clinical, legal, or financial complexity. The escalation point is part of the product, not a fallback.

Design for jurisdiction early  

Regulatory regimes differ, so a single agent design will not carry across markets unchanged. Governance has to understand each jurisdiction’s rules before the agent is built.

Nandan tied all of this back to trust as a foundational capability. In her framing, any decisioning system in a regulated environment has to give the customer assurance first, before it attempts to influence behavior, offer options, or build loyalty. Nandan was direct when asked about decisioning logic in high‑stakes settings:​

“Any technology has to build trust. It has to be able to say: you’re dealing with a bot, you’re dealing with AI, you’re dealing with technology, but you’re safe. Your information is safe, and you are in good hands. I think it’s important to build that trust as a foundational capability.”​

— Shri Nandan, VP of AI Products and Experiences at Comcast

Unified Customer Data for Reliable AI Context

Asked whether the real blocker in these sectors is data, regulation, or organizational readiness, Nandan said it is all three, but focused on data and on the organizational behavior behind it. In a large institution, customer data is typically owned by many different parts of the company. The first problem is building a single source of truth from those disparate holdings; only then can a team think about making that data AI-ready.​

Even with unified, AI-ready data, leaders need to consider where the processing that turns that data into customer context takes place. Nandan described this as the “gravity” of the computation. If processing sits too far from the customer interaction, it can introduce latency and performance issues, and a poorly designed data architecture also increases operating costs as AI usage scales. For enterprise leaders, this makes architecture an early scaling decision rather than a technical consideration to address after deployment.​

Nandan was unequivocal that data is “one hundred percent important” and the key to a good customer experience, but she located the hard part somewhere other than the technology:​

“The problem with creating good data, creating data with integrity, and creating single sources of truth is more cultural than anything else. If you have five different sets of disparate teams owning certain aspects of the data, it’s very difficult for you to say you need to give up that data. So there’s a bit of organizational change that needs to come into place that allows the data team to say: this is the data we have, this is how we all come together, this is how we create a source of truth, this is how we keep it fresh, and this is how we can use it in our decisioning systems and to create context.”​

— Shri Nandan, VP of AI Products and Experiences at Comcast

​The executive implication is that AI-ready customer data is an organizational ownership problem before it is a technical integration problem, and Nandan argued that the change has to come from the top. Without that alignment, and the data architecture and strategy to support it, institutions end up building AI on fragmented customer context, producing inconsistent experiences and additional friction for the customer.​

Once the data foundation exists, Nandan’s guidance on which AI capabilities matter is simple: any capability that solves a customer problem quickly. What the unified data adds is the ability to hyper-personalize. With fresh, integrated context about a customer’s history and journey, an agent can render experiences almost in real time, recognizing, for example, that a customer has asked for help with the same problem repeatedly without resolution, and routing them down a different path as a result. Evaluating the agent’s performance then feeds the product roadmap: features that do not work are dropped, features that do are extended, and the experience improves iteratively.​

For leaders trying to move from fragmented data to a usable customer context, Nandan’s account suggests a sequence:​

• Treat data consolidation as an executive mandate: Leadership needs to establish customer data as an enterprise resource, with enterprise-wide rules for access, stewardship, and accountability, rather than leaving it under the control of whichever business unit happens to hold it.
• Define freshness and ownership alongside the single source of truth: A consolidated record that is stale, or that no one is accountable for keeping current, does not produce reliable context for models or agents.
• Decide where computation sits before scaling: Positioning the processing that builds customer context close enough to the customer to avoid latency, without duplicating heavy infrastructure, is an architectural choice with long-term cost consequences.
• Use agent evaluation as the roadmap input: Rather than planning AI features in the abstract, build the agent, measure how it performs against real customer problems, and let the results determine what gets built next.

Nandan also noted that the governance picture has fragmented in the same way the data has. What was a single umbrella of digital governance fifteen years ago has become data governance, AI governance, and context governance, each of which now needs its own guardrails in a regulated business.

Centralized AI Governance to Ensure Faster Operating Scale

On the practical steps toward an operating model where AI improves service quality at scale, Nandan laid out a clear order of operations. The first requirement is an AI governance practice; without guardrails, everyone goes off in different directions and the result is chaos. Running alongside that is a sound data strategy. Only once both exist should an organization turn to experimentation and innovation. ​

Nandan advocated reserving explicit capacity for experimentation, where teams can test new technologies and build proofs of concept (POCs) without committing to enterprise-scale deployment. Governance and data strategy then provide the criteria for deciding which experiments warrant further investment and which should stop. An innovation lab of that kind, in her view, is how organizations keep pace with a technology landscape she described as changing at “lightning speed.”​

She was equally clear that none of this eliminates failure. Things will go wrong, and the differentiator is how quickly leaders adapt. Returning to a theme from earlier in the conversation about red flags, Nandan offered a rule for when to stop:​

“If you see that none of the KPIs are moving, have the courage to stop and say, what do I need to change? We don’t need to keep pushing at something if things are not moving in the right direction. So a lot of it is strong leadership, along with all of the governance and the basic foundational capabilities in place.”​

— Shri Nandan, VP of AI Products and Experiences at Comcast

​The most pointed part of the discussion concerned what the organizations reaching real operating scale in BFSI or healthcare do differently. Nandan cautioned that organizations are still in the beginning stages of building enterprise AI operating models and described this phase as nascent, with even leading institutions still experimenting and encountering failures. The pattern she sees among those progressing faster is more concentrated AI decision‑making, and she explained why spreading it too broadly across the enterprise backfires:

​Nandan’s view is that concentrating responsibility for AI within one part of the organization lets enterprises move faster by reducing competing decision paths and organizational politics. That position also aligns with the governance-first sequence she laid out earlier: guardrails come from a governing body that understands the industry, relevant laws, and internal policies, while leadership retains clear authority to act within those boundaries.

​The operating model Nandan describes can be summarized as a set of standing commitments:

​“When you say, I want democratization of AI, and I want everybody to work on it, and I want everybody to have opinions and everybody to make decisions, then I think that creates a lot of chaos and uncertainty, and then people end up having to report to five different managers. But when organizations create a tight AI unit and have strong leadership making very courageous and quick decisions without worrying about the optics, I think that is really helpful in moving things forward, so that we’re not wasting time on politics.”

​— Shri Nandan, VP of AI Products and Experiences at Comcast

• ​Stand up AI governance and data strategy before funding use cases: These are the guardrails within which every subsequent experiment is judged, and building them after the fact means retrofitting controls onto systems already in production.
• Reserve explicit capacity for experimentation, and gate scaling on governance: A POC earns the right to scale by fitting the data strategy and governance framework, not by enthusiasm alone.
• Define success thresholds before scaling: Leadership should agree in advance which customer and business outcomes determine whether an AI initiative advances, changes direction, or stops, so that halting one is a matter of discipline rather than a prolonged negotiation.
• Concentrate AI decision authority during early scaling: A tightly accountable AI unit with clear leadership and the authority to make rapid decisions removes the political overhead that broadly distributed decision-making creates while governance practices and operating models are still maturing.