According to the U.S., one of the first sectors where banks, insurance companies, and healthcare organizations have used AI directly in front of clients is customer service. Office of Government Accountability.
According to the Consumer Financial Protection Bureau, over 98 million American consumers engaged with a bank chatbot in 2022. All 10 of the nation’s biggest commercial banks now utilize chatbots to interact with their clients.
The CFPB has cautioned that badly built chatbots may give inaccurate information, fail to identify instances in which customers are exercising their federal rights, and prevent users from contacting a human agent.
Adoption and readiness are similarly divided in the healthcare industry. According to the Office of the National Coordinator for Health IT, 71% of U.S. hospitals currently use predictive AI, and the percentage using it for scheduling increased from 51% to 67% in just one year. However, according to a nationwide survey published in the Journal of the American Medical Informatics Association, health system leaders cite immature AI tools as their top barrier to adoption at 77%, with regulatory uncertainty coming in second at 40%.
The regulatory oversight itself is having difficulty keeping up with the rate of adoption. The GAO has determined that the federal agency in charge of overseeing credit unions lacks some of the resources necessary to monitor their use of AI, resulting in a disconnect between the speed at which the technology is being used and the extent to which it is being regulated.
Yolandi de Weerdt recently sponsored a discussion with Shri Nandan, VP of AI Products and Experiences at Comcast, to explore how CX is grounded in governance, clean data, and distinct boundaries between humans and AI in order to analyze how AI scales in regulated industries.
This post looks at three key takeaways from that discussion that are most important for CX, digital, and AI professionals in the banking, insurance, and healthcare industries:
- Determine what AI can handle on its own and when human escalation is necessary before deploying agents into customer journeys that are therapeutically, financially, or legally sensitive to secure high-stakes engagements.
- Unified customer data for trustworthy AI context: Before expecting AI to provide uniform personalization across business divisions, establish organizational ownership, freshness requirements, and a shared customer record.
- Centralized AI governance to guarantee operational scale: To enable successful use cases to grow without increasing organizational risk, combine data strategy and governance with regulated experimentation and unambiguous decision-making authority.
Visitor: Comcast Vise President of AI Products and Experiences Shri Nandan
Proficiency in Artificial Intelligence, Digital Products, Customer Experience, and Product Strategy
Brief Recognition: Previously employed at Momentum Financial Services Group, Main Line Health, and MetLife, Shri Nandan is a technology and product executive with over 20 years of experience leading digital and AI initiatives across telecommunications, healthcare, financial services, and insurance. She graduated from Mississippi State University with a master’s degree in computer science.
Limiting the Use of AI to Protect High-Stakes Conversations
A lot of the excitement surrounding AI in customer service is predicated on a general business environment. Nandan begins by stating that there are differences between BFSI and healthcare, starting with the conversation’s emotional register. Before a line of code is generated, the architecture of any AI system must take into account the fact that a contact center representative assisting someone in purchasing an insurance policy is handling a transaction, whereas an agent determining why a patient needs an appointment may be dealing with something far more sensitive.
She maintained that having an open and honest discussion about what the company wants AI to do for the client and where it should end is the crucial first step.
“You need to be open about what you want your AI to do to assist the consumer when you’re creating your agentic system. Is it something very basic, like viewing the results of your lab work, or is it just making and changing appointments? How would you anticipate AI to assist the consumer if it’s a little more complex, particularly in cases like cancer or anything more serious? I believe the company must realize that in some circumstances, AI is limited and human assistance is necessary.
— Shri Nandan, Comcast Vise President of AI Products and Experiences
On the risk side, financial services provide a similar issue. After describing the allure of an AI financial advisor, Nandan raised the issue, “How does the institution know the advice is sound, and how does the agent have considered every option that could earn more for the customer?” According to her, creating such a comprehensive agent is harder than it first appears, and regulations make it worse. She claimed that her experience at MetLife, where regulations differed between nations, taught her that an agent designed to adhere to U.S. requirements might not meet UAE regulations.
According to Shri Nandan, the practical implication for CX executives is the necessity of carefully defining the scope of an AI agent and treating the distinction between automated and human handling as a component of the design rather than an afterthought:
Sort the interaction weight.
Prior to selecting use cases, arrange encounters according to the emotional and legal stakes. On one end are information retrieval and routine scheduling. On the other side are discussions on cancer, financial guidance, and fraud challenges. Nandan’s test is whether the company can clearly explain what AI should perform for the client at that particular time.
Consider human escalation to be a feature.
Clearly define the handoff. Bounded, repeated interactions are handled by AI. Higher emotional, clinical, legal, or financial complexity is handled by humans. The escalation point is not a backup; rather, it is a feature of the product.
Early jurisdictional design
A single agent design will not be consistent across markets due to differences in regulatory regimes. Before the agent is constructed, governance must comprehend the regulations of each jurisdiction.
Nandan connected all of this to the fundamental skill of trust. According to her framing, any decision-making system in a regulated setting must first provide the client with assurance before attempting to change behavior, provide options, or foster loyalty. When questioned about decision-making logic in high-stakes situations, Nandan was straightforward:
“Any technology must foster confidence. It must be able to indicate that you are safe even though you are interacting with a bot, artificial intelligence, or technology. You are in capable hands, and your information is secure. Building that trust is crucial, in my opinion, as a fundamental skill.
— Shri Nandan, Comcast Vise President of AI Products and Experiences
Consolidated Client Information for Dependable AI Context
When asked if data, regulations, or organizational preparation are the true obstacles in these fields, Nandan responded that they are all three, but he concentrated on data and the organizational behavior that underlies it. Customer data is usually owned by numerous departments within a major organization. Before a team can consider making that data AI-ready, the first challenge is to create a single source of truth from those different holdings.
Leaders must take into account the location of the processing that transforms unified, AI-ready data into consumer context. Nandan referred to this as the computation’s “gravity.” A poorly designed data architecture raises operating expenses as AI usage grows, and processing that is located too far away from client engagement may cause latency and performance problems. This means that architecture is no longer a technical issue to be resolved after deployment, but rather an early scaling decision for company leaders.
Nandan was adamant that data is “one hundred percent important” and the secret to a positive customer experience, but she found the difficult part outside of technology:
“More than anything else, the issue of producing high-quality data, data with integrity, and single sources of truth is cultural.” It’s really hard to say you have to give up data that is owned by five distinct sets of disparate teams. For 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,” some organizational changes must be made.
— Shri Nandan, Comcast Vise President of AI Products and Experiences
Nandan contended that the shift must originate at the top and that the executive consequence is that AI-ready consumer data is an organizational ownership issue before it is a technical integration issue. Without such alignment, as well as the data architecture and strategy to enable it, organizations wind up developing AI on a fragmented consumer context, which results in uneven customer experiences and more friction.
Nandan’s advice on which AI capabilities are important is straightforward once the data foundation is in place: any capability that promptly resolves a customer issue. The power to hyper-personalize is what the unified data adds. An agent may provide experiences virtually instantly with new, integrated context regarding a customer’s history and journey. For instance, they can see that a customer has repeatedly requested assistance with the same issue without finding a solution and direct them in a different direction. The product plan is then informed by the agent’s performance evaluation; things that don’t work are removed, those that do are expanded, and the user experience gets better over time.
Nandan’s account offers a step-by-step guide for leaders attempting to transition from fragmented data to a useful consumer context:
- Consider data consolidation as an executive mandate: Instead of leaving customer data in the hands of whichever business unit happens to own it, leadership must create it as an organizational resource with enterprise-wide regulations for access, stewardship, and accountability.
- Define ownership and freshness in addition to the single source of truth: A consolidated record that is outdated or that no one is responsible for maintaining up to date does not provide models or agents with trustworthy context.
- Choose the location of computation before scaling: It is an architectural decision with long-term cost implications to place the processing that creates customer context close enough to the customer to prevent latency without duplicating heavy infrastructure.
- Use agent assessment as the roadmap input: Instead of designing AI features in advance, create the agent, assess its performance against actual customer issues, and allow the outcomes dictate what is developed next.
Nandan added that just as the data has been fragmented, so too has the governance picture. Data governance, AI governance, and context governance—all of which now require separate safeguards in a regulated business—have replaced the unified canopy of digital governance that existed fifteen years ago.
Centralized AI Governance to Guarantee Quicker Operational Growth
Nandan outlined a defined order of operations for the practical stages toward an operating paradigm where AI enhances service quality at scale. An AI governance approach is the first prerequisite; without safeguards, everyone goes in various ways and anarchy ensues. A good data strategy goes hand in hand with that. An organization should only start experimenting and innovating after both are in place.
In order to allow teams to test new technologies and create proofs of concept (POCs) without committing to enterprise-scale deployment, Nandan recommended setting aside specific capacity for experimentation. The criteria for determining which experiments should continue and which should be discontinued are then provided by governance and data strategy. According to her, an innovation center of that type helps businesses stay up to date with a technological landscape that is evolving at “lightning speed.”
She was equally clear that failure is not eliminated by any of this. When things go wrong, leaders’ ability to quickly adjust is what makes a difference. Referring back to the topic of red flags from earlier in the discussion, Nandan provided a guideline on when to stop:
“Have the guts to pause and ask yourself what needs to be changed if you notice that none of the KPIs are changing. If things are not going in the proper way, we don’t need to keep trying. Thus, in addition to all of the governance and fundamental core competencies in place, a significant portion of it is great leadership.
— Shri Nandan, Comcast Vise President of AI Products and Experiences
The discussion’s most insightful section focused on what BFSI and healthcare firms do differently when they achieve genuine operating scale. Nandan warned that businesses are still in the early phases of developing enterprise AI operating models, calling this stage “nascent” and pointing out that even top universities are still trying and failing. More focused AI decision-making is a pattern she observes among those making faster progress, and she emphasized why implementing it too widely throughout the company backfires:
According to Nandan, businesses can advance more quickly by consolidating AI responsibilities under a single department, which eliminates conflicting decision-making processes and organizational politics. That stance is also consistent with the governance-first sequence she previously outlined: leadership has unambiguous power to operate within certain bounds, while guardrails are provided by a governing body that is knowledgeable about the business, pertinent laws, and internal regulations.
Nandan’s operating model can be summed up as a series of enduring commitments:
“I believe that when you say, ‘I want democratization of AI, and I want everyone to work on it, and I want everyone to have opinions and make judgments, that causes a lot of turmoil and uncertainty, and then individuals end up having to report to five different managers. However, I believe that it is incredibly useful to move things forward and avoid wasting time on politics when businesses establish a tight AI unit and have strong leadership making very brave and swift decisions without thinking about the optics.
— Shri Nandan, Comcast’s Vise President of AI Products and Experiences
Establish AI governance and data strategy before funding use cases. These are the standards by which all subsequent experiments are evaluated, and developing them after the fact entails adding controls to systems that are already in operation.
A POC gains the right to scale by fitting the data strategy and governance structure, not just by being enthusiastic. Set aside explicit capacity for experimentation and gate scaling on governance.
Establish success thresholds before scaling: To ensure that stopping an AI endeavor is a question of discipline rather than a protracted negotiation, leadership should establish beforehand which customer and business outcomes decide whether an AI initiative advances, changes course, or ceases.
Concentrate AI decision authority during early scaling: While governance procedures and operational models are still developing, a highly accountable AI unit with clear leadership and the power to make quick judgments eliminates the political overhead that comes with widely dispersed decision-making.