What does success look like when an enterprise launches its own generative AI (GenAI) initiative? For Swatee Singh, EVP, CIO – Data, AI and Corporate Functions at TIAA, it isn’t just about the number of tools deployed or the lines of code generated — it’s about measurable impact across people, clients, and operations. 

That philosophy is shaping GAIT, TIAA’s enterprise-wide GenAI platform which provides secure access to multiple large and small language models — including those from Anthropic, Meta, Google, and OpenAI — within a secure environment.

A Fortune 100 leader in retirement, wealth management, and asset management solutions, TIAA manages $1.4 trillion in assets, serves millions of customers, and supports thousands of institutions — a scale that makes adopting trusted, enterprise-ready AI essential.

In this interview, Singh frames AI success as employees mastering prompt engineering, customers gaining faster resolutions, and teams reclaiming lost hours. She highlights the GAIT initiative as business transformation, not a tech experiment, and emphasizes agentic AI’s role in pairing human associates with specialized AI agents across functions — anchored in governance and trust.

Edited Excerpts

Q: You have been at TIAA for over 3 years now, previously serving as the Chief Data Officer as well. Could you share some examples of data-driven digital transformation initiatives you’ve led over the last few years?

Let me start with data. The industry has been moving toward data mesh architectures and reusable data products, which become critical as organizations embrace AI. Over the past three years, we’ve created high-value data products serving businesses across the enterprise. These products are modeled, governed, and available in ready-to-use patterns, making them trusted assets.

For example, in our Salesforce implementation, data for sales teams, relationship managers, and wealth advisors comes directly from these products. All of this is housed in our go-to Enterprise Data Cloud platform. While silos remain, the platform — with strong governance, quality, and controls — has given us a trusted, enterprise-wide data foundation impacting retirement, wealth management, and shared services.

On the AI side, a key step was developing a Responsible AI Policy, as GenAI entered mainstream consciousness. Unlike many financial services institutions embedding AI within IT or cyber risk policies, we created a standalone policy focused on ethical and fair use of data and AI. This governs all AI initiatives at TIAA.

It ensures no data leaks while giving employees access to cutting-edge tools. By this year, it’s rolled out to over 6,000 employees, supporting tasks like drafting emails, summarizing meetings, and creating user stories. Governance is automated through GAIT, making it scalable.

The value extends beyond productivity. We are reimagining workflows using AI agents. In shared services and training, agents now help train new call center staff on products and processes. In insurance, where unstructured contract data is common, GenAI agents process information more efficiently, reducing effort and errors.

We always keep a human in the loop and are experimenting with small language models to reduce hallucinations. Over two and a half years, we’ve filed more than 30 patents, several addressing hallucinations and responsible AI use. These efforts span shared services and all major business lines. We continue to invest in technology and talent to ensure clients see the value of this transformation.

Q: GAIT was launched about a year ago. How do you measure the success of such an initiative, and what kind of impact has it had so far?

Success is measured in several ways. Sometimes it’s employee adoption — are people getting comfortable with prompt and context engineering, and prepared for a hybrid workforce where agents work alongside humans? Sometimes it’s client experience, as with improved search on TIAA.org, which now leverages GenAI to provide clear answers on retirement topics. And sometimes it’s efficiency — freeing researchers or developers to focus on high-value work.

When we launched GAIT, the first step was bringing large language models into a secure environment. No client data could leave or enter except through approved ingestion channels. At first, we kept the “blast radius” small. Marketing teams got early access, since creativity lent itself to experimentation. We also used it in cybersecurity. As scammers leveraged GenAI for phishing, our team used GAIT to simulate attacks and better train employees. Our CISO calls it “AI for cyber, cyber for AI.”

As GAIT gained multimodal capabilities, use cases expanded. In Learning & Development, regulatory training is mandatory, but repeating the same test questions each year made it less effective. Using GAIT, the team now generates fresh, high-quality questions, keeping training more engaging and effective.

From there, we scaled. By year-end, more than 9,000 employees — over half of TIAA — will be using GAIT. Many apply it for productivity tasks like summarizing meetings, drafting communications, or generating user stories. Others use specialized agents.

One example is our “empathy agent.” We often work with clients during difficult times — losing a spouse or facing financial hardship. This agent ensures interactions remain empathetic while staying compliant and factually accurate. It draws from compliance standards and uses AI to maintain the right tone throughout complex conversations.

Q: What lessons or best practices from leading the GAIT initiative would you share with other organizations trying to establish similar initiatives?

I’ve learned three key lessons from leading TIAA’s AI-first digital transformation. First, always start with client needs. Direct engagement through TIAA Client Tech Labs and initiatives like our intern-led ideation programs has given us practical insights to shape strategy.

Second, build consensus and secure leadership buy-in. Our AI-first approach is embraced across the C-Suite because it aligns with TIAA’s three pillars — lifetime income, customer delight, and operational strength. I see myself as both educator and evangelist, ensuring AI is viewed as business transformation, not just technology.

Finally, success requires people, process, and culture. Through our Guild Network, associates develop core skills while exploring new ones, fostering innovation, collaboration, and long-term talent empowerment.

Q: You mentioned “GAIT for Developers.” Is it similar to tools like GitHub Copilot that we see in the market?

It’s similar in spirit, but our goal goes beyond just coding assistance. We’re encouraging our technologists to create their own agents to automate repetitive tasks. Take software testing as an example. If I’m a tester repeating certain steps every day, why not design an agent to perform those tasks?

The vision is to embed agentic AI across the entire software development lifecycle. From the very beginning — say, a design or jam session with the business — we can record the meeting, and then GAIT can generate the first draft of user stories. Product owners can then refine, edit, and validate those drafts, saving significant time.

Designers, too, can use agents within tools like Figma to accelerate experimentation and prototyping. From there, agents can help translate designs into UI code, providing a strong first draft for developers.

I wouldn’t call this full automation, because developers still need to know how to prompt effectively and engineer solutions — but it changes their role. Developers are evolving from writing everything from scratch to becoming compilers of code snippets, applying their expertise to refine and assemble solutions rather than reinventing the wheel each time.

Q: While major tech firms like Google and Microsoft report that 25–30% of their code is AI‑generated, has TIAA tracked or benchmarked similar metrics internally?

Although I can’t offer specific metrics, I can say that we’re increasing the use of this emerging technology enterprise-wide. At this point, every developer in our company, employee or vendor, has access to pair programming tools. In the coming months, we aim to create an ecosystem of agents that standardize some activities (e.g., Testing) in the software development life cycle.

Where we find value is the use of these tools for closing out vulnerabilities faster and more completely, in creating excellent documentation that is then searchable by Op teams. Creation of architecture diagrams and data flows brings tremendous benefit to our developers and architecture associates.

That said, we have a “human-in-the-loop” philosophy for all software built using these tools, as we do find limitations of the tech and want to ensure security and quality of our software base. 

Q: Given that TIAA is more than 100 years old, with a lot of legacy systems, have you been able to solve that challenge?

We’re well on our way. Over the past few years, we’ve made a major push to reduce technical debt. Interestingly, because TIAA hadn’t invested in certain areas for a period of time, it allowed us to leapfrog a few generations and build a state-of-the-art data platform with high availability, strong governance, and trusted data quality.

All our new initiatives are now built on this modern platform, and we’ve been migrating legacy assets into it as well. While no company can realistically claim to have eliminated all data silos, we’re steadily moving in that direction. Our goal is to ensure clear ownership of data, establish true sources of truth, and create a transparent understanding of how data should be used and by whom. It’s very much a journey, but we’re making strong progress.

Q: Culture often becomes the biggest challenge in driving transformation, sometimes even more than technology itself. Do you agree?

Culture is always both the biggest challenge and the greatest enabler. You can build the best data platform in the world, but without the right culture, it won’t deliver its full value.

On the data side, culture starts with ownership. Every business must recognize that it owns its data, not in the technology sense, but in terms of responsibility: What the data is used for, how it should be improved, and how it adds value, always within the right legal and privacy boundaries. Driving this mindset of data ownership is ongoing work, because data is rarely black and white — there’s always a lot of gray.

The second aspect is building repeatable, trusted patterns. Our Enterprise Data Architect plays a key role here, ensuring we don’t create manual bottlenecks but instead embed best practices back into the ecosystem. We also embed solution architects in each CIO and shared services team to help scale that consistency.

To reinforce culture, we’ve created guilds for data, analytics, emerging tech, and AI. These guilds act as communities of practice where we share good practices, train people, and create learning paths. For example, we provide specific training for data owners and stewards on how to identify and protect critical data elements within TIAA.

On the AI side, the cultural challenge is similar. We can build sophisticated agents or enterprise platforms like GAIT, but their success depends on how we bring people along. To do this, we use multiple approaches. For instance, we have a very active Viva Engage channel on Teams where thousands of employees ask questions about AI tools. What’s great is that it’s not just my team responding; we have champions and evangelists across the business, including frontline sales staff, who share answers and encourage peers.

Q: Today, there is a lot of talk about Agentic AI and what it could mean for enterprises. How does TIAA see this iteration of Gen AI?

We already have agents embedded in the workforce. For example, every one of our software developers has access to pair-programming agents within GAIT.

But we’re also building out specialized agents across functions.

In marketing, for instance, we’re working on agents that can generate compliant, personalized campaigns in real-time. In audit, we’re exploring task-specific agents powered by small and large language models. In sales, we’re leveraging Salesforce agents to enhance productivity. And on the client side, we’re testing agents that can assist bilingual users or guide customers through complex self-service processes on our website.

Looking ahead, we’re asking bigger questions:

• How will the role of a manager change when teams include both human associates and AI agents?
• Who owns an agent?
• How do we measure its performance, decide when to decommission it, and provide feedback for improvement?

We have a clear goal of developing hundreds of agents over time. But we’re doing it thoughtfully — always balancing innovation with governance, performance monitoring, and responsible adoption. The workforce of the future will be hybrid, and we’re laying the foundation for that today.

Q: Do you think we’ll soon see AI agents in financial services that can perform major actions like selling an insurance policy, without any human involvement?

I don’t see that happening fully end-to-end in the near future, at least not in financial services. In areas like shopping or travel, you can already imagine personal agents that compare options and even book something for you. For example, finding the cheapest flight with the fewest layovers. Those decisions can be safely automated.

But in financial services, the threshold is much higher. Decisions like canceling or selling insurance are significant, often with long-term impacts, so I don’t see them being completely automated in the next two to three years. What we’re more likely to see are ambient, autonomous agents that can act in the background, complete tasks, and even make recommendations, but with a human in the loop to review, intervene, or make the final decision. For now, humans will remain central to the process.

Q: What core challenges do you see in deploying these agents at scale?

One of the first challenges with AI agents is ensuring they are truly useful and reliable. Agents must consistently perform the tasks assigned to them, which means carefully selecting the right model for the job. Sometimes, a massive 70-billion-parameter model isn’t necessary; a smaller, more specialized model may deliver better results for a specific use case.

The next hurdle is deployment. A decade ago, the hardest part of machine learning wasn’t building models; it was moving them into production. In many companies, that process took months, even up to nine. Agentic AI could face a similar bottleneck: How do we safely and efficiently transition from development to large-scale production?

But the biggest challenge lies in governance and trust. Orchestration agents today can spin up other agents, but how do we control that? How do we ensure they access only the right data, for the right purpose, in a compliant way?

In industries like financial services, trust is paramount — client data must be safeguarded, and agent performance must be monitored over time.

Finally, there’s the risk of “spaghetti agents.” Without proper oversight, companies could end up with duplicate agents, inconsistent quality, and no clear accountability. Guardrails, governance, and quality control will ultimately define whether enterprises achieve success — or chaos — with agentic AI.

Q: What’s one trend in enterprise data and AI that excites you most — and one that’s overhyped?

The most exciting trend today is the rise of agentic AI, powered by AI-ready, high-quality data infrastructure — what many call “Agentic Data.” Unlike traditional AI that delivers insights, agentic AI can act, optimize, and learn in real time. This leap requires unified, governed, and trusted data platforms to ensure accuracy, context, and compliance. Enterprises are now rapidly investing in such infrastructure to support AI-centric operations.

In contrast, talks about Artificial General Intelligence (AGI) remain overhyped. Despite media buzz, the research community agrees it’s far from reality. For now, autonomous AI agents running the world is more science fiction than an imminent business transformation.