What has to change before an AI pilot can become part of a regulated financial workflow?
In this episode of AI at Work, I speak with Tom Carey, President of Broadridge's Global Technology and Operations Business, about the work required to move AI beyond personal productivity tools and prototypes. Tom brings an unusual perspective to the conversation. He studied AI in the 1990s, long before current computing power and language models made commercial deployment practical, and now works where financial market infrastructure, technology and operations meet.
Tom describes four levels of AI productivity. The first helps an individual complete a task. The second turns successful personal patterns into tools that other employees can use. The third places agents inside production workflows while a person verifies the result. The fourth allows agents to operate without constant human review, but only within a controlled environment with monitoring, policies and evidence that the system is behaving as expected.
One of the strongest lessons from our conversation is that AI can force an organization to examine a process it may have accepted for years. Tom says Broadridge removed 30 percent of the email traffic entering one operational group after finding duplicates, messages that required no response and work that should not have existed. His own estimate is that around half the benefit of an AI program may come from simplifying the current process before advanced models are applied. That is an important counterweight to the assumption that every inefficiency needs a language model.
We also discuss the economics of enterprise AI. Tom compares current token-cost surprises with the early cloud projects that moved existing workloads without redesigning them for a different computing model. Broadridge baselines costs, tests use cases through its operations business and monitors token consumption. He shares one example in which a team's default configuration automatically moved to a newer and costlier model, showing why model selection and configuration belong in the operating discipline around AI.
For financial institutions choosing their first agentic workflow, Tom recommends high-volume, rules-based work with known inputs and outputs. A person or an automated test should be able to verify the result. He also favors several smaller agents over one large agent responsible for an entire process because individual steps are easier to inspect, replace and stop when something goes wrong.
As models become widely available, Tom believes business advantage will increasingly depend on the proprietary data, platforms, workflows and governance surrounding them. Does your organization have the process discipline, cost visibility and control framework required to put AI into production? Listen to the episode and share your thoughts with me.
Useful Links
Executive: Tom Carey, President of Broadridge's Global Technology & Operations business
Broadridge Deploys Agentic AI at Institutional Scale Across Capital Markets and Wealth Operations (May 11, 2026)
Broadridge Joins Anthropic's Project Glasswing (June 17, 2026)
GenAI Delivering Now, Tokenization Is Next: Financial Services Enters Period of Accelerating Transformation, Landmark Broadridge Study Finds (February 25, 2026)
LTX Launches Agentic AI in BondGPT, Turning AI Insights into Trading Action (June 16, 2026)
Broadridge Invests in DeepSee, Further Harnessing Agentic AI to Transform Post-trade Operations (January 8, 2026)