For much of the past decade, enterprise AI has been discussed as a technology problem. Build the model, test the output, ship the product and improve it over time.
Banking and insurance do not work that way.
In life insurance, an underwriting model that produces an incorrect decision can create consequences that surface years later. In banking, a customer-facing AI assistant that gives the wrong answer is not just a user experience issue. It can become a compliance and liability problem.
That is the operating environment behind AI Beyond the Pilot: The Enterprise Blueprint for Trusted AI in Banking and Insurance.
Abhishek Kumar, corporate vice president at New York Life and a Forbes Technology Council member, works across AI product strategy, financial services technology and regulated enterprise architecture.
Why pilots stall
The book opens with a familiar enterprise problem. Business units have AI use cases, vendors, models and projected returns. But when compliance, legal, product and technology leaders begin asking who owns the risk, how decisions are recorded and what happens when the system is wrong, momentum slows.
Kumar frames this as an institutional architecture problem.
In regulated firms, AI products need clear data ownership, evaluation methods, audit trails, access controls and defined accountability.
That is where many pilots fail. They prove that a model can work in a narrow setting, but they do not create the structure needed to make the product safe, repeatable and governable across the enterprise.
Platform, pattern, practice and proof
Kumar organizes the book around four ideas: platform, pattern, practice and proof.
Platform refers to the technical and organizational foundation that lets AI products exist inside a regulated firm. That includes shared infrastructure, security controls, data governance and model-risk processes.
Pattern is the layer of reusable components that allows teams to avoid rebuilding the same approvals, retrieval pipelines, evaluation tools and guardrails for every new use case.
Practice covers adoption. AI systems have to fit into the way business units actually work, which means product teams need legal, compliance, technology and operational partners involved early.
Proof is the measurement layer. A regulated enterprise needs to know whether an AI system is producing value, whether it is behaving as expected and whether the evidence is strong enough to support continued use.
Why consumer-tech AI playbooks do not transfer cleanly
One of the book’s main arguments is that methods borrowed from consumer technology often break inside financial services and insurance.
Consumer-tech teams can often iterate quickly, tolerate some errors and rely on incremental exposure. In regulated industries, a change to a customer-facing AI product may require model review, updated disclosures, access controls, documentation and additional testing before release.
When regulatory questions are addressed late, teams can spend months rebuilding products that were technically promising but institutionally unusable.
For AI leaders in banking and insurance, the lesson is that responsible deployment has to be designed into the system early.
Building AI products that can be governed
The book’s more technical sections draw on Kumar’s work with enterprise AI platforms.
One example centers on retrieval architecture. If two business units maintain different versions of the same source document, and two AI tools give different answers, a firm needs to know which answer was correct and why.
That problem is not solved by a better model alone. It requires source authority, ownership, review dates, version histories and clear links between AI outputs and the material used to generate them.
Kumar uses that example to show why trusted AI depends on more than prompts or model choice. It depends on the systems that control what the model is allowed to see, how answers are evaluated and how decisions are traced after the fact.
The organizational work behind AI scale
The second half of the book focuses on the work technical platforms cannot do by themselves.
Kumar argues that enterprise AI requires cross-functional product leadership. Legal, compliance, data, engineering, risk and business teams cannot operate as separate review gates if the goal is to ship responsibly. They need defined decision rights and a cadence for resolving issues before they stall the project.
Kumar distinguishes between drafting agents, action proposers, supervised actors and more autonomous systems. Each category carries a different risk profile and requires different design controls.
The key question is not only what the system did. It is whether the organization can explain what happened, why it happened and who was responsible for approving the conditions under which it happened.
From AI experiments to operating models
“AI Beyond the Pilot” reframes enterprise AI as an operating-model challenge.
The book does not argue that models are unimportant. Instead, it argues that models are only one part of production readiness. For regulated firms, the surrounding structure determines whether an AI product can move beyond a pilot and become part of normal business operations.
That structure includes technical platforms, reusable patterns, adoption practices, governance proof and clear accountability.
For banks and insurers, the path to trusted AI may depend on building the foundation correctly before rushing more pilots into production.
©2026 Cox Media Group
