The science and engineering behind NTT DATA AI for Insurance

Today, NTT DATA launched AI for Insurance, an enhanced service-as-software solution powered by the NTT DATA AIVista platform.

Service-as-software solutions give carriers an ongoing, managed way to deploy AI agents that perform defined insurance work, integrate with existing systems and improve over time without requiring the carrier to build and operate the full technology stack itself. Our solution transforms complex insurance workflows into governed, repeatable AI-delivered agents. It is built for operationalizing the last mile of enterprise AI, where a general-purpose model has to be specialized to a client’s own policies, rules and systems before it can be deployed in production.

Insurance is one of the most complex use cases for AI deployment. Every decision, whether it’s pricing a risk, approving a claim or servicing a policy, carries regulatory weight and financial consequences. This complexity compounds globally: Requirements vary across regions, product lines, distribution models, languages and regulatory frameworks. The knowledge that drives those decisions is rarely straightforward or centralized. It is found in underwriting guidelines, claims playbooks, regulatory filings and the institutional expertise of people who have spent their careers in the organization or industry.

Getting AI to work reliably in this environment requires more than a capable model. Today’s foundation models can synthesize broad industry information at scale, but they lack the carrier-specific domain context, sufficient knowledge about compliance requirements, and the system integrations that regulated insurance operations demand. They also lack an insurance ontology: a structured representation of the (often client-specific) entities, relationships, rules and exceptions (in other words, the “last mile knowledge” about the client) that define how the business actually works.

The challenge is therefore simultaneously technological, operational, scientific and contextual. This post explains how we addressed that combined challenge, why the approach works and what it took to move from experimentation to enterprise execution.

The problem we set out to solve: Operationalizing AI-delivered outcomes

Bruno Abril, Global Lead: Insurance Industry at NTT DATA

Insurance carriers run on processes that weren’t designed for the speed or complexity of today’s risk environment. Submissions arrive faster than teams can assess them, claims queues grow, regulatory demands multiply, and the data that should drive better decisions is scattered across systems that do not talk to each other.

Friction rarely appears in just one place. It is found in the handoffs between humans and systems, between structured data and judgment calls, and between what AI can do and what today’s operating model allows it to do.

As a result, carriers face mounting pressure to increase underwriting capacity, accelerate claims, reduce cost to serve, and meet rising governance expectations, all at the same time. Most have experimented with AI, but few have converted those experiments into durable operating value. Those pain points set the standard for what insurance AI must do: not simply demonstrate model capability but also improve the performance of real workflows under real business constraints.

Most carriers are still stuck in pilot mode

According to research conducted for NTT DATA’s 2026 Global AI Report: A Playbook for AI Leaders, 86.7% of insurance AI leaders support AI across back- and mid-office workflows. Yet our Insurtech Global Outlook 2026 found only 22% have scaled AI to production, even as 66% of the insurance workforce has adopted AI tools. The gap between experimentation and enterprise value is real, and it is widening.

The limiting factor is not merely technology. It is also the ability to combine that advanced technology with disciplined execution. Scaling AI in insurance requires three elements that many pilots skip: governance controls designed from the start, operating models rebuilt entirely around AI-enabled work (not just augmented by it) and a knowledge foundation grounded in the carrier’s actual policies, rules, systems and risk appetite rather than a generic industry approximation.

Forward-thinking insurers have stopped asking where they can try AI and started asking how they redesign operations around it. That shift changes everything. The opportunity is bigger than automating individual tasks. We can redesign insurance work so expertise reaches more decisions, customers receive faster and more consistent service, and carriers learn continuously from every interaction.

What AI-delivered operations mean in practice

AI-delivered operations are not automation for its own sake. They form a new operating model in which AI agents perform defined work across underwriting, claims, servicing and compliance, while people retain authority over judgment, exceptions and accountability. It means underwriting agents can assess submissions against carrier-specific risk rules to flag exceptions; use automated claims workflows to coordinate triage, validation and settlement; and rely on compliance functions that monitor decisions as work occurs rather than after the fact.

In a regulated business, confidence must flow from the way the system is designed. Clear decision rights, enforceable rules and traceable evidence allow carriers to expand AI into higher-value workflows without relinquishing control. Human expertise is amplified rather than replaced. Experienced professionals focus on judgment, customer impact, complex exceptions and the evolution of business rules, while AI handles repeatable analysis and coordination at scale.

The result is AI for Insurance, a service-as-software solution purpose-built not for a single carrier workflow but for the operating reality of a large, complex insurance enterprise. It combines insurance operations expertise, AI science, enterprise engineering and managed execution so carriers can move from isolated pilots to production outcomes.

What it took to build AI for Insurance

Mukesh Karki, CTO at NTT DATA AIVista

Building NTT DATA AI for Insurance required solving engineering problems that generic platforms are not designed to handle. That work is the last mile of enterprise AI: the distance between a capable foundation model and a system that can be trusted inside a regulated insurance operation. Closing it takes deliberate specialization rather than a bigger model, and last-mile specialization is what the AIVista platform is built to deliver. Three problems define that gap:

• Context: Foundation models possess broad public knowledge but lack the specific policies, regulatory rules and institutional knowledge that drive individual insurance decisions. We addressed this by building an AI-native knowledge base and insurance ontology that encode the relationships among policy terms, coverage conditions, regulatory rules, carrier-specific exceptions, submission data and underwriting criteria. Agents don’t search for relevant text; they reason against a structured representation of how the insurance business works.
• Governance: Insurers operate under strict regulatory oversight (such as the EU AI Act’s high-risk obligations and, in the US, the National Association of Insurance Commissioners’ model bulletin on AI use for insurers), so we built specialized guardrails where every agent’s decision passes through a policy enforcement plane. Regulatory requirements, carrier guidelines, authority limits and internal controls are evaluated before an action is taken or a recommendation is made. Human-in-the-loop review is placed at decision points where business judgment, customer impact or material risk warrant it. Each step produces structured evidence of what the agent received, which rules it applied, what it proposed and why. That evidence is generated as the work happens, so an audit is a query rather than a reconstruction.
• Agent orchestration: Insurance workflows don’t run in isolation. For example, a submission triggers a document extraction, which surfaces a coverage question. In turn, that requires a review by an underwriter who may uncover a regulatory flag, which then loops back to the originating system. The AIVista agent orchestration layer is event-driven at its core. This means agents activate on real workflow events, hand off to humans at defined decision boundaries, handle exceptions through configurable escalation paths and resume action without data loss when control returns. Each workflow is configurable to a carrier’s specific systems and operating mode, and the platform adapts to how the carrier works, rather than requiring the carrier to adapt to the platform.

Powered by NTT DATA AIVista, the new solution is designed for open integration with existing carrier systems and enterprise technology environments, enabling insurers to deploy AI capabilities without core system lock-in.

A look at the NTT DATA AI for Insurance console

The science behind the AIVista differentiation

George Karypis, Chief Scientist at NTT DATA AIVista

What separates NTT DATA AI for Insurance from a general-purpose agentic framework is the science underlying its knowledge and control architecture. To understand that difference, it helps to see why the standard approach falls short.

Most enterprise AI deployments rely on prompting and model-level checks to manage risk. Better prompts can reduce error rates. But in a regulated business, “AI usually follows the rule” is not a control. The moment you give an AI system autonomy to act inside a regulated workflow, you create a management problem that is different from a technology challenge. For example, an unchecked error can become a big compliance issue, a customer complaint or financial loss.

To make AI deployment successful in today’s agentic era, that control must come from the architecture.

The scientific foundation for our approach is straightforward and recognizes that neural models and rule-based systems solve different parts of the trust problem. Neural models are very good at understanding messy, real-world inputs such as submissions, policies, claims notes, emails and regulatory text. Rule-based and knowledge-driven systems are very good at applying explicit constraints, preserving consistency and producing reasoning that can be inspected. In high-stakes, regulated enterprise settings, the strongest architecture is not one that asks the model to behave deterministically. It’s one that combines probabilistic reasoning with deterministic enforcement. It means we let the model do what it is good at and implement a deterministic layer around it that the model cannot bypass.

In insurance, this distinction matters because the workflow itself is governed. Every action must respect carrier policy, authority levels, regulatory requirements, customer commitments and escalation rules.

AI for Insurance is built around this fundamental reality. The agent can interpret information, identify gaps, draft recommendations and coordinate next steps, but proposed actions are checked against the rules before they are executed. The result is not just faster processing. It’s AI that operates within business boundaries and leaves behind the evidence needed to explain what happened, why it happened and how the rules were followed.

At AIVista, we call this implementation Specialized Guardrails, a deterministic enforcement layer that checks every proposed agent action against the client’s policies, controls and regulatory requirements before execution.

The result is AI that does more than predict outcomes. It applies constraints, explains its decisions and produces the evidence required for internal review and regulatory scrutiny. In a domain where a single underwriting or claims error can carry real financial and legal consequences, reliability must be engineered rather than assumed.

What AI for Insurance looks like in practice

Rohit Puri, Executive Managing Director, Global Head: Business Process Services at NTT DATA

When one of North America’s largest life and annuity third-party administrators (TPAs) needed to modernize their policyholder services operation, NTT DATA AIVista delivered measurable results. Third-party administration refers to the outsourced management of an insurance carrier’s operations, where the TPA owns the strategy, technology and day-to-day operations to handle all incoming policyholder requests on behalf of their insurance clients. NTT DATA runs this function for major life and annuity carriers across North America, processing nearly 4 million customer transactions per year across more than 70 transaction types, touching six systems per request.

• The challenge: Every incoming customer request, whether an address change, beneficiary update or form submission, passed through a fully manual pipeline spanning six systems. A 30-person indexing team manually extracted data from scanned documents and categorized each request before it could reach the 2,500-person processing staff downstream. Classification errors cascaded into compliance risks. Business rules had been codified for only three of the more than 70 transaction types. And with no unified observability across systems, there was no way to measure performance, identify bottlenecks or build a case for improvement.
• The solution: AIVista deployed specialized AI agents to replace the repeatable, mundane manual chain with a single automated pass. AIVista’s Intelligent Document Processing Agent extracts and classifies every incoming request, including multitype documents, with confidence scores on every field. The AIVista Rules Validator Agent then connects to existing systems of record, retrieves the required context and runs full IGO/NIGO determination in seconds. Professionals concentrate on exceptions, low-confidence cases, customer-sensitive decisions and the improvement of operating rules. The entire pipeline runs on top of existing infrastructure, with no system replacement required.
• The results: The system achieved 98.5% accuracy compared with 50% to 72% for leading off-the-shelf frontier models, thereby raising automation potential to 98%, from a frontier-model-only baseline of 70%, with 25 times fewer documents sent for manual review and an estimated $7 million in annual savings for a partial TPA implementation in 2026. This was achieved at near cost parity with commodity models. The platform also includes confidence thresholds, enforceable controls and end-to-end traceability, which are industry prerequisites for AI-driven processing in regulated production environments.

Getting the last mile right

AI for Insurance is not a feature update. It is the result of deep collaboration among NTT DATA insurance domain experts, AI scientists, enterprise engineers and researchers working on a shared problem: How do you make AI work reliably inside one of the world’s most complex and regulated industries?

NTT DATA brings those disciplines together from design to day-to-day execution, closing the last-mile gap between a capable model and a dependable production system.

What carriers can do next

• Choose one high-friction workflow with measurable business value and clear decision boundaries.
• Map the carrier-specific rules, data, systems, exceptions and human authority that govern the work.
• Design controls, escalation paths and performance evidence before moving into production.
• Build for operational ownership, including monitoring, rule updates, model changes and continuous improvement.

Executive takeaway: Trustworthy AI in insurance must be engineered, governed and industry-specific. Carriers will create durable value only when advanced models are combined with insurance knowledge, enforceable business controls, human judgment and an operating model capable of sustaining performance at scale.