At VB Transform 2026, NTT DATA AIVista CEO Bratin Saha spoke with VentureBeat CEO and editor‑in‑chief Matt Marshall about the operational challenge of putting frontier models into regulated production. Their discussion focused on the practical question confronting companies investing heavily in AI: how to convert that spend into measurable enterprise value.
Saha emphasized that deploying AI is more than installing a model: "It's not just a model, you're building a system around the model." The ‘‘last mile’’ consists of wrapping a frontier model in a firm's own data, workflows and guardrails so it becomes a reliable agent in production.
Why frontier models stall in enterprise workflows
According to Saha, most enterprise AI projects fail during implementation because of poor integration, gaps in domain specialization, lack of governance, and unclear ownership of outcomes. Frontier models such as Fable 5, Opus 4.8, and GPT‑5.5 often fall short of production‑grade accuracy on many real insurance workflows — especially those that are complex, multinational, or involve handwritten forms and many exception cases.
Saha argued that the biggest improvements come from specializing the entire AI system rather than only fine‑tuning the foundation model. That system specialization uses the customer's proprietary data, their workflows and often the tacit, undocumented tribal knowledge that doesn't exist in operating manuals.
The three components of the last mile
Saha laid out three practical elements of the work:
- capturing the enterprise’s context and making it consumable by AI;
- running an ensemble of models so that costs remain controlled;
- adding specialized guardrails that validate the model output and force rework when the model is wrong.
He noted that fine‑tuning ranks low among companies' model‑selection priorities in recent VentureBeat enterprise survey results, reinforcing that the last mile centers on domain expertise and private workflows rather than model tuning.
What agentic AI actually requires
The core of the last mile is using proprietary corporate data to build a system that can steer the model correctly and place appropriate guardrails around it. In enterprise AI, the objective is to move a workflow from point A to point B rather than to deploy a standalone technology; value is created by the workflow that is moved, not by the model alone.
NTT DATA’s approach pairs AI experts with subject‑matter experts: teams interview workers to learn how they actually perform tasks and then encode that knowledge into agents. Across regulated industries like insurance and manufacturing, success depends simultaneously on technology, domain expertise, and change‑management capability — and Saha contends that technology itself is rarely the bottleneck.
Where solutions stay bespoke and where they scale
Keeping intelligence in the surrounding system rather than embedding it solely in the model preserves swappability and allows enterprises to use open‑weight and open‑source models as they mature. Saha's team runs an ensemble that mixes frontier and open‑source models, with the expectation that open weights will be used where the cost of error is low and frontier reasoning reserved for high‑risk cases where a few percentage points of additional accuracy matter.
On the platform side, guardrail generation and neurosymbolic components can scale across customers, but capturing each organization’s tribal knowledge remains a bespoke effort. Saha pointed to NTT DATA’s position as one of the world's largest insurance third‑party administrators as a durable advantage in acquiring that domain expertise built up over about 20 years.
Bottom line: where ROI comes from
Saha concluded that the return on enterprise AI investment comes not from the foundation model itself but from the work built around it: domain specialization, guardrails, and disciplined change management. Organizations that focus investment only on the model risk leaving most of the potential return on the table.
(Sponsored content: this article was presented by NTT DATA AIVista.)
