The gap between buying frontier models and shipping production AI just got a name: the last mile. And it's where most enterprise AI budgets go to die.
The Summary
- NTT DATA AIVista CEO Bratin Saha laid out why frontier models like GPT-5.5 and Opus 4.8 fail in regulated enterprise workflows — they lack domain-specific guardrails, institutional knowledge, and the context to handle complex real-world tasks like multinational insurance claims
- The "last mile" isn't about better models. It's about wrapping those models in enterprise data, workflows, compliance rules, and the tribal knowledge that never made it into documentation
- Most enterprise AI projects fail during implementation because of poor integration and unclear ownership, not because the underlying model isn't smart enough
The Signal
We're past the demo phase of enterprise AI. The models work fine in controlled environments. They impress in pitch decks. But regulated production environments expose a brutal truth: a frontier model that scores 89% on benchmarks might hit 60% accuracy on your actual insurance claim forms, which are handwritten, checkbox-heavy, and built around domain logic that exists nowhere in the training data.
The gap isn't a model problem. It's a system problem. Saha's framework is simple: you're not deploying a model, you're building a system around the model. That system needs to understand your risk appetite, your client classifications, your regulatory interpretations, and the institutional knowledge that's stuck in someone's head three desks over.
"It's not just a model, you're building a system around the model."
This is where the money gets real. Consider what "last-mile specialization" actually means in practice:
- Domain-specific workflows that reflect how work actually gets done, not how the org chart says it should
- Guardrails tuned to your regulatory environment, not generic safety theater
- Integration with the legacy systems that still run 80% of enterprise operations
- Ownership models that make it clear who's responsible when the AI screws up
The insurance example Saha uses isn't hypothetical. Multinational claims involve forms that are complex, often handwritten, full of checkboxes, and governed by regulations that vary by jurisdiction. A foundation model trained on the open internet has no idea what to do with that. It doesn't know your underwriting standards. It doesn't know which claims need three signatures versus one. It doesn't know that the handwriting on line 47 overrides the checkbox on line 12 if they conflict.
This is the unsexy part of the agent economy that nobody talks about at conferences. Not the breakthrough model releases. Not the AGI timelines. The actual work of making AI agents useful inside organizations that have compliance departments, legacy systems, and people whose jobs depend on these things working correctly.
The Implication
If you're running enterprise AI, this is your checklist. Before you buy another model API subscription, ask: do we have the data infrastructure to specialize this model for our actual workflows? Do we have clear ownership when something goes wrong? Can we explain to regulators why the AI made the decision it made?
The companies that crack last-mile specialization won't just deploy better AI. They'll build moats around institutional knowledge that foundation model providers can't replicate. Watch who's building systems, not just licensing models.