Your Company Bought the AI. Now Comes the Harder Part.
NTT DATA AIVista's CEO says most enterprise AI projects fail not because the model is bad, but because no one does the unglamorous work of wrapping it in real workflows, real data, and real guardrails.
Key points
- Most enterprise AI projects fail during implementation because of poor integration, missing domain expertise, and unclear ownership of outcomes.
- NTT DATA AIVista CEO Bratin Saha says spending aimed only at buying a frontier model leaves most of the financial return unclaimed.
- Fine-tuning, the process of retraining a model on new data, ranked last in VentureBeat's latest enterprise survey of model-selection priorities.
- NTT DATA AIVista runs a mix of frontier and open-source models, reserving the most expensive reasoning for decisions where a mistake carries serious cost.
- Saha says enterprises need technology, domain expertise, and change management together, and that technology is the least of the three bottlenecks.
Every week, another company announces a big AI budget. And every week, a quieter story goes untold: the cheque cleared, the model got deployed, and almost nothing changed.
Bratin Saha, CEO of NTT DATA AIVista, the AI division of one of the world's largest IT services firms, spoke about exactly that problem at VB Transform 2026. His argument is blunt. Buying a capable AI model is the easy part. Everything that surrounds it is where projects live or die.
"It's not just a model," Saha said. "You're building a system around the model."
Why do so many enterprise AI projects stall?
They stall because frontier models, the most powerful AI systems available today such as GPT-5.5 or Fable 5, are trained on general knowledge. Real business workflows are anything but general.
Saha used multinational insurance claims as an example. Those forms carry handwriting and regulatory rules that shift by country. A general-purpose model reads them poorly. Accuracy stays too low for a regulated business to trust.
The fix is not simply buying a bigger model. Saha's team calls the fix "last-mile specialisation": taking a capable foundation model and wrapping it in three layers. First, the company's own proprietary data and context, fed in so the model understands the specific business. Second, an ensemble of models, meaning several AI systems working together, so computing costs stay manageable. Third, specialised guardrails, automated checks that catch errors and force a redo before a wrong answer reaches a customer or a regulator.
"The biggest bang for the buck comes from the specialisation and then these specialised guardrails," Saha said.
What does this cost, and who pays for what?
| Layer | What it is | Who typically owns it |
|---|---|---|
| Foundation model | The core AI (GPT-5.5, Fable 5, open-source) | Model vendor |
| Domain context | Company data, undocumented workflows | Enterprise + integrator |
| Model ensemble | Mix of frontier and cheaper open models | Integrator |
| Guardrails | Automated error-checking layer | Integrator + enterprise |
| Change management | Training staff, rewriting processes | Enterprise |
The model licence fee is often the smallest line on that budget. Capturing the undocumented "tribal knowledge", the stuff veteran workers know but never wrote down, costs the most. Saha says NTT's team gets at it by sitting with human workers and asking, plainly, how they actually do their jobs.
"The only reason is because we go and talk to those human workers," he said.
That scarcity of people who can do this work is real. Our 30 July story on the engineering talent gap reported that demand for specialists who connect AI to real business workflows is projected to surge by 2,100 percent. Saha's point about domain expertise being the bottleneck looks less like a sales pitch and more like a structural problem.
What does this mean for companies weighing an AI investment?
Budget for the whole system, not just the model subscription.
Saha says the sequence matters too. Start by embedding AI into existing workflows rather than redesigning them from scratch. Workers running mission-critical operations will not accept a vendor tearing out a process mid-flight. Once trust is built and results show up, the deeper redesign becomes possible, and that is where the largest returns sit.
There is an honest caveat worth naming. Saha was speaking at a sponsored session; NTT DATA AIVista paid for the platform. His advice points, conveniently, toward hiring an experienced integrator. Many smaller businesses will not need a firm of that scale. The underlying principle holds regardless of who does the work: the model is table stakes, and the workflow is the prize.
Takeaway: Before signing any AI contract, map the workflow you want to change from start to finish. If you cannot describe what changes at each step, and who owns the outcome, the technology budget is running ahead of the plan.
