Why most Enterprise AI initiatives never scale — And what can fix it

Enterprises have no shortage of AI models. What they lack, according to Sanjoy Ghosh, Global Head of Technology, Engineering and R&D Services at HCLTech, is the engineering data foundation to run them at scale. In a wide-ranging conversation, Ghosh argues that the real barrier to enterprise AI isn’t the model — it’s a data problem hiding in plain sight, inside the walled-off world of engineering and operational systems that most enterprise AI strategies never touch.

The pilot-to-scale trap
Pilots succeed because they’re built on a single, curated dataset. Production AI has no such luxury — it has to work across an organization’s full complexity: fragmented IT systems, operational technology, engineering technology, and decades of accumulated institutional knowledge. “That is a fundamentally harder problem,” Ghosh says, “and it is the one HCLTech has invested in solving.”

HCLTech’s answer is what it calls an AI-ready intelligence layer that unifies enterprise IT, operational technology (OT), and engineering technology (ET) data — giving R&D teams a real-time, trusted view of information that has traditionally lived in silos. The payoff shows up in hard numbers: for a leading European automotive OEM, HCLTech applied AI/ML to vehicle test planning — matching test requests to variants, automating workflows, optimizing asset use — and delivered a 30% productivity gain along with fully secured data transactions.

Three gaps standing between enterprises and AI value
Ghosh points to a consistent pattern behind stalled AI initiatives:

A disconnect between enterprise and engineering data. ERP, CRM, and finance systems are well integrated; product telemetry, test results, and PLM data are often still siloed or unstructured. AI trained only on enterprise data is effectively blind to the engineering reality of the business.

Data that records history instead of driving decisions. Much engineering data is archival — hard to query, disconnected from live systems. Ghosh argues it needs to become active: continuously updated and available at the point of decision.

Governance built for the wrong era. Legacy IT governance frameworks weren’t designed for the volume and diversity of engineering data — sensor feeds, simulations, structured and unstructured formats — so it often goes untracked entirely. The fix, he says, is governance built for engineering scale that preserves traceability without throttling the teams AI is meant to serve.

Agentic AI raises the stakes
As AI shifts from generating insights to taking autonomous action — running test plans, making design trade-offs, managing workflows — the cost of bad data changes character. “An agent acting on poor data does not just produce a wrong answer,” Ghosh warns, “it can derail a program.”

He lays out three requirements for deploying agentic AI safely:

1. Real-time, contextualized data unifying IT, OT, and ET so agents can act on live, complete information.

2. Explicit governance guardrails — permissions limiting what an agent can touch, approval gates for high-impact decisions, and kill switches for human override. In regulated sectors like aerospace and medical devices, Ghosh notes, this isn’t optional — it’s a certification requirement.

3. Engineering infrastructure built for observability, capable of detecting drift and escalating anomalies before they compound.

A new mandate for engineering — and for CIOs
Ghosh sees data engineering shifting from a back-end IT function to a strategic R&D capability in its own right. Engineering organizations, he says, are increasingly taking ownership of data across the full product lifecycle — design, simulation, testing, manufacturing, field deployment — creating a feedback loop where test failures inform design and field telemetry accelerates the next development cycle.

His advice to CIOs: change the question. Not “how do we manage our data infrastructure,” but “how do we build an engineering intelligence layer that makes our best people more powerful?” That means investing in data quality as an engineering discipline and partnering with organizations fluent in both the engineering domain and the AI stack — an area where HCLTech points to more than 150 proprietary accelerators and 300-plus global customers as evidence of execution at scale.

Where AI maturity is already real
Advanced manufacturing, aerospace and defense, semiconductors, and automotive lead the pack — not because their AI models are more advanced, but because their engineering data ecosystems are. The common thread: these organizations have dissolved the wall between enterprise and engineering data, and their AI is embedded directly in engineering workflows rather than layered on top of them. Just as important, Ghosh notes, is organizational alignment — when the chief engineer and chief data officer share ownership of the data strategy, AI maturity tends to follow.

The key metrics to watch out for

The metric CXOs should be watching, Ghosh argues, isn’t the number of models in production — it’s engineering velocity: how fast teams move from hypothesis to insight, from design to validated prototype. “AI success is not an IT outcome,” he says, “it is an engineering outcome.”

Getting the data foundation right, in his framing, isn’t a prerequisite to the AI journey — it is the AI journey. “The enterprises that will lead in product innovation and engineering excellence over the next decade are those building this intelligence layer today.”