40% faster engineering, and still no illusions about what AI costs, says Bosch India CIO and CDO

At Bosch India’s engineering centres, a document called DRBFM (design review based failure mode analysis) has long been the preserve of experience. Only engineers certified at L1, typically after at least five years on a product line, were trusted to produce one, mapping every way an engineering change could fail once it reached production.

AI has changed who can draft that document, but not who remains accountable for it. “Even somebody who is just six months into that product can also do this DRBFM document now,” says Shankar Rao, Chief Information Officer and Chief Digital Officer at Bosch India. Trained on years of engineering knowledge and historical failure data, an AI system prepares the first draft. The certified engineer no longer spends weeks assembling it from scratch. Instead, the engineer reviews, validates and signs off.

“Now he spends a few hours, maximum a day or two,” Rao says. “It is human in the loop.” The example captures Bosch India’s broader approach to enterprise AI. It is neither a story about replacing engineers nor one about chasing automation for its own sake. It is about identifying work where AI compresses effort without transferring accountability. That distinction becomes more important as enterprises move beyond pilots and begin confronting a harder question than whether AI works, the questions of whether it pays or not.

AI’s most difficult conversation

For Rao, the biggest challenge in enterprise AI is no longer building models. It is understanding their economics. Most organisations know how much they spend on cloud infrastructure. Far fewer know how much each AI use case actually costs relative to the value it creates.

He illustrates the problem with a deliberately simple example. A company deploys an AI application that eliminates work equivalent to one full-time employee, saving perhaps ₹10 lakh to ₹15 lakh annually. Meanwhile, the underlying token consumption costs close to ₹1 crore.

“You are spending one crore on an AI solution to reduce one FTE,” Rao says. “You need to look at it holistically.” That level of visibility, he argues, still does not exist.

Hyperscalers report overall AI consumption, but CIOs increasingly need to understand costs at the level of individual applications, where investment decisions are actually made. “I want to control it at a use case level,” Rao says. “That is still a gap.”

The economics become even harder once models begin drifting away from the data on which they were trained. A system that performs well today may quietly deteriorate over time, demanding continuous monitoring and retraining long after the initial deployment. Those realities have forced Bosch to become more selective about where it applies generative AI.

Every proposed AI application is first logged into an internal AI agent registry before development begins. Many never proceed. “Not every problem shape needs to be solved by GenAI,” Rao says, describing instances where engineers instinctively reach for a chatbot to solve problems that require nothing more sophisticated than a calculator. “People are using GenAI just because it is easy, rather than asking if it is really the right use case.”

Where Bosch says AI is working

The caution does not mean Bosch has struggled to find measurable gains. Engineering has produced some of the clearest results. DRBFM is one example within a broader redesign of Bosch’s engineering V-cycle, spanning requirements gathering, design, testing and validation.

Customer requirements often begin as fragmented information, a short email, a conversation, scattered screenshots and follow-up messages spread across multiple teams. Engineers traditionally spent days reconstructing that context before meaningful design work could even begin. AI now assembles those interactions into a structured draft specification while identifying gaps and proposing the questions engineers should clarify with customers.

Across the engineering cycle, Rao estimates productivity has improved by roughly 40% within Bosch’s own operations. Maintenance and manufacturing have produced similarly measurable improvements.

Predictive AI systems now identify equipment failures before breakdowns occur. On some production lines, Rao says, weekly unplanned downtime has fallen from roughly four hours to only a few minutes.

The same approach extends into logistics, where AI analyses shipment patterns, consolidates partial truckloads and strengthens negotiations with Bosch’s network of logistics providers.

Not every initiative has produced the same outcome. Digital twins remain one of Bosch’s more cautious experiments. The company has deployed them on individual production lines, including a valve manufacturing stream, but has deliberately stopped short of wider rollout. “The real value does not come from a pilot, it comes from scale,” Rao says.

Scale, however, has proved more expensive than expected, while demonstrating measurable business value has been harder than anticipated. “I am not happy with the way we have deployed it.”

For Rao, the lesson is straightforward. Pilots demonstrate technical feasibility. They do not necessarily prove commercial viability.

Manufacturing intelligence is evolving faster than product intelligence

One assumption surrounding industrial AI is that intelligent factories and intelligent products will naturally converge. Rao believes they remain separate journeys. “As things stand today, it is on parallel tracks,” he says.

Within manufacturing, Bosch is steadily building intelligence across planning, procurement, logistics and production. A disruption reported by a supplier in the middle of the night can automatically trigger revised production schedules before engineers arrive the next morning.

Customer-facing products, however, are evolving under different constraints. Here the ambition is continuous improvement after purchase, enabling vehicles and industrial products to personalise behaviour, receive new capabilities and adapt to customer usage over time.

Rao expects the two worlds to converge eventually, but believes they are progressing at different levels of maturity today.

Governance is becoming the real enterprise challenge

The expansion of AI into operational technology has also changed Bosch’s security priorities. Manufacturing environments were traditionally isolated from external networks. AI, IoT and increasingly connected production systems have expanded that attack surface considerably.

Bosch’s response has been to isolate critical production environments inside highly segmented security zones, ensuring that a compromise within one environment cannot easily propagate across the wider manufacturing network.

Technical controls, however, address only part of the problem. Bosch has also adopted an AI code of ethics that governs how AI systems are deployed across the organisation. Its central principle is straightforward. Decisions affecting employees or customers must ultimately remain human decisions.

For Rao, that philosophy is becoming increasingly relevant as AI systems evolve from conversational assistants into autonomous agents capable of executing multi-step tasks independently. The technology may be becoming more capable, but the governance is becoming more difficult.

The CIO job is getting harder, not easier

Despite years of digital transformation, Rao says AI has made the CIO’s role more complex rather than simpler.

Traditional enterprise systems operated within predictable permission models. Large language models introduce a different challenge altogether. Access controls must account not only for what users are authorised to see, but also for what an AI system might infer, synthesise or inadvertently disclose.

Guardrails now extend beyond cybersecurity into judgement. They must ensure an AI assistant cannot reveal confidential employee information regardless of how the question is phrased. They must define who remains accountable when autonomous systems complete work on behalf of humans. And they must evolve as quickly as the technology itself.

“Last year, if we had this discussion, I would have questioned you only on GenAI,” Rao says. “Now GenAI has moved to agents.”

The next wave of enterprise AI, he believes, will not be judged by how many tasks machines can perform autonomously. It will be judged by how confidently organisations can govern them.

“The challenge,” Rao says, “is only increasing.”