A chip engineer’s day is mostly reading. Specifications, verification logs, somebody else’s design from four years ago, and the assumptions nobody wrote down. That is precisely the kind of work language models handle well, which is why nearly every semiconductor company started an AI pilot. Many of those pilots never progressed beyond the proof-of-concept stage. A review of 300 publicly disclosed enterprise AI deployments found that 95% produced no measurable effect on profit or loss. The technology performed in the demo. It did not survive the organization.
Ajay Kumar Govindaram is a Senior Solutions Architect at a leading cloud technology company and a Senior Member of the IEEE. His clients include Fortune 500 semiconductor manufacturers. In practice, he is often brought in after a chip company decides its engineers should use AI and then discovers no consensus exists on what that would actually involve. Since January 2025, he has served as the primary architect of a unified generative AI platform for one of the world’s largest semiconductor companies.
The Tool Sprawl Nobody Budgeted For
The failure pattern is consistent enough to be dull. A survey of 650 enterprise technology leaders in March 2026 found 78% running AI agent pilots and only 14% with anything scaled to organization-wide use. RAND put the broader failure rate at 80.3%, with roughly a third of projects abandoned before production and another quarter shipping but underdelivering. IBM researchers found that at least half of generative AI projects get dropped after the proof of concept. The models are rarely what breaks. What breaks is data scattered across departments and a cost curve nobody forecast.
Govindaram encountered a familiar pattern. Multiple engineering groups had already launched independent AI pilots using different vendors, architectures, and governance models. His assessment was straightforward: the company did not need another tool. It needed a foundation on which tools could reliably sit. He designed a unified generative AI platform built on managed foundation-model infrastructure, with a single governed path to the company’s proprietary codebases, design data, and internal knowledge bases, then let each team build its own use case on top of that.
“Everyone wants to argue about which model is best. That is the least interesting question in the room,” Govindaram says. “They already had several pilots running on different stacks. Not one of them could touch the design data safely, so none of them could scale beyond isolated pilots.”
Chip Engineering Has the Least Slack
The pressure to make engineers more productive lands harder in semiconductors than in most industries. Deloitte projects the sector will need more than 1 million additional skilled workers by 2030 and will grow over 80% across the same period. In the United States, the Semiconductor Industry Association estimates that 67,000 of roughly 115,000 new technical jobs risk going unfilled at current graduation rates. Europe expects to lose close to 30% of its semiconductor workforce to retirement by 2030, which takes decades of undocumented institutional knowledge with it. The work keeps getting harder while the people who know how to do it get scarcer.
The organization Govindaram advises employs more than 20,000 engineers, and the platform now carries a wide span of their work: code generation and review, design queries, hardware verification, technical documentation, and analysis that used to sit in a queue behind a data science team. Rather than building around a single flagship use case, he designed the platform so every team could bring its own workflow to the same governed foundation. As a result, the second use case cost a fraction of the first, and the tenth cost less than the second. That compounding is the whole argument for a platform.
“A point tool solves one problem and leaves you with the same integration work every time someone has a new idea,” Govindaram explains. “Build the foundation once and the marginal cost of the next use case collapses. That is not a technical preference. It is the difference between four use cases and forty.”
Adoption Depth Beats Tool Choice
What the evidence rewards is depth of adoption, not quality of procurement. Across more than 600 organizations McKinsey tracks, over 60% report at least a 25% productivity improvement from AI. The more useful number came from their research with OpenAI: companies where 80% to 100% of developers had actually adopted the tools recorded gains above 110%. Partial adoption returns partial results, and most enterprises never move past partial. The organizations that plateau are usually the ones that bought well and rolled out badly.
Govindaram moved the platform from proof of concept into production across 20,000 users in under 12 months, achieving a roughly 20% productivity gain across that population. AI solution deployment velocity improved 96% against the pre-platform baseline, and solution delivery improved 75%. Those results came from sequencing rather than mandate. He started with the teams whose workflows were closest to ready, got them working, and used what they produced to bring the next group over. The platform is now expanding toward the full 20,000.
“Adoption is not a rollout email. You earn each team,” Govindaram notes. “Pick the group most likely to succeed first, even if their use case is unglamorous, because their result is the only argument that moves the skeptical team next door.”
The IP Question Decides Everything
For a chip company, enthusiasm is rarely the constraint. DZone’s 2026 survey found 79% of organizations using large language models but only 22% at what the researchers call enterprise-integrated maturity, leaving 64% parked somewhere between a pilot and a working system. IBM traced most abandoned generative AI projects to poor data quality and weak risk controls, not to model performance. The gap between experimenting with AI and trusting it with proprietary engineering data is where most enterprise programs stall.
Semiconductor design data carries export-control exposure and represents the company’s core intellectual property. Govindaram therefore designed the platform around that constraint from the outset. Proprietary design data never leaves the company’s own cloud environment. Model invocations are logged, audited, and attributed back to the team and use case that triggered them. Foundation model selection excluded any model whose terms permitted data retention or training on customer inputs, which ruled out otherwise capable candidates. He then benchmarked the survivors against the company’s own semiconductor datasets instead of public leaderboards, on the reasoning that a general reasoning score tells you very little about whether it can correctly interpret a semiconductor timing report.
“A benchmark score is somebody else’s exam. It says nothing about your data,” Govindaram observes. “We disqualified capable models over their retention terms. In this industry, the leak you cannot rule out is worse than the capability you give up.”
What Comes After Assistance
The next step is systems that act rather than answer, and almost nobody has arrived. A 2026 Wakefield Research study of 1,000 senior technology and data leaders found that only 7% of enterprises had reached the operational stage where agentic AI delivers real business outcomes. 77% of executives said a fifth or less of their enterprise data was described well enough for agents to use it at all, and 78% found unifying data across business functions difficult. The bottleneck moved, but it did not move far. It is still the foundation.
This is where the architecture begins to pay a second dividend. Agentic workflows are already running in production on the same platform, handling incident detection and automated response across engineering systems. Govindaram presented that work at AWS re:Invent in December 2025. The agents did not require a new stack. They needed governed data access, logged invocations, and attribution, which the platform had from the beginning because the alternative was never going to clear a semiconductor company’s security review. Teams that skipped those layers are now rebuilding them under pressure.
“The unglamorous work you do in year one is what lets you say yes in year two,” Govindaram reflects. “Nobody gets promoted for access controls and audit logs. But when the company asks whether autonomous agents can touch production, you either already have that answer or you spend 18 months getting it — and by then somebody else has shipped;
