The Pilot Trap Most Enterprises Never Escape
Gartner research cited in this conversation puts the problem in stark relief: 85 percent of organizations have adopted AI in some form, but only five percent of those initiatives make it to production. By 2027, Deloitte projects that 40 percent of even that tiny surviving group will fail. Hemang Upadhyay, a product and AI leader with 16 years of enterprise experience, sat down with Robin on episode 265 of the Localization Fireside Chat to explain why that number is not a technology story. It is a leadership alignment story. The teams that get stuck, Hemang argues, are the ones that start by selecting a model rather than by defining the problem they are actually trying to solve. They run a controlled pilot on 10 simulated queries, get excited in the demo room, and push to production before anyone has asked who owns the outcome if something breaks at scale. The moment real customers arrive with real language, the system starts to crack, and the project quietly disappears.
What Measurable Governance Actually Looks Like
Hemang draws a sharp line between activity metrics and outcome metrics, and that distinction is where this conversation gets practically useful. A localization company reporting that its AI model generated thousands of words per second is measuring activity. The questions that matter are whether turnaround time actually improved, whether human correction effort decreased, whether terminology consistency held across markets, and whether cost per completed project moved in the right direction. He organizes measurement into three layers: quality, which covers accuracy, relevance, safety, and cultural appropriateness; operational performance, which asks whether time and cost improved without creating hidden rework; and business impact, which looks at retention, conversion, and customer satisfaction. The point he keeps returning to is that faster translation is not a success if critical errors are climbing, and lower service costs are not a win if customer trust is quietly eroding. Both sides of the equation need an owner before the project scales, not after something goes wrong. He is equally direct about governance structure: legal, security, and cross-functional stakeholders need to be in the room before the pilot launches, not invited to review the wreckage afterward.
The Independent Voice That Enterprise Work Cannot Give You
One of the more honest exchanges in the episode happens when Robin asks Hemang why he chose to build a public platform at hemangai.com alongside a demanding enterprise career. Hemang’s answer lands cleanly: staying inside one organization means staying bound to that organization’s problems, and you stop learning from the friction that only comes from talking to people who see the world differently. The connections he made through publishing his work led to a live TV panel in India, a recent TED Talk focused on AI’s role in education, and the open source community he now runs to help small and medium businesses access AI at a cost they can actually absorb. Robin draws a parallel from his own experience, describing a weekend email outreach tool he built himself using Codex after IT told him the request would take years to fulfill. The open rate beat every CRM output the team had produced. Both of them are making the same argument from different angles: the tools are accessible enough now that waiting for institutional permission is a choice, not a requirement, and the people who are learning by doing are building a compounding advantage that the people waiting for the approved process simply are not.
If you are working through an AI initiative right now and wondering why it feels like it is stalling somewhere between the demo and the real world, this episode will give you a concrete diagnostic framework to work with. Hemang’s core recommendation before you scale anything is to get every stakeholder, the data owner, the business owner, the technology owner, and the risk and legal team, into a room and surface the disagreements before you build. If those people cannot agree on what success looks like, that misalignment is the first problem to solve, and no additional AI demonstration will fix it. You can Watch on YouTube or Listen on Simplecast and choose whichever format fits your workflow best.
Robin Ayoub 
Leave a comment