Unni Nambiar, CTO – Aeries Technology, drives AI and innovation through Global Capability Centers for PE-backed firms.getty​If you want to know whether a company is getting real value from AI, stop reading their press releases. Look at what is running in production, because that is where most of this falls apart.A lot of AI work looks impressive in demos, sounds compelling in presentations and creates genuine internal momentum. But when you plug it into real operations, it breaks, stalls or gets abandoned before anyone officially calls it a failure.MIT's NANDA Initiative found in 2025 that nearly 95% of enterprise generative AI pilots deliver little to no measurable P&L impact. ISG's State of Enterprise AI Adoption found that only 31% of prioritized AI use cases have reached full production. These numbers describe the dominant outcome, not the exception.Why Pilots Die Before They Reach Your Operations TeamThere is a pattern that shows up consistently enough to be called structural. Teams build pilots that work in controlled environments, show strong results on curated datasets and produce a story that holds together in a presentation. And then nothing happens. The model never gets embedded into a live workflow, or it gets deployed in a limited way and slowly stops being used. Check back 12 months later, and the organization is still "exploring AI."There is a more intentional version of this, too. Companies announce AI programs before the foundations are ready because it creates visibility and reassures stakeholders. But underneath, the fundamentals are missing. Data is fragmented, processes are loosely defined and ownership sits with the small group that built the thing. It looks like progress from the outside while changing nothing about how work actually gets done.When The Real Bottleneck Is Execution, Not AmbitionMost organizations already have an AI strategy. Where things fall apart is in execution, because AI lives inside the business, not on top of it. It depends on how clearly work is defined, how consistent the data is and how well systems hold together under real conditions. If those pieces are weak, no model compensates for that.Fixing those foundations is not glamorous work—documenting how work actually flows, cleaning fragmented data, defining ownership across teams. None of that registers as progress in a board update. So the pressure to show results fast pushes teams to skip it, build on unstable ground and then wonder why the pilot never scaled.Gartner's 2025 research predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. That is not a modeling problem but a data and process problem that existed long before AI arrived.What The Companies Getting This Right Actually DoWhen AI adoption works well, it does not look like most of what gets talked about.Take automated reconciliation in a financial operations team, for example. The visible part is the model. The actual work is the six months before the model, spent mapping exceptions, standardizing data feeds and getting three teams to agree on what a resolved record means. That groundwork is invisible from the outside and entirely responsible for why the model holds in production.The consistent pattern is that these organizations start with process clarity rather than technology selection. They understand how work really flows, where decisions get made, where things break and where exceptions show up.McKinsey's 2025 State of AI found that high performers, the roughly 6% of organizations where AI meaningfully contributes to EBIT, are nearly three times more likely to have fundamentally redesigned their workflows before selecting any technology. The sequence matters more than the sophistication of the tool.From Process Mapping To AgentificationIf there is one practical sequence that separates organizations making real progress from those stuck in pilot mode, it is this:1. Map the process as it actually runs. Not the version in your documentation, but the real one. Walk through it with the people who do the work every day. Identify where decisions happen, where exceptions pile up, where handoffs create delays and where someone's institutional memory is holding the whole thing together. This is the step most teams rush past, but without it, everything that follows is built on assumptions.2. Automate to augment. Layer AI into the existing human workflow so that the person doing the work is still the primary actor, but the repetitive, data-heavy and pattern-recognition tasks are handled by the system. The gains here can be significant.3. Flip the lens. Instead of asking how AI can help the human do their job, assume the AI agent is doing the entire process end-to-end, and then ask what it needs the human to do. That reframe changes everything, from automating around the human to designing the process around the agent, where the human's role is defined by what the agent genuinely cannot do.The Role Question Nobody Is AskingOnce you start building processes around what the agent can do, an uncomfortable question follows: What happens to the roles that used to own those processes? The answer is not elimination but redefinition. The person who used to run a reconciliation process end-to-end now handles exceptions the agent cannot resolve, validates edge cases and sets the rules the agent operates within. That is a fundamentally different job, and most organizations are not designing for it.AI's biggest impact is not at the task level. It is in how entire workflows get reshaped and how tasks are sequenced, grouped and handed off between humans and machines. Getting that right is an organizational design challenge, not a technology one. Most of the investment in AI right now is going into the visible parts. The durable value is being built in the work nobody talks about. That gap is where competitive advantage builds, and most organizations are not even looking for it there.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?