Your Best AI Asset Has Been Working Here for Years | Opinion

A big summer storm was rolling in. The local utility’s AI dispatch system pinpointed the county with the highest forecast wind speed as the utility’s greatest exposure. The model recommended stationing crews at a regional depot there to be ready.

A 24-year veteran line supervisor disagreed.

He focused on a service area two counties away, an easily overlooked narrow stretch with old feeder lines beneath mature trees, a substation near a flood-prone creek, a major hospital, and a recently closed bridge on the major access route—a fact not reflected in the outage-management data. So, he overrode the recommendation. His more measured plan restored power to the critical circuit hours ahead of the algorithm’s plan.

This scenario illustrates the most pressing challenge now for AI: Without a deep, nuanced understanding of how things work in the real world, even the best AI model will be functionally hamstrung. Which is why so many AI pilots and proofs of concept don’t graduate to the real world.

In one global study, the percentage of companies abandoning most of their AI initiatives before they reach production surged from 17 percent to 42 percent year over year, with organizations on average reporting that 46 percent of projects are scrapped between proof of concept and broad adoption. Boards are looking for returns on their investments and coming up short.

This is happening even though the technology is getting better, faster, cheaper, and more widely available. Today, a company can test powerful models without building its own from scratch. Yet getting from a promising demo to a fully functioning solution that people can use, trust, and rely on remains elusive.

That’s because an AI system needs more than just data. It needs to understand the work—not what the work is, but how, in the reality of day-to-day operations, it actually gets done. The system needs an understanding of the semantics of the business and how work is prioritized, how work gets done and how failures are managed.

For example, an AI system needs to know how a bank settles a payment when a routine instruction turns out to have a sanctions implication or why a claims professional treats one discrepancy as a clerical issue and another as a warning sign. It needs to know which maintenance alert can wait until morning, and which one signals a safety risk. Those distinctions rarely show up in a process map. They live in the experience of people who have handled the exceptions, seen the consequences, and learned what the formal rules can’t convey.

The next phase of AI depends on tapping into that expertise. Successful AI will bring a company’s most experienced people into the AI work early—not to learn how to use a new tool, but as partners in deciding what the AI system should know, what it should do, and where it should stop.

That begins with finding the exceptions. Most workflows look orderly on paper. In practice, they contain judgment calls, workarounds, customer history, regulatory requirements, and conditions that change by season, geography, product, or circumstance. A model trained only on the standard path or simple variations may perform admirably—right up to the moment when the standard path no longer applies. Also, buried among routine training data are a few exception patterns, making it difficult for models to learn from those exceptions.

It also means measuring what matters. A model may reduce the time needed to complete a task, but speed alone tells little about business value. The right measures might include avoided losses, fewer repeat calls, faster restoration of a critical service, improved compliance, or more time for skilled people to focus on difficult cases. The measures must be holistic beyond direct productivity and recognize and reflect the peculiarities of business the company is actually in.

Finally, organizations need to decide how many errors a process can absorb. Recommending a product to a shopper and releasing a high-value payment are both forms of automation. They should not be treated in the same way. Experienced employees can help determine when AI should suggest, when it can take action, and when a person must remain responsible for the decision. The final accountability always lies with a human.

We’ve seen this play out in practice. Working with a U.S. specialty insurer, leveraging Wipro Intelligence, we built an AI system to help underwriters process the dense, inconsistent documents that accompany complex commercial-risk submissions: inspection reports, binders, forms, and risk-evaluation reports.

The technology could extract more than 100 relevant data points and identify anomalies, but the real work was defining what those signals meant. Experienced underwriters helped distinguish a harmless inconsistency from a material exposure, identified the narrative details that should trigger review, and set clear boundaries for when the system could accelerate a decision—and when it had to hand the case back to a person. The result was not an AI underwriter. It was a faster underwriter: handling time fell from an hour to minutes, while human judgment remained where it mattered most.

For CEOs, the message is clear. AI can’t be treated as a technology project handed off to the technology team. It’s a business redesign project which needs to be owned at the top. Start with a workflow where a better decision or a faster resolution has clear value; put the people who know that work best beside the people building the technology; give them a shared definition of success; and match the model to the actual task at hand based on the knowledge of your operational experts. Not all workflows require the biggest, most sophisticated models.

The companies that make AI work will be the ones that recognize where their most valuable AI asset already resides—in the heads of the people who know how things really get done.