The Agentic ROI Reset: Why Autonomy Is Becoming the Defining Measure of AI Value

Most enterprise AI spend today is being measured the wrong way. Companies are grading their AI on accuracy and adoption, the same scorecard used three years ago, while the technology itself has quietly moved from assisting decisions to executing them. The result is a growing number of organizations that can point to impressive pilots and rising budgets, but not to a business process that actually works differently. 

The pattern shows up everywhere. Most large companies have models running somewhere, whether for document summarization, ticket routing or anomaly detection, and McKinsey estimates generative AI could add $2.6 to $4.4 trillion a year in economic value. But ask a leadership team what AI has changed about how they operate, and the answer tends to be careful and non-committal. Teams have demos, often impressive ones, yet BCG reports that nearly three quarters of organizations have yet to move AI value past the pilot stage.The first wave of enterprise AI mostly improved judgment at specific points in a workflow. This next wave is starting to take over parts of the workflow itself. 

Aviation offers a useful comparison. Early cockpit automation helped pilots calculate routes and hold altitude, useful tools that still left the pilot flying the plane. Modern aircraft work differently: autopilot manages long stretches of a flight on its own, while the pilot monitors and steps in only when conditions demand it.The job remained. What it requires changed entirely. Enterprise AI has crossed a similar line, though most organizations have not noticed because they are still grading their systems on how well they assist people rather than on how much work those systems complete independently. 

For years, enterprise AI functioned as analytical support. A model predicted churn or flagged a risk, and people did the rest, interpreting the output, deciding what to do, and pushing the process forward. Agentic systems work differently. They hold context across steps, coordinate between applications, and keep a workflow moving until something unusual surfaces, carrying decisions through into action. 

A mid-size insurer processing 30,000 claims a month makes the difference concrete. Under the older model, AI might classify an incoming document as a motor claim or flag it for possible fraud, while the adjuster still did the heavy lifting: pulling the policy record, cross-checking coverage terms, calling the garage for a repair estimate, requesting a police report if needed, and finally keying the recommendation into a separate approval system. An agentic system handles that entire chain. It extracts the claim data from the submission, matches it against the policy, pings the repair network for an estimate, flags if a police report is required, and stages the file for approval, so that routine claims clear without anyone opening them. That gap matters more than any accuracy improvement. 

Most companies still evaluate AI using the same metrics they relied on three years ago: accuracy, latency and adoption. None of these indicate whether the organization actually runs differently. All three can improve while the business stays exactly the same. RAND reports that, by some estimates, more than 80 percent of AI projects fail, twice the failure rate of IT projects that do not involve AI. The model itself is rarely the bottleneck. The bottleneck is everything wrapped around it. A model scoring 95 percent on a test set does not matter if the workflow still requires six handoffs, two email approvals and a manual entry into a legacy system before anything reaches the customer. 

Once AI begins executing work rather than merely informing it, the metrics have to change. Nobody evaluates autopilot by counting how often the pilot grabbed the controls. The test is whether the flight completed reliably. The same logic applies inside the enterprise. What matters now is workflow completion rate, how exceptions get routed, the cognitive load on the humans still in the loop, and cycle time from request to resolution. The relevant question becomes whether the process reaches a reliable outcome with less human intervention. 

Building the model is rarely the hard part. Redesigning a claims workflow so the system can run it end to end, defining where autonomy should stop, and wiring it into a policy administration system built in 2007 that nobody wants to touch, that is the hard part. Most organizations stall before any of that happens. Teams get through the pilot and even the demo, and then the system never becomes part of the operating workflow itself. 

In the companies that get this right, AI starts to look less like a tool and more like infrastructure. The question becomes whether the organization has rearranged itself to let the system run. In one insurer case the authors reviewed, the average claim cycle fell from eleven days to under three. The model was unchanged. The gain came from removing four approval steps that had existed only because nobody had questioned them in a decade. 

Autopilot did not take pilots out of the cockpit. It changed what the job required. Agentic AI is doing the same thing to claims processing, procurement and customer operations, and to any process where a routine decision currently bottlenecks on a person waiting to click approve. The companies that benefit most will be the ones that redesign how work gets done around these systems, and the quality of the model will rarely be what separates them. The most useful question to ask of any AI initiative is how far the workflow runs before a person has to step in.