AI agents are in production. The business case often is not.

On 12 November 2026 Genpact published a survey of more than 2,000 senior executives at companies with revenues above $1 billion. The headline is striking. 79% have AI agents in production, but only 20% use them broadly and consistently across functions, and only 15% have fully tied their AI investment to profit and loss.

Most enterprises have deployed AI agents. Few can show the return.

AI agents in production 79%

AI initiatives that delivered the expected value 49%

Agents used broadly across functions 20%

AI investment fully tied to profit and loss 15%

The same report says that, on average, 49% of the AI initiatives funded in the past two years delivered the value expected of them. Review data points the same way. G2 reports that the average share of users who adopt an AI agent once it is bought has fallen from 68% in 2024 to 55% in 2026.

Why the value goes missing

Genpact's own conclusion is that the barrier is operational more than technological. That matches what we see when a business asks us to look at an AI project that has stalled. Three causes come up repeatedly.

• The project started from the tool. A team was given an agent and asked to find uses for it. Nobody owned a number that the agent was supposed to move.
• The process was never written down. AI was laid over a workflow full of unrecorded exceptions, so staff kept doing the work by hand to be safe.
• Nobody measured the starting point. Without a baseline for time, cost or error rate, there is nothing to compare the result against.

What the projects that pay back have in common

None of this requires a large programme. It requires discipline at the start.

1. Choose one process with a cost you can already state. Hours spent keying invoices, enquiries waiting more than a day, calls reviewed per week.
2. Record the baseline before building anything.
3. Decide where a person signs off. People trust a tool sooner when they can see what it did and correct it.
4. Review the number after 30 and 90 days, and stop or extend on that evidence.

It is also worth noting what the survey did not find. Among companies with agents in production, 92% kept their workforce the same size or grew it. The return is coming from capacity, not from headcount cuts.

How we approach it at NexAura

Our own UMAR product is a worked example. It was built for one process, call quality assurance, with one measure: calls reviewed for a given cost. QA managers still review the output. The result was roughly ten times the throughput at a similar operating cost, and it can be stated in a sentence because the target was set before the build.

We apply the same approach to client work. Every AI development project we take on starts with the process and the measure, and the technology is chosen afterwards.