AI agents are past the pilot stage. According to a Genpact survey published on 12 November, 79 percent of large enterprises now have AI agents running in production. The second number is the one that matters: only 15 percent have fully tied their AI investment to P&L outcomes. The technology is inside the companies; the measurable return, in most cases, is not. That lesson applies to a 20-person business just as much as to a billion-dollar one.
The news in brief: deploying is easy, running is hard
Genpact’s report, “Autonomy in Practice”, draws on more than 2,000 executives across 17 industries and 14 business functions at companies with revenues above 1 billion dollars. The key figures:
- 79 percent have AI agents in production, but only 20 percent use them broadly and consistently across functions.
- 49 percent of AI initiatives funded in the past 24 months delivered the expected value, meaning roughly every second project fell short.
- Just 9 percent have completed a significant operating-model transformation.
- AI now leads or fully runs 40 percent of decisions, more than double a year earlier, yet humans still decide or audit in at least 87 percent of cases at every decision stage.
- 92 percent of companies with agents in production kept their human workforce the same size or grew it.
The report’s central line: enterprises have become effective at deploying AI agents; running them is the hard part. Two caveats are worth keeping in view: the findings are self-reported, and Genpact itself sells agentic operations services.
Why this matters for SMBs
On the surface this is an enterprise story. In practice, SMBs fall into the same trap at a smaller scale: they buy a chatbot or an automation tool, switch it on, and three months later nobody can say how much money it made or saved. The 49 percent “expected value” figure shows the problem is not the technology but deployment without process redesign and without measurement.
SMBs do have one real advantage: a short decision chain. What only 9 percent of large companies manage, actually changing how work gets done, a small business can decide and execute within a week. The second lesson is the 87 percent: a human in the loop is not a weakness, it is standard equipment in systems that work. Anyone promising “full autonomy” is promising something even the enterprises in this survey do not practise.
Practical steps
- One process, one number. Give every AI agent a single business metric before launch: calls answered, appointments booked, cases closed, hours of manual work saved.
- Connect it to the P&L. Translate the metric into money: what a missed call costs, what a manually processed order costs. That is the difference between the 15 percent and everyone else.
- Redesign the process, don’t just bolt on the tool. If the agent answers the call but your team still types the details into the CRM by hand, half the value is lost.
- Design the human checkpoint. Decide which decisions the agent makes alone and which it hands to a person, and track how many cases get handed over.
- Start with one function, then expand. Broad, consistent use (the 20 percent) does not arrive all at once; the second and third process build on what the first one taught you.
- Review monthly. Look at the numbers every 30 days and switch off what is not delivering. It is the fastest way to stay on the right side of the 49 percent.
Where to start
The cleanest, most measurable entry point is inbound call handling. An AI Voice Agent logs every call: how many came in, how many it resolved alone, how many it handed to a person, and how many appointments or quote requests it generated. That is exactly the metric 85 percent of companies in the Genpact survey have not connected to their results. If you want to see what this looks like on your own call volume, book a consultation. The numbers start talking in week two.
Source: Genpact / PR Newswire – “79% of Enterprises Report AI Agents in Production, but Only 15% Have Fully Tied AI Investment to P&L Outcomes”, 12 November 2026.