Most Enterprises Run AI Agents in Production, but Few Can Tie Them to P&L
Genpact's survey of 2,005 executives found 79% with agents live but only 15% tying AI investment to specific P&L lines. AT&T's account of nearly 500 production workflows shows what earned autonomy and governance look like in practice.
By The AI Enterprise Magazine Team4 min readAnalysis · Weekly briefing

Key takeaways
- Deployment is common. 79% report agents live in production, but only 20% report broad, consistent use across functions.
- Value lags. Only 15% fully tie AI investment to P&L lines, and initiatives delivered expected value 49% of the time on average.
- Redeployment is the barrier. 77% have a plan to redeploy freed capacity, but 46% say they cannot execute it.
- Earned autonomy. AT&T lets agents run alone only after accuracy reaches the mid- to high-90% range.
Week of November 16, 2026
Four in Five Enterprises Run AI Agents, but Only 15% Tie AI to P&L
In Genpact's survey of 2,005 executives, 79% report agents live in production, yet far fewer report broad use, lifecycle governance or AI value mapped to specific profit-and-loss lines.
View the data
| Share of respondents | |
|---|---|
| Have AI agents live in production | 79% |
| Use agents broadly and consistently across functions | 20% |
| Embed governance across the agent lifecycle | 17% |
| Fully tie AI investment to specific P&L lines | 15% |
| Have full-stack AI architecture at enterprise scale | 12% |
| Completed an operating model transformation | 9% |
Most large enterprises now run AI agents in production, but few can show where those agents show up in the financial results. That is the central finding of Autonomy in Practice, a Genpact survey of 2,005 executives published on November 12. Its figures, set alongside AT&T's account of running nearly 500 agentic workflows that was reported the same week, show that the hard part of agentic AI has shifted from building agents to governing them and proving what they are worth.
What did the Genpact survey find?
According to the Genpact report, 79% of respondents report AI agents live in production. That headline hides a wide range of depth. Only 20% describe broad and consistent use across functions, while 32% have agents live in some functions and 27% in isolated use cases. The rest are in pilots or planning.
The value side is thinner. Initiatives funded over the past 24 months delivered the business value originally expected in only 49% of cases on a weighted average. Just 15% of respondents have fully integrated their AI investments into specific profit-and-loss lines, which is the figure Genpact put in the headline of its press release. Only 9% have completed a significant operating model transformation, 12% report full-stack AI architecture capability at enterprise scale, and 17% embed governance across the full agent lifecycle.
The respondents work at companies with annual revenue from $1 billion to more than $50 billion, across 17 industries and 14 business functions. The results are self-reported and have not been independently verified.
Why is the gap between deployment and value so wide?
The report points to operating models more than technology. Genpact found that 77% of respondents say they have a defined approach for redeploying capacity freed up by AI, yet 46% name the inability to actually redeploy it as their leading barrier. An agent that saves hours has not created value until those hours are used for something else, and most organizations have not changed roles, budgets or targets to make that happen.
Governance follows the same pattern. Genpact reports that about half of respondents apply governance at deployment with only periodic monitoring afterward, and that 21% use real-time guardrails and continuous monitoring. Accountability for agent errors is split, with 26% holding the technology team responsible, 23% the business leader and 20% a dedicated, named agent owner.
The report's own Leaders segment, 17% of respondents, shows what closes the gap. Among Leaders, 49% fully integrate AI value metrics with P&L lines, compared with 7% of other enterprises, and 47% have governance embedded across the agent lifecycle, compared with 11%. Leaders are also more than twice as likely to have created net-new roles for human-agent operations. As Newsweek noted, Genpact added value realization as a criterion in its maturity index this year, which complicates comparisons with last year's results.
What does AT&T's approach add?
AT&T offers a working example of the controls the survey describes. Luke Corbin, assistant vice president of digital transformation and AI at AT&T Business, said at an AWS event in Boston that the company runs nearly 500 agentic workflows in production, handling more than one million transactions a month and about 45 billion AI tokens a day, Let's Data Science reported, summarizing SiliconANGLE. AT&T plans to triple that number of workflows next year.
The governance model is what stands out. Every production workflow starts with human oversight, and an agent runs autonomously only after its accuracy reaches the mid- to high-90% range within an acceptable risk profile, with continuous monitoring for drift afterward. Customer-facing workflows must go through AT&T's Business Agent Foundry and pass security, privacy and compliance reviews, while employees can build personal productivity agents on their own. An analysis by Rajesh Beri notes that AT&T's reported fivefold return is a company-stated figure that has not been audited, and Let's Data Science said the report does not explain how it was calculated.
What should data and AI leaders take from this?
The combined picture suggests three priorities for 2027 planning. The first is to measure value in the finance system, not in the AI program. If an agent fleet cannot be tied to a cost center or revenue line, it will be hard to defend once budgets tighten, and the survey suggests most organizations are in that position.
The second is to make autonomy something an agent earns. AT&T's practice of keeping a human in the loop until accuracy is proven, then monitoring continuously, is a concrete version of the lifecycle governance that only 17% of Genpact's respondents report.
The third is to plan for redeployment before deployment. The gap between the 77% with a stated approach and the 46% who cannot execute it is where expected returns disappear. Decisions about which roles change, which targets rise and which budgets shrink belong in the business case for each agent, not in a review afterward. Genpact's figures show how many organizations already have agents running. The harder task now is deciding what those agents are for and assigning ownership of the results.
Frequently asked questions
What percentage of enterprises have AI agents in production?
79% of the 2,005 executives in Genpact's Autonomy in Practice survey, published November 12, 2026, report AI agents live in production.
How many companies tie AI investment to P&L?
Only 15% of respondents have fully integrated AI investments into specific P&L lines, according to Genpact.
How does AT&T govern its AI agents?
Every production workflow starts with human oversight, and agents run autonomously only after accuracy reaches the mid- to high-90% range, with continuous monitoring for drift.
Sources
- Genpact: Autonomy in Practice
- Genpact: Autonomy in Practice: How agentic operations can turn AI progress into business value (PDF)
- Genpact via PR Newswire (Finviz): 79% of Enterprises Report AI Agents in Production, but Only 15% Have Fully Tied AI Investment to P&L Outcomes
- Newsweek: Genpact Study Finds the Real AI Challenge Begins After Deployment
- Let's Data Science: AT&T runs nearly 500 agentic AI workflows in production
- The Daily Brief (Rajesh Beri): AT&T Keeps Human Review on Nearly 500 AI Workflows
Filed under: AI Strategy & ROI, Agentic AI, AI Governance & Trust
Become a Sponsor







