Summary
- Genpact surveyed 2,005 senior enterprise executives, with 79% reporting AI agents in production but only 20% using them broadly across business functions.
- Respondents estimated that 49% of AI initiatives funded over the previous two years had delivered expected value, while just 15% fully linked AI investment to profit-and-loss outcomes.
- Only 9% reported significant operating model transformation, suggesting that deploying agents and reorganising how businesses work remain separate stages.
Artificial intelligence agents have entered production across most large businesses surveyed by Genpact, but comparatively few organisations have established a clear connection between their AI investments and measurable financial performance.
The professional services company’s Autonomy in Practice research, based on responses from 2,005 senior enterprise executives, found that 79% reported AI agents operating in production. However, only 20% said agents were being used broadly and consistently across business functions.
Respondents estimated that, on average, 49% of AI initiatives funded during the previous 24 months had delivered the business value originally expected. Just 15% reported fully connecting AI investment and value measures to specific profit-and-loss outcomes, suggesting that the technology is being deployed more quickly than organisations are establishing how to account for its commercial contribution.
Published on 12 November, the research covers businesses with annual revenues ranging from $1 billion to more than $50 billion. Its findings describe the assessments of senior executives rather than independently audited returns achieved through individual AI projects.
Production use has moved ahead of wider integration
With 79% of respondents reporting agents in production, AI deployment has progressed beyond isolated experimentation within many large businesses. Such systems are being introduced into established activities, although the extent of their operational responsibility varies considerably.
The smaller proportion reporting consistent use across business functions indicates that production deployments remain concentrated in particular areas. An organisation may have several functioning agents without using them across departments or connecting their activities into a coordinated process.
That distinction can affect the benefits achieved. Automating part of a finance workflow may reduce the time required to prepare information, but the process may still depend on manual checks, approvals, and transfers between different applications.
Where those surrounding activities remain unchanged, the wider organisation may gain less capacity than the improvement in an individual task initially suggests. Employees can complete one stage more quickly while waiting for information or authorisation elsewhere.
The challenge becomes more complicated when agents operate across systems that were originally designed around separate departments. Data definitions, approval procedures, and access rights may differ, requiring integration work before information can move reliably between automated processes.
Genpact found that only 9% of respondents reported completing significant operating model transformation associated with AI. The result suggests that changing how organisations allocate work and responsibilities has progressed more slowly than introducing software agents.
Those findings do not prove that a particular organisational redesign will deliver better financial results. They establish a gap between the extent of reported deployment and the smaller proportion of businesses saying they have made substantial changes to their operating structures.
Financial outcomes remain difficult to attribute
Genpact’s finding that 49% of funded initiatives had delivered their expected value requires careful interpretation. The figure represents the average proportion of initiatives that respondents believed met their objectives, not a 49% return on investment or an independently calculated success rate across all enterprise AI spending.
Different organisations may also have set very different expectations. One project could be judged successful because it reduced document processing time, while another might be evaluated against sales performance, service quality, or the cost of delivering a business process.
Operational improvements do not always translate directly into financial outcomes. An agent may reduce the time employees spend on a recurring task without lowering payroll expenditure, particularly if the company retains staff and reallocates their capacity to other activities.
Alternatively, automation might allow a business to complete more work with its existing workforce, creating value through increased capacity rather than direct cost reduction. Such a benefit requires measurement against demand, service output, and other commercial factors before its financial contribution can be established.
Only 15% of respondents said AI value was fully connected to specific profit-and-loss lines. The finding indicates that most surveyed organisations had not completed the work needed to attribute their reported gains directly to financial results.
Doing so can require cooperation between finance, operations, and technology teams. Technical staff may measure accuracy, processing time, and system availability, while finance departments need to establish how those changes affect costs, revenue, or profitability.
These assessments must also account for the continuing expense of model usage, infrastructure, software integration, monitoring, and employee supervision. Measuring the cost of developing an application without its operating expenditure can produce an incomplete assessment of its returns.
Human oversight remains part of deployment
The survey found that AI now leads or fully performs 40% of decisions across eight stages of decision-making examined by Genpact. The figure concerns the activities measured within the research and should not be generalised to every decision made by businesses.
Despite the growth in automated activity, humans still decide or audit in at least 87% of cases at every stage, according to respondents. The findings indicate that operational autonomy frequently involves a combination of automated actions and employee oversight.
The appropriate division of responsibility can differ according to the consequences of an individual decision. A system recommending how to prioritise routine administrative work presents different risks from one changing contractual terms, financial records, or customer access to a service.
Employees may therefore remain responsible for approving consequential actions, reviewing exceptions, and correcting outputs that do not meet established requirements.
Genpact also found that 92% of respondents with AI agents in production had maintained or increased their human workforce, while 69% reported increasing the number of agents used during the previous year.
These results should not be treated as proof that AI prevents redundancies or creates additional employment. Organisations may increase or reduce staffing for reasons unrelated to their automation programmes, and the survey does not isolate those effects.
However, the findings show that expanding the use of agents has not universally coincided with a reduction in employee headcount among the participating businesses.
Implementation involves connected organisational changes
Genpact identifies several requirements associated with more advanced AI adoption, including redesigning processes, defining responsibilities between employees and agents, establishing reusable data and technology foundations, governing automated systems, and connecting their results to financial outcomes.
These recommendations reflect the company’s interpretation of its research and its professional services experience. They should not be presented as independently validated evidence that every organisation adopting the same approach will achieve equivalent results.
Nevertheless, the implementation requirements are relevant to enterprises operating complex software environments. Agents may need access to information held in different applications, while their actions must comply with existing permissions, controls, and business procedures.
The accuracy of underlying records can affect whether an automated recommendation is suitable. Where data is incomplete or inconsistent, a system may produce a technically coherent response that does not reflect the organisation’s actual operating conditions.
Governance must also continue after initial deployment because software, business requirements, and the information available to agents can change. An application that performs reliably during a limited pilot may encounter different conditions when used more widely across departments.
Global findings have limits for individual markets
Genpact surveyed executives across 17 industries and 14 primary business functions, supplementing the quantitative research with interviews involving enterprise and technology leaders.
The study concerns large businesses and does not provide a representative assessment of smaller organisations. Its headline findings should also not be treated as specific measurements of AI adoption or financial performance within the UK or European markets individually.
Responses depend on executives’ understanding of terms such as production deployment, expected value, and operating model transformation. These definitions can vary between organisations, particularly where agents perform different levels of independent activity.
Although the results describe reported experience rather than audited performance, they indicate that large enterprises have moved further in introducing AI agents than in demonstrating their contribution to financial results.
The continuing expansion of AI deployment will place greater demands on organisations to measure both operating costs and commercial outcomes. Genpact’s findings establish the extent of that measurement gap among surveyed executives, while the returns achieved by individual businesses will depend on the processes changed and the evidence used to assess them.












