A global survey of 937 business and IT leaders published today finds half of respondents (50%) work for organizations that already have 10 or more artificial intelligence (AI) agents in production.
Conducted by International Data Corp (IDC) on behalf of Amazon Web Services (AWS), the survey also finds only 3% work for organizations that are scaling AI agents across multiple departments. A total of 46% said more accurate and trusted results are needed for broader agentic AI adoption within their organization.
More than half (55%) also said there is a need for improved user experience, while 41% report concerns over unpredictable costs and return on investment (ROI) are holding back additional budget allocations. More than half (51%) also expressed concerns about data privacy and a total of 61% said they are encountering some type of technical limitation.
Nevertheless, nearly two thirds (65%) are not expecting to fully complete their roll out of agentic AI applications until 2027, compared to 23% that plan to complete within the next 12 months. A total of 65% said their organizations are prioritizing investments in additional agentic tools.
While it’s clear that organizations will broadly adopt AI agents as the underlying technologies rapidly evolve, organizations are also carefully evaluating specific use cases for deciding to deploy them, says Jeffrey Hammond, worldwide product strategist for independent software vendors (ISVs) at AWS.
For example, 61% of respondents said they are comfortable allowing AI agents to complete data analysis and reporting tasks autonomously, but only 42% said they would allow AI agents to autonomously provide customer support. A total of 60% are deploying agents in operations functions, compared to 40% that are deploying them across marketing/sales, customer service and IT functions. Only 2% are embedding agents into their own products, the survey finds.
However, 41% said their organization does use collaborative multi-agent systems that share intelligence and coordinate in real time.
A full 67% also recognize there will be a need for more training to fully harness agentic AI, with the lack of skilled personnel (55%) cited as the top implementation challenge.
Specific top challenges organizations adopting AI agents are encountering are observability (52%), followed by ensuring that output generated is accurate and trustworthy (46%) and multi-agent interoperability (30%). “There’s a need for more telemetry data,” says Hammond.
Top investment areas for the next 12 months, meanwhile, are tools (65%), data strategy and integration (62%), and infrastructure and compute (47%). Nearly two thirds (63%) are testing and deploying AI agents in public cloud infrastructure, with 22% using multiple foundational models and large language models (LLMs).
However, the survey suggests more organizations are interested in customizing AI agents rather than building their own from the ground up. A full 91% said there is a need for industry/domain customization, with 62% of organizations planning to buy pre-built agents that they will then customize.
A total of 44% also noted their organization is already, however, mixing custom-built and purchased agents to complete tasks, with 37% having deployed multi-agent solutions. In most of those instances, organizations are chaining together multiple agents to sequentially complete a more complex task, notes Hammond.
Less clear is to what degree the rise of AI agents might spur a massive wave of business process re-engineering. Organizations are clearly breaking processes down into a smaller set of tasks that can be completed by an AI agent. Those outputs are then being chained together to drive a larger process, says Hammond. How often organizations might use that opportunity to re-engineer those processes remains to be seen.
In the meantime, however, organizations are aggressively evaluating AI agents. The issue is that many of the processes that organizations rely on today are deterministic in the sense they need to be done the same way every time. AI agents, however, are based on large language models (LLMs) that are probabilistic, which means they almost never complete the same task the same way. As a result, organizations need to better understand exactly where in any given workflow an AI agent can, with a significant amount of context engineering, be relied on to automate a workflow, especially if, for example, there is a tendency to report that a task has been completed when in fact it has not.
There is, of course, no putting the proverbial agentic AI genie back in the bottle but as more organizations are exposed to these advances they are also starting to have a greater appreciation of both its potential and limitations.
