# AI Agents in the Flow of Work: New Research | Asana > New Asana research of 4,502 workers finds coordination overhead, agent overlap, and misaligned deployment slow AI's productivity gains at enterprise scale. Source: https://asana.com/inside-asana/ai-agents-flow-of-work-which-agents-you-use ## AI agents are in the flow of work. Whether you find value depends on which agents you use. _New research from Asana finds that employees report coordination demands, overlapping agent work and gaps in the help they need most — highlighting what organizations need to solve to turn agent activity into value._ AI agents are becoming a regular part of work. But look closely at the numbers, and a more complicated story emerges about what it takes for them to deliver on their productivity promise. We undertook new research, in partnership with Censuswide and 4,502 CIOs, AI leaders, and AI-using office workers in the US and UK. The highlights show three challenges: - Employees are carrying the extensive coordination burden - Agents are doubling up on work already done - Agent work isn't supplying what employees actually need. For leaders, the challenge goes beyond deploying agents. It's putting agents to work on the problems employees most need help solving — and redesigning the flow of work to make that happen. ### **Agents are active. But are they doing the right work?** The research finds a wide gap in agent usage between leaders and workers. Nearly seven in ten leaders (68%) use three or more agents in a typical week, compared with 36% of workers. Leaders report an average of 4.6 agents per week; workers, 2.6. That gap matters because agent deployment decisions typically flow from leadership — but the workflows those agents touch are run by employees. When you ask workers directly whether agents are being deployed where they're most needed, the picture shifts. Only 23% of surveyed AI-using office workers say the agents being introduced address the tasks and problems they most need help with. A further 33% say agents address some important needs but leave significant gaps, while around one in five say they mostly address lower-priority or convenient tasks. This isn't a verdict on the technology. Enterprise work — vendor negotiations, cross-functional campaigns, compliance reviews — has qualities that differ from what personal AI tools are built for. It spans teams, decisions and systems. It depends on institutional context: what the organization has decided before, what it paid, what worked. Agents need that context, a clear role in the process and permission to act, with people retaining control over decisions that require judgment. ### **A new kind of overhead** Among workers who use at least two agents in a typical week, 46% estimate that switching between tools, checking which agent to use, managing outputs, or resolving conflicting results takes more than 10% of their working time. Coordination isn't free. When multiple agents operate in parallel without shared context or clear ownership, someone — usually a person — has to reconcile the results. As people work with more agents, that coordination only becomes more complex. Overlapping work adds another layer. Among workers at organizations that use two or more agents, 38% estimate that more than a quarter of agent-completed work is duplicated, repeated, or substantially overlaps work by another agent. Overlap isn't always unnecessary — some redundancy is deliberate and useful. But these estimates raise questions that every organization scaling agent use needs to answer: Who owns each part of the work? Which output should the team use? How does the next person — or the next agent — know what has already been done and what needs to happen next? ### **Visibility matters as much as autonomy** Then there's trust in what agents do beyond workers' direct view. More than four in five leaders (81%) say they would be comfortable allowing an agent to access information and complete actions without reviewing every step. Fewer than two-thirds of workers (65%) say the same. The gap is more interesting than it first appears. Workers who work alongside agents day-to-day are more likely to notice when an agent acts on incomplete context, surfaces conflicting outputs, or escalates something that needed human judgment. Their caution is calibrated, not uninformed. The specific concerns workers raise reflect this. 30% worry that agents may access sensitive or confidential information without people realizing. 25% worry that access permissions could expand without sufficient review. Another 25% worry that an agent may act on information it should not have accessed. Security and privacy are also among the most commonly cited barriers to scaling agents, selected by 32% of leaders and 31% of workers. What workers are describing is a need for clarity, not resistance to AI. They want to know what information an agent can access, what actions it can take, who can step in, and who is ultimately accountable. That isn't an argument against agent autonomy — it's an argument for making the boundaries legible inside the workflow itself. ### **Starting from outcomes, not agent counts** The data doesn't establish a single cause of enterprise AI's growing pains, and it doesn't prescribe what every organization should do. But taken together, the findings suggest a useful reframe: rather than asking how many agents to deploy, start by asking what work needs to reach completion — and design from there. That means identifying specific workflows with clear owners and defined outcomes. Mapping where human judgment is irreplaceable and where an agent can act. Ensuring that context — historical decisions, approved policies, data from connected systems — travels with the work, rather than staying locked in a separate conversation window. It also means measuring the right things. Faster individual outputs are useful. But the more revealing indicators are: whether coordination overhead went down, whether work reaches completion without rework, and whether the people involved can actually see what happened and why. This is the opportunity behind orchestrated execution: getting the right work to the right person, agent or system, with the context and authority to carry it forward. People set direction and exercise judgment. Agents take on work within clear boundaries. Their contributions move a shared goal closer to completion. At Asana's [Work Innovation Summit in New York](https://workinnovationsummit.asana.com/event/nyc/summary) today, enterprise leaders, AI practitioners, and customers are exploring how to put that approach into practice — and build human-agent teams that deliver on AI's productivity promise. ### **Research methodology** Research conducted by Censuswide, September 2026, among 4,502 respondents in the UK and US: 1,501 CIOs and AI leaders and 3,001 office workers who use AI at companies with more than 50 employees. The coordination finding uses a base of workers who use at least two agents weekly (n=1,809); the overlap finding uses a base of workers who report at least two agents in their organization (n=2,071). Percentages are rounded. Leader and worker samples are not matched by employer. The research reflects respondents' reported experiences, estimates, and attitudes — it does not measure changes over time or establish causation. - [How we cut a browser agent's cost 76x and made it 5x faster by keeping its cache intact](/inside-asana/cut-browsers-agent-cost) Engineering Artificial Intelligence (AI) #### CTO of StackAI A study with GPT-6 Astra in Codex across four models, from investigating the code to measuring the resultsFigure 1. Humans direct GPT-6 Astra in Codex, which logs its work in Comm ... - [Deliver great client work at scale with Asana Client Management, now GA](/inside-asana/asana-client-management) Artificial Intelligence (AI) Clients stay informed, teams stay focused, and leaders always know where every account stands.The client relationship is everything to a professional services organization. Busine ... - [Command by Asana, the planning and product development system for builders, now available](/inside-asana/command-by-asana) Artificial Intelligence (AI) Engineers stay in flow, releases hit their dates, and leaders always know where to look for the answer without a meeting. Coding agents changed how software gets built, but the sy ... - [Streamline your service operations in one platform with Asana Service Management, now Generally Available](/inside-asana/asana-service-management) Artificial Intelligence (AI) Answer the quick questions instantly, route the ones that need human approvals to the right teams, and coordinate the complex, cross-functional work, all in the same system.Legacy ... - [AI agents are in the flow of work. Whether you find value depends on which agents you use.](/inside-asana/ai-agents-flow-of-work-which-agents-you-use) Artificial Intelligence (AI) - [Content writer](/author/team-asana) New research from Asana finds that employees report coordination demands, overlapping agent work and gaps in the help they need most — highlighting what organizations need to solv ... - [Artificial Intelligence (AI)](/inside-asana/artificial-intelligence)