Researchers have begun to test a question that has hovered over the AI industry for years, but with a new level of realism: not whether models can write convincing text or answer questions, but whether they can actually behave like office workers inside messy, multi-step workflows. In a recent experiment, AI “agents” were deployed to carry out typical tasks associated with office work. The headline result was both encouraging and sobering—these systems could complete some assignments effectively, yet they struggled with others, often in ways that matter in day-to-day operations.
The study’s framing is important. Many earlier demonstrations of AI in workplaces focused on narrow capabilities: drafting an email, summarizing a document, or generating a spreadsheet formula. Those are useful, but they don’t fully capture what makes office work difficult. Office tasks are rarely single actions. They involve context switching, interpreting ambiguous instructions, checking constraints, coordinating with other documents and systems, and knowing when something is “done enough” versus when it requires escalation. They also require reliability under time pressure—an area where even small failure rates can become operationally significant.
In this experiment, researchers treated AI agents as workers rather than as tools. The agents were given sets of tasks that resemble what employees might be asked to do—work that typically includes reading information, making decisions based on that information, producing outputs in the right format, and sometimes following up when something doesn’t go as expected. The goal wasn’t to see whether the agent could produce an answer at all. It was to see whether it could complete the job in a way that would hold up in a real office environment.
What emerged was a pattern that many workplace leaders will recognize immediately: AI agents are already capable of meaningful contributions, but their performance is uneven across task types. Some tasks were handled well enough to suggest that automation—or at least assisted automation—could be practical today. Other tasks were missed, incomplete, or unreliable, revealing gaps that go beyond raw language ability.
One of the most striking findings is that success depended heavily on how “structured” the task was. When instructions were clear, inputs were consistent, and the expected output could be validated against explicit criteria, agents performed better. In those cases, the system could follow a chain of reasoning, use available information, and generate outputs that matched the required format. This is where modern AI shines: it can process large amounts of text quickly, maintain internal state across steps, and produce coherent drafts that humans can review.
But office work isn’t only structured work. A large portion of it is semi-structured or ambiguous. Employees often receive requests that are incomplete, contain implicit assumptions, or require judgment calls. They may need to interpret what a requester really means, decide which details are relevant, and determine whether a response should include caveats. In the experiment, tasks that demanded that kind of judgment were more likely to fail. Sometimes the agent produced something plausible but not aligned with the underlying intent. Other times it simply didn’t complete the task as specified, either because it couldn’t identify the correct next step or because it lacked the operational grounding to verify what it had done.
This distinction—between producing text and completing work—may be the core lesson. An AI agent can generate a response that reads like it belongs in an email thread, but that doesn’t mean it has actually completed the workflow. Completion implies that the output satisfies the business requirement, that the steps taken were appropriate, and that the agent recognized when it needed more information. In the study, agents were more reliable when the workflow itself provided guardrails: explicit instructions, clear success metrics, and limited degrees of freedom.
Another factor influencing performance was the presence of “hidden work.” Office tasks often include steps that aren’t always written down. For example, an employee might need to check a prior conversation, confirm a policy, verify a number against a source of truth, or ensure that a document is formatted correctly for downstream systems. Humans do this automatically because they know where to look and what counts as evidence. AI agents, by contrast, must infer what to check and how to validate it. If the workflow doesn’t explicitly require verification—or if the agent doesn’t have access to the necessary tools or data—errors can slip through.
In the experiment, some failures appeared to stem from this mismatch between what the agent was asked to do and what the task implicitly required. When the agent lacked a reliable way to confirm facts or to cross-check outputs, it could still produce something that looked reasonable. But “reasonable” is not the same as “correct,” and in office settings, correctness is often the difference between a smooth day and a costly correction later.
There’s also the issue of escalation. In real workplaces, when something is uncertain, employees escalate. They ask clarifying questions, flag risks, or route tasks to specialists. The study suggests that agents did not consistently perform this escalation behavior. Instead, they sometimes continued with an assumption rather than stopping to request clarification. That’s a subtle but important difference. A human worker might say, “I’m not sure which version you want,” while an agent might proceed with the best guess. In low-stakes contexts, that might be acceptable. In high-stakes contexts—compliance, finance, legal, customer commitments—it can be unacceptable.
This leads to a broader insight about how AI agents should be deployed. The question isn’t only “Can the agent do the task?” It’s also “Can the agent recognize when it cannot do the task—and then behave appropriately?” Reliability isn’t just about accuracy; it’s about calibration. Agents need to know the boundary between what they can complete confidently and what requires human intervention.
The experiment’s results also highlight that not all office tasks are equally automatable. Some tasks are essentially information transformation problems: take input text, apply a rubric, produce a formatted output. These are well-suited to current agentic systems. Other tasks are coordination problems: they involve multiple stakeholders, evolving requirements, and dependencies on external systems. Even if an agent can draft a message, it may not be able to coordinate the surrounding steps reliably without deeper integration and robust feedback loops.
Consider the difference between drafting a summary and managing a project update. A summary can be evaluated by readability and coverage. A project update involves timing, status tracking, and alignment with what other team members have reported. If the agent doesn’t have access to the authoritative sources of status—or if those sources are inconsistent—the agent can easily drift away from reality. The study’s pattern of partial success fits this reality: agents can handle parts of the workflow, but the full end-to-end job often requires more than language generation.
Another unique angle in the experiment is the emphasis on “agent” behavior rather than model behavior. Modern AI systems can be used as chatbots, but agents are designed to take actions: to plan steps, call tools, retrieve information, and produce outputs in sequence. That shift introduces new failure modes. A chatbot can be wrong and still be “done.” An agent can be wrong and also take the wrong action—sending an incorrect email, updating the wrong record, or failing to follow a required procedure. Even when the final text looks fine, the path matters.
In the study, some tasks were missed or not performed reliably, which suggests that the agents’ planning and execution layers were not always robust. Planning failures can occur when the agent misinterprets the task structure or chooses an inefficient or incorrect sequence of steps. Execution failures can occur when the agent cannot access the right tools, encounters unexpected formats, or fails to maintain consistency across steps. These are not purely model-quality issues; they are system design issues. They point to the need for better tool integration, stronger validation mechanisms, and clearer interfaces between the agent and the workplace environment.
So what does this mean for workplaces right now? The most realistic takeaway is not that AI agents will replace office workers soon. It’s that they can already serve as partial automation engines—especially for tasks that are bounded, repeatable, and supported by clear inputs and measurable outputs. In other words, the near-term value is likely to come from augmentation rather than replacement.
But augmentation is not a vague promise. It can be operationalized. Organizations can start by identifying task categories where agents have demonstrated strength: drafting, summarizing, extracting structured information from documents, preparing first-pass responses, and generating checklists or templates. Then they can wrap those tasks in human review processes that are designed for speed and safety. Instead of reviewing everything from scratch, humans can focus on exceptions, factual verification, and judgment calls.
The study also implies that task selection will be a major determinant of success. If a workflow is too open-ended, too dependent on tacit knowledge, or too reliant on external verification, agents will struggle. That doesn’t mean the workflow can’t be automated; it means it needs redesign. Many organizations will need to make their processes more explicit—clarifying inputs, defining success criteria, and ensuring that the agent has access to the information it needs to validate its work.
This is where the “unique take” becomes practical: the future of office automation may depend less on making agents smarter and more on making work legible. Work legibility means converting informal processes into structured procedures that machines can follow. It means documenting decision rules, standardizing formats, and building systems that provide the agent with reliable signals. In that sense, the study is not only about AI capability; it’s about organizational readiness.
There’s also a governance dimension. If agents can complete some tasks but not all, then the risk profile changes. Partial automation can create a false sense of security. Teams might assume that because the agent succeeded on one task type, it will succeed broadly. The study’s uneven results argue against that assumption. Instead, organizations should treat agent deployment like introducing a new junior employee: useful, fast, and capable, but requiring supervision, training, and clear boundaries.
That supervision should include monitoring for failure patterns. If agents tend to miss certain steps, organizations can add checks. If agents tend to proceed without escalation, organizations can enforce “stop conditions” that trigger human review. If