AI Agents at Work: The New Shape of the Office
How AI agents are entering real workplaces, why adoption is uneven, and what leaders should watch before scaling.
The first office computers did not replace the office. They changed what counted as office work.
Spreadsheets changed finance. Email changed coordination. Search changed memory. Cloud software changed where records lived. AI agents may do something similar to the small decisions and handoffs that fill the day.
The office is full of half-tasks. A meeting creates notes that need actions. A customer call creates a follow-up. A bug report creates a reproduction step. A sales conversation creates CRM updates. A policy change creates dozens of edits across documents and help pages. Moving context from one system to another can consume time that people would rather spend on the expert parts of their jobs.
Agents are aimed at that middle layer.
The first wave: assistant plus action
An assistant may answer questions and draft text. A workplace agent can also use tools to carry out steps in a workflow: look up a customer, read a ticket, check a policy, and prepare a response. What it can actually access or change depends on its integrations and permissions. It can still misunderstand the task, miss an exception, or act on stale information.
This is why companies talk about agents as a digital workforce. The underlying idea is concrete: software can take on parts of repeated workflows when the task, data, and permission model are clear. A software agent does not carry the accountability of an employee; a person or team still needs to own the outcome.
Salesforce introduced Agentforce for customer service, sales, commerce, and marketing tasks. Google described Agentspace as a way to search organizational knowledge and use agents across information silos. These are vendor descriptions of products, not independent evidence that a particular deployment will work. The shared ambition is to make organizational information actionable.
The adoption gap
The technology is moving faster than operating habits.
In McKinsey’s 2025 survey, 23% of respondents said their organizations were scaling an AI agent system in at least one business function, while another 39% reported experimentation. Nearly two-thirds said their organizations had not begun scaling AI across the enterprise. These are respondents’ reports from a survey, not a census or proof of successful deployments. Teams can make impressive prototypes; scaling requires data access, permission design, evaluation, change management, cost control, and leaders who can decide which workflows matter.
Agents do not fix messy operations by magic. They expose the mess.
If policies conflict, the agent will stumble. If the knowledge base is stale, the agent will repeat bad information. If nobody owns the process, the agent will automate confusion. If every useful action requires a credential nobody wants to grant, the agent will be trapped as a summarizer.
Where agents fit best
The best early workplace agents have several traits:
- The task happens often.
- Success is easy to inspect.
- The data is available.
- The cost of a mistake is limited.
- The agent can draft before it acts.
- A person can approve high-impact steps.
- The workflow has a clear owner.
Customer support triage, internal IT help, sales research, code maintenance, and document comparison can fit when scoped to a specific task. Compliance evidence gathering can fit if the sources are controlled and the review process is strict. None of these categories is automatically low-risk: a ticket may contain sensitive information, and a code change can affect production.
The weak cases are vague executive wishes. “Make us agentic” is not a workflow.
New jobs around agents
Agents will create work as well as absorb it.
Someone must define the workflow. Someone must decide tool permissions. Someone must write instructions that match policy. Someone must evaluate outputs. Someone must monitor costs. Someone must handle failures. Someone must improve the knowledge base that the agent depends on.
The responsibilities may sit with an existing team or with roles such as workflow owner, evaluation engineer, or knowledge steward. An agent with ongoing access can keep acting after its instructions or sources become outdated, so someone needs authority to pause it and correct the process.
Human skill does not disappear
McKinsey’s 2025 report on people, agents, and robots argued that future work will become a partnership between humans, agents, and robots, with many human skills enduring but being applied differently. That is a useful frame.
People will still need judgment, taste, accountability, negotiation, empathy, domain knowledge, and the ability to notice when the system is solving the wrong problem. Agents may take over more collection, comparison, formatting, first drafting, routing, and retrying. In a particular workflow, the human role may shift toward setting intent, reviewing results, handling exceptions, and maintaining relationships. Review itself takes time and requires enough subject knowledge to catch mistakes.
This will not happen evenly. Some roles will change quickly. Some will barely change. Some organizations will use agents to remove drudgery. Others will use them to create surveillance and pressure. The technology does not guarantee the labor model.
The leadership test
A serious agent program should be able to answer these questions:
- Which workflows are we changing first?
- Who owns each agent?
- What data may it access?
- What actions may it take alone?
- What actions require approval?
- How do we evaluate quality and safety?
- How do workers challenge or correct the agent?
- What happens to the time the agent saves?
The last question is the most human one. If saved time becomes only more volume, people will notice. If saved time becomes better service, deeper work, faster learning, or fewer late nights, they will notice that too.
A small pilot: support reply drafts
Suppose a support team wants help with routine delivery questions. The owner selects a defined set of tickets and lets the agent read only the fields and current delivery policy needed to draft a reply. The output includes the proposed answer, links to the policy and order information used, and any missing facts. A support worker checks the evidence, edits the draft, and sends the reply through the normal process.
Give the agent a restricted account and a review location that cannot send messages or trigger updates. A prompt saying “do not send” is not an access control. Confirm what the connected tools can actually do, where ticket data and logs go, and who may see them. Limit access to what the job needs, following the principle of least privilege . The permissions guide explains how to define and test those boundaries.
Tell the agent to stop and hand the case to a person when the policy is missing or conflicting, the order information is incomplete, or the customer asks for an exception. A retrieved ticket can also contain instructions aimed at the agent; customer text is evidence for the case, not authority to change the workflow. Keep consequential actions such as refunds and customer messages behind their own controls.
Decide whether the pilot helps
Before the trial, record how the team currently handles the same kind of work and agree on a review rubric. Sample routine cases and include difficult exceptions separately. Compare the agent’s proposed reply and cited evidence with the normal process; a different answer may reveal a human mistake as well as an agent error. Count failed and timed-out runs, not just completed drafts. The shadow mode guide describes how to make that comparison while keeping the agent out of the live action path.
Track a few measures that matter to the team:
- Quality: How many drafts are accurate and usable after review? Record serious errors and unsupported policy claims separately from minor edits.
- Time and workload: How long do preparation, review, correction, and escalation take in total, compared with the existing process? Ask reviewers whether the new queue makes their work easier or harder.
- Exceptions and cost: How often does the agent stop appropriately or miss an exception, and what do the tools and human review cost per completed case?
Record the number and type of cases, the review criteria, and failures. A small or convenient sample can identify problems but cannot establish a reliable rate for all tickets. Before the trial, set acceptable quality and risk thresholds, a review date, and who can stop it. If results fall short, revise the workflow or keep the work with people. If results are promising, expand the sample and continue monitoring before considering any additional authority.
For a reusable assignment, the delegation playbook offers a brief for the goal, context, boundaries, and proof. The aim is a workflow people can inspect and correct, with enough evidence to decide whether the agent is actually helping.