Finding Work an Agent Can Improve

The useful question is not whether an agent could replace a customer-service representative. It is whether there is a particular stretch of work that repeatedly slows that representative down—and whether an agent could handle part of it safely.

That distinction matters. A job title bundles together many different activities: listening, interpreting, deciding, writing, checking, coordinating, and taking responsibility. Some activities are repetitive; others depend on trust, judgment, or a sensitivity to context that is hard to reduce to instructions. An agent may be useful for one slice of the job without being a substitute for the person doing it.

Look, instead, for friction: work that consumes time because information is scattered, a process has many handoffs, or someone must keep nudging the next step along. Friction is not proof that an agent is the answer. It is a clue that a workflow is worth examining.

In the small company handling customer inquiries, an employee might read a message, identify the product, search a policy document, check the customer’s order, draft a reply, and remember to follow up if the issue remains unresolved. The message itself may take a minute to answer. The surrounding work—finding the right facts and carrying the case across several steps—may take much longer. That surrounding work is often where the opportunity lies.

Repetition with variation

A useful starting heuristic is that a promising task may be neither perfectly fixed nor entirely novel. If every case follows exactly the same rule, ordinary automation may be simpler and more reliable. If every case calls for a fresh judgment with serious consequences, a person should probably remain in charge. Between those poles lies repetitive work with manageable variation: familiar patterns, but enough differences that a rigid script breaks down. These characteristics help identify work worth examining; they do not, by themselves, predict that an agent will succeed in deployment.

Consider sales. A team receives inquiries from prospective customers in several formats. Someone must identify what each prospect wants, gather relevant company information, and prepare a useful first response. An agent might organize the incoming details and draft a tailored message for a salesperson to review. It need not decide that the prospect is worth pursuing, promise a discount, or send a consequential offer on its own. The task is not “do sales”; it is “reduce the time spent preparing for a human conversation.”

Customer service offers another example. An agent might collect order details, compare a complaint with the company’s return policy, and suggest a response. Routine cases could be handled under clear rules; unusual complaints, safety concerns, or angry customers could go to staff. The value comes partly from speed, but also from making sure the necessary information reaches the person who can act.

Follow the handoffs

Some work is slow not because any one step is difficult, but because the work travels between people and systems. A service request may wait in an inbox while someone looks up an account, then sit again while another team checks availability. An agent could assemble the relevant facts, create a concise case summary, and route it to the right team. It might also notice that a promised update is overdue and prepare a follow-up.

Operations teams face similar friction when tracking delayed shipments, reconciling discrepancies, or chasing missing documents. Researchers may spend hours gathering information from approved sources and organizing it for comparison. In knowledge work, an agent might collect background material for a meeting, summarize open questions, and draft a briefing from documents a team has authorized it to use. In each case, the human work remains important: confirming the facts, resolving exceptions, and deciding what the information means.

A useful clue is the sentence, “I can do my actual work once I’ve…” Once I’ve found the latest version. Once I’ve copied the details into the right place. Once I’ve heard back from three people. Once I’ve remembered to follow up. The “once I’ve” work is not always suitable for an agent, but it often reveals hidden costs in a process.

Find the slice, not the profession

To assess an opportunity, describe the work as a sequence rather than a job: What arrives? What information must be gathered? Which steps repeat? Where does judgment enter? What happens when the usual path fails? This makes it easier to see whether an agent could take on a bounded segment while a person retains the decisions that matter.

The aim is not to maximize the number of tasks handed over. It is to reduce avoidable delay and busywork without making the process harder to trust. In the customer-inquiry workflow, an agent may be valuable not because it can act like an employee, but because it can reliably gather context, prepare the next step, and keep a routine case moving—while recognizing when the case is no longer routine.

That is a more practical place to begin than asking which jobs AI will change. Start with the work that gets in the way of the work.

Once you have found friction in a workflow, the next question is not automatically, “Could an agent do this?” It is, “What is the simplest kind of help that would solve the problem?” An agent can be powerful precisely because it can make choices and take actions—but those abilities also add complexity, uncertainty, and ways to go wrong. Sometimes the sensible design is to give the work to a person, use a fixed rule, or ask AI for advice without letting it act.

Imagine the small company receives a message asking for its opening hours. The answer is on the company website, and it is the same every day. A rule can recognize a common request and send the approved hours. If the company’s hours change, someone updates the rule. There is little value in asking an agent to interpret the request, search several systems, decide what to do, and send a reply. That is like hiring a detective to read a clock.

This is the natural territory of fixed automation: work with stable inputs, clear rules, and predictable outcomes. It can be a scheduled reminder, a form routed to the right team, or a confirmation sent when an order is received. Because the steps do not need to change from case to case, a conventional rule is often easier to test and maintain, depending on the process and the organization’s tools. It may also behave more consistently than a system asked to infer what the rule probably ought to be; that advantage is not universal and still depends on how the rule is designed and maintained.

But the same company may receive a message like, “The replacement part you sent doesn’t fit, and we need the machine running by Friday.” Now the request contains a complaint, a deadline, and a possible product issue. The customer’s order history and the company’s replacement policy may matter. A fixed rule can still handle narrow pieces—such as logging the message or alerting the service team—but the whole situation does not fit neatly into a single if-this-then-that path.

Even here, an agent may not be the best first move. If a support representative needs to understand the case, weigh the options, and decide what to promise, but does not need AI to change an order or contact the customer, an AI assistant may be enough. It can summarize the message and relevant history, find the applicable policy, and draft a response for the representative to review. The human retains the decision and the act of sending.

That distinction matters. A request for judgment is not, by itself, a request for delegation. AI can help someone think without being allowed to do anything outside the conversation. An assistant is often the right fit when the work benefits from synthesis or drafting, while a person must still choose what is appropriate. For example, a sales manager might ask AI to compare meeting notes with a proposal and flag unanswered questions. The system contributes useful analysis; the manager decides what to say to the client.

Manual work also remains a real option. If the inquiry is rare, sensitive, or unusually consequential, adding an AI system may create more checking than it saves. A person may already be able to handle the case quickly, especially when its significance depends on a relationship or a subtle piece of history that is not captured in the company’s records. “Can be automated” and “should be automated” are different claims.

A useful test is to match the shape of the work to the kind of help:

• If the steps are stable and the exceptions are few, consider a rule.
• If the work requires judgment or synthesis but not independent action, consider an assistant.
• If the work involves several changing steps across tools—and the system can check progress and stop or escalate when needed—an agent may be worth considering.
• If the stakes are high, the situation is unusual, or the human review would erase the time saved, keep the work with a person.

These are not rigid categories. A good workflow can combine them. The company might automatically acknowledge every inquiry, have an assistant prepare a summary, and let an agent gather order details for routine replacement requests. A human could approve any refund, exception, or promise about a deadline. The point is not to choose one technology for the whole process; it is to give each part only the capabilities it needs.

The extra capability of an agent is most valuable when the path cannot be fully scripted but still has a clear destination. If the route is obvious, a rule is simpler. If someone needs help reaching a judgment, an assistant may be safer. If the task is too consequential or too personal to delegate responsibly, a person should remain in charge. The best tool is not the one that can do the most. It is the one that adds the least new complexity while making the work meaningfully better. That means matching action limits to the consequences of error and deciding in advance what should make the system pause, ask, or hand the work back.

A customer writes to the small company that sells commercial coffee equipment: “The replacement connector for our machine doesn’t fit, and we need it running before our new location opens on the 18th. Can you help?”

The message looks simple. It is not. The company must identify the order, work out which connector the customer needs, check whether it is available, determine whether the requested timing is realistic, and respond without promising more than it can deliver. The work crosses several small decisions—and several places where a confident guess could become an expensive apology.

A small delegation, with a clear finish line

Suppose the agent can read the inquiry, search the order system and product guide, check inventory and shipping estimates, and draft a response. It may also update the customer-service record and, once the required facts are verified, send routine follow-ups under an explicit approval policy. Those are meaningful tools, but they do not grant unlimited authority. The agent cannot issue refunds, change shipping charges, or promise an exception without approval.

It first uses the email address to locate a possible order, then checks the order number and customer details before using the record. An address match alone does not establish identity or authority to change an order. The order record shows a particular machine model. The product guide identifies two compatible connector types, while the order contains no note about the customer’s setup. The agent has found relevant facts, but not enough to choose between the connectors. A weak system might treat the most common option as certain. A careful one recognizes the missing information.

It can still make progress. It checks the inventory and shipping estimate, then drafts a concise reply asking for the machine’s serial number and a clear photo of the connector and connection point. It explains that the company can identify the right part once it has that detail, and gives a realistic estimate for delivery if the part is in stock. Because the opening date is close—and the exact part remains unknown—the agent leaves the reply for a person to review rather than sending it automatically.

A customer-service employee checks the draft, removes an unnecessary sentence, and sends it. The agent records what it found, the question asked, and the reason it paused. When the customer replies, a system equipped to inspect images can use the photo alongside the serial number and approved compatibility records; otherwise it routes the photo to a specialist. A visual resemblance alone does not establish fit. If compatibility is verified and standard shipping meets the customer’s timeline, it can send the permitted routine follow-up. If the part is unavailable, the delivery estimate is uncertain, or the customer asks for a special arrangement, it stops and hands the case back.

That is not a dramatic display of artificial intelligence. It is a practical division of labor: the agent gathers and connects information, handles routine steps, and keeps the case moving; a person handles ambiguity or commitments outside the agent’s authority.

Let the agent move the case, not the goalposts

The inquiry has a clear destination: identify the right part and give the customer an accurate next step. The agent can save an employee from searching several screens and composing the same basic explanation again. Its value is not merely faster typing; it is reducing the small delays that let an inquiry sit unanswered.

The finish line should be observable. The inquiry has been matched to an order; the relevant product and delivery information has been checked; the customer has received a response or the case has been escalated; and the record says what happened. A message that sounds polished is not, by itself, proof that the work is done.

The pause points matter just as much as the routine path. Missing information, conflicting records, uncertain delivery, a request for an exception, or signs of an unhappy customer can all change what a sensible response requires. The agent should not hide those uncertainties behind fluent prose. It should surface them, preserve the work already completed, and make the next human decision clear.

In a successful follow-up, the customer does not need to know which steps were handled by software. They notice something more ordinary and more valuable: the company understood the question, did not invent an answer, and replied with a useful next step. An action is not an outcome: a reply sent or a record updated matters only if it helps resolve the customer’s problem accurately. That is the standard to aim for—not maximum independence, but dependable progress within sensible limits.