Your sales agent is prospecting. Who checks its work?

On Friday, your new agent sent 240 emails. On Monday, your sales director finds 37 messages to review, eight replies to handle and a meeting with a company you cannot serve. Prospecting is running. But who saved any time?

The question has become timely. In October, Microsoft opened a public preview of an agent that researches prospects, personalises outreach and manages follow-ups. That is a real change. It tells us little about what happens inside a sales team after the send.

The work may not disappear. It may move desks.

I build and use sales automations myself. I know the pleasure of watching a long task run without me. That is also why I distrust the sent-email count. It is flattering, immediate and only loosely connected to selling.

Consider a company selling management software to service groups. Its agent finds accounts, drafts messages and follows up. In the demo, the seller saves two hours of research. In practice, they check job titles, remove existing customers, correct a misplaced reference and rewrite replies that deserve a real conversation. Some of the saved time returns as a less visible, more urgent job.

That is not necessarily a failure. A message to the wrong finance director, or a promise about a feature you do not have, can cost more than twenty good messages earn. Some actions should be checked because they commit your reputation or your commercial terms. The point is to count the checking.

In its October 2026 report, G2 says 31% of surveyed B2B buyers spend at least six hours a week overseeing agents. That number does not describe your company. It does remind us that supervision is real work, and belongs in the calculation.

Do not approve every email

The instinctive answer is to make a seller approve every message before it goes out. It feels safe. It can also turn a seller into the proofreader of a machine that writes faster than they do.

I would distinguish three situations. Account research and draft preparation are easy to sample-check. A first message to a named strategic account deserves a human read: the context, recipient and proposition matter. A reply about pricing, security, timing or product capability belongs with someone authorised to commit the company.

Those are not universal autonomy levels. They are places where the cost of an error changes. A team selling a standard service to hundreds of small accounts will draw the line differently from one pursuing fifteen tenders a year.

For example, an agent can prepare a follow-up after a missed meeting. If the buyer replies, “We have moved this to January; finance blocked the budget,” it should not continue with three polite reminders. It has found commercial information. A person now needs to decide what to ask, whom to ask and when.

Measure accepted conversations, not machine output

I would start with one month and one segment, not the whole contact database. Choose accounts where your sellers already know a good prospect when they see one. Otherwise you are testing your targeting and the agent at the same time, and a poor result will teach you little.

Each week, look at four things: replies that open a real conversation; meetings the seller accepts after reading the context; time spent reviewing, correcting and taking over; errors that could have damaged a relationship. Compare with how you prospect without an agent on similar accounts. Not to claim a perfect experiment. To avoid calling a moved queue “productivity.”

A booked meeting is not necessarily an opportunity. If the buyer is a poor fit, nobody understands the problem, or the seller declines the meeting, the calendar filled but the pipeline did not move. Ten serious exchanges may be worth more than a hundred automatic replies. Your sales cycle decides that, not the vendor's enthusiasm.

If review takes too long, do not ask only for a better prompt. First inspect the account list, exclusions, what evidence the agent may use, and the situations where it must stop. Often the missing piece is a sales rule nobody wrote down because a human already knew it.

Sales leadership keeps the wheel

The point of a sales agent is not to replace human prospecting indiscriminately. It is to remove an identifiable share of research, preparation and follow-up without inventing a hidden supervision job.

The question for a sales meeting is not “how many emails did it send?” but “which useful conversations would not have happened without it, and how many hours did we spend getting them?” If nobody can answer, it is too early to raise the volume.

I like building these systems. I like them even more when the sales director can close the dashboard and get back to customers. If you are testing an agent in B2B prospecting, let's look at one specific workflow: I will tell you where I would let it act and where I would take back control.