What can AI agents genuinely do in sales today — and what can’t they?

What sales work do agents genuinely do well?

Everything repetitive and evidence-based. Enriching a lead from public sources, ranking a list by likelihood of converting, writing a first-touch email that reflects what the prospect does, following up on a schedule, and keeping records accurate. This is most of the hours in a sales week.

Ranking deserves particular mention because it is where trained models outperform language models decisively. A language model can write persuasively about which leads look promising. A model trained on outcomes can rank four hundred of them and attach a confidence figure that means something. Those are different capabilities and they are frequently conflated.

Follow-up is the other underrated one. Most outbound fails because nobody sends the third message, and an agent has no feelings about sending the third message.

What can they not do?

Anything requiring a commitment. They cannot decide what to charge, agree terms, judge whether a badly-fitting customer is worth taking, or read a room. They also cannot tell you that a deal is dead when the prospect is still politely replying, which is a specific and expensive human skill.

They are also poor at the thing sales people call qualification. An agent can check whether a lead matches criteria you defined. It cannot decide that this particular customer will be a nightmare to serve despite matching every criterion, which is a judgment built from experience nobody wrote down.

And they should not be allowed to discount. A discount is a pricing decision, and pricing is the clearest example of something that should always stop for a person. On newc0 it does, by default and with a threshold you set.

Do AI SDRs actually book meetings?

Yes, at rates that depend far more on your list and offer than on the tool. Agents make a mediocre offer fail faster and at greater volume. If outbound was not working manually, automating it produces the same result with more people annoyed and a worse sending reputation.

This is the least commercially convenient thing on this page and it is the most useful. Before automating outbound, send fifty emails yourself. If none of them work, the problem is the offer, and no agent fixes an offer.

There is also a compliance dimension that gets skipped. Sending automated outbound to people who have not opted in is a legal question in most markets, not a tooling question, and it is the sender’s responsibility regardless of what sent it.

What about phone calls?

Voice agents work for narrow, transactional calls — confirming appointments, qualifying against fixed criteria, chasing a documented outstanding item. They do poorly at anything that becomes a conversation, and in most markets there are disclosure rules about telling the person they are speaking to software.

The technical quality is no longer the limiting factor. The limiting factor is that a phone call is the channel people choose when something is wrong or complicated, which is precisely the situation agents handle worst. A voice agent answering your main line intercepts your most important conversations.

Outbound voice deserves particular care. Calling people who did not ask to be called is regulated differently from emailing them, the rules vary by market, and responsibility sits with the business rather than the vendor whose software dialled.

A reasonable position for a small company: use voice for confirmations and reminders where the call is expected, and let a person handle anything inbound that is not routine. newc0 does not currently document a voice capability, and this site does not claim one.

What this looks like in practice

In a newc0 company, the sales function runs continuously: leads found and enriched, scored by a trained lead-scoring model, outbound written and sent to opted-in contacts, replies chased, calls booked. What arrives in the morning is a short list — the leads worth calling today, and any discount or commitment that needs a person.

The agent workforce page sets out what each function runs and what it never does without approval.