What an AI sales agent actually does (and the two jobs it hands back to you)

An AI sales agent is software that does the top half of an outbound sales job. It finds accounts that match your profile, decides which are worth contacting, and writes the message. That's three of the five jobs a human sales development rep does. The other two, sending without damaging your email domain and handling replies, are the ones vendors say least about, and they're where most deployments go wrong.

So buy one for finding, qualifying and drafting, keep a person on sending and replies, and the maths can work. Buy one as a straight replacement for a rep, and you risk producing apologies at scale.

Outbound is five jobs, not one

Outbound isn't one activity. It's five, and they have very little in common. A tool that's excellent at the first isn't automatically any good at the fourth, but pricing pages tend to sell all five as a single seat.

Evaluating one gets much easier once you stop asking "is this AI SDR good?" and start asking "which of these five jobs is it good at, and what happens to the rest?"

Finding accounts: largely solved

Pulling a list of companies that match a set of attributes has been a solved problem for years. What's changed is that the filters are no longer limited to fields in a database.

An older tool could find you US software companies with 50 to 200 staff. A newer one can find companies whose engineering blog mentioned moving off a particular tool last quarter. That's a genuinely different kind of result, because the second is a reason to write and the first is a demographic.

This stage is worth paying for. It's repetitive, it's high volume, and getting it slightly wrong is cheap.

Qualifying: only as good as the evidence

This is where tools start to differ.

Qualifying means deciding that this particular company, this week, is worth a message. Doing it well needs evidence. Doing it badly only needs a confidence score, and confidence scores are easy to generate and hard to check.

The most common failure is worth describing precisely. Give a language model a thin record with empty fields, ask it why the account is a good fit, and it won't tell you the record is thin. It'll write something plausible. "They're scaling their engineering team" is a sentence a model can produce from a company name and a category alone, with nothing behind it.

That invented reason then becomes the first line of the email. The prospect reads a confident claim about their own company that isn't true, which is worse than sending nothing.

So the test for this stage is one question: can the tool show you the source for a qualified lead, as a quote and a link that still opens? It's the same evidence test you should already apply to your own leads. This is also why we built Revtive to reject any link that doesn't appear word for word in the pages it actually fetched, and why it often finds nothing. Across the 792 opportunity-finding runs our agents made between July and September 2026, 52% came back empty. A tool that never comes back empty is probably filling the gap with guesses.

Writing: solved, with one catch

Language models write a competent cold email, and that's no longer much of a differentiator.

The catch is that a competent email isn't what gets replies. What gets a reply is a first sentence that couldn't have been written about any other company, and that sentence depends on the evidence from the previous stage, not the prose in this one. Give the writer good evidence and even a modest model produces something that works. Give it nothing and the best model available produces fluent, well-organised, completely generic email.

So if the output reads as generic, the problem is usually two stages upstream, not in the writing.

Sending: where your domain reputation gets spent

This stage gets undersold, and it's where you can do damage that takes months to undo.

Volume is the whole appeal of automating outbound, and volume is exactly what mailbox providers watch. A tool that makes it easy to send two thousand emails a week makes it just as easy to wreck a domain that took years to build, and if that's your main company domain, your invoices and password resets can start landing in spam too.

What matters here is unglamorous:

• Send from a separate domain, never the one your company runs on.
• Authenticate it properly. Gmail requires SPF or DKIM for every sender, and SPF, DKIM and DMARC together once you send more than 5,000 messages a day to Gmail addresses (Google's sender guidelines spell out the rest).
• Keep volume per mailbox low, and add mailboxes rather than raising it.
• Keep an eye on reply rate as a health measure, not just a success measure. A sudden drop is often an early sign that messages have started going to spam.

None of this is specific to AI. Automation just removes the natural friction that used to limit the damage.

The reply: still yours

The reply is where the deal really starts, and it's the stage least suited to an unsupervised agent.

Most replies to cold outreach aren't "yes" or "no". They're "what is this?", "we already use something", "not now, try me in March", "who gave you my email?", and "I'm not the right person, but Sam is". Each needs a different response, several need a judgement call about whether to push, and one needs an apology.

An agent working that queue on its own does what it does with a thin record: it produces something plausible. At the reply stage, plausible and wrong costs you the account, not just one email. It's worth reading this alongside how a follow-up sequence should behave.

What should you ask an AI sales agent vendor?

Run these in the demo. The answers sort the category quickly.

The last question is the most revealing. A vendor who's never seen it go wrong probably hasn't had enough customers yet.

What it costs when it goes wrong

The obvious cost is a wasted subscription. The real costs are slower and worse.

A burned sending domain can be recovered over months, if it was a separate domain. If it wasn't, your company email becomes unreliable.

A market you've already annoyed doesn't recover at all. In a narrow niche, two thousand generic emails can reach a real share of everyone who might ever buy from you, and those people talk to each other.

And your numbers can mislead you. Activity goes up when you send more, but meetings don't necessarily follow. A team can run this for two quarters, watch the activity charts climb, and only notice at the pipeline review that very little of it turned into anything.

Who shouldn't buy an AI sales agent?

If you haven't written fifty cold emails by hand and had replies to some of them, do that first. An AI sales agent scales a process, and you can't tell whether the process works until you've run it slowly enough to watch.

If your whole market is under a thousand companies, automation doesn't have much room to help either. At that size, the limit isn't how many messages you can produce, it's how many of those thousand you can afford to get wrong.

And if you want one because outbound isn't working, the tool won't fix that. It'll do the same thing faster.

Something to try this week

Pick ten accounts you'd genuinely like as customers. If you can't name ten quickly, that's an ideal customer profile problem rather than a tooling one. For each, try to write one sentence naming something specific that happened there recently. Not "they're growing", but something with a date on it.

Count how many of the ten you managed. That's roughly what an AI sales agent will be automating. If it's two, the tool will find you two good messages and eight generic ones, and you'll know exactly what to test in a demo. Here are the five questions worth asking in it, and if the term itself is the confusing part, here's a plain definition of an AI SDR.