How to evaluate AI SDR tools (five questions a demo will not answer)

Evaluating AI SDR tools comes down to five questions, and a scripted demo only answers the first two: can it find accounts that fit, and can it write a clean sentence? Nearly every vendor can do both now. The three that decide whether you'll still be paying in six months are about sources, replies, and whose email domain takes the damage when it gets something wrong.

A quick disclosure before going further: we make one of these, so read this the way you'd read any vendor writing about its own category. What we won't do is end with a ranked list that puts us first. Most of the lists currently ranking for "AI SDR tools" are published by companies selling one, and it's rare for the publisher not to come out on top.

If you're still working out what the category even is, start with the definition and come back.

What should you ask an AI SDR tool before you buy it?

1. Can it find accounts that fit? Every tool can. It's a filtered search against a data provider, and many vendors draw on the same handful of providers. It tells you very little.

2. Does it write a clean sentence? Also solved. Modern models write fluently, and fluency was never why cold email failed.

3. Can it show a source for every claim? This is where tools start to separate. Ask for one drafted message, then ask where each factual claim in it came from. You want a link that opens and a date. What you often get instead is a confidence score. The evidence standard should be the same one you'd hold a human researcher to.

4. What happens when someone replies? Demos end when the email goes out. Real outbound starts at the reply, and replies are rarely a clean yes or no. They're "we looked at something like this in March, what's different?", or a question about a compliance requirement. Ask to see how it handles a real thread, not the happy path.

5. Whose domain suffers when it's wrong? If it sends from your domain, your sender reputation is the collateral for its mistakes. That can be a fair trade if the mistakes are rare, but you should know it's the trade you're making.

Questions one and two get answered in a twenty-minute call. Three, four and five need your own accounts and a couple of weeks, which is the real argument for a trial over four more demos.

Why a confidence score isn't a source

A score tells you how sure a system is, not whether it's right, and it can only be as good as what it measures. We ran into this in our own product. In August 2026 we found a workspace holding a Reddit thread that was more than five years old, rated high-impact, because the agent had scored how well the thread matched the product and nothing else. A perfect match to a conversation started in 2021 isn't an opportunity. We now cap anything older than eight months at low priority, but the lesson generalises: a score hides whatever it forgot to measure, and a source with a date doesn't.

So when a vendor shows you a score, ask what goes into it, then ask to open the sources behind it.

Test what it does when there's nothing to find

Here's a question most demos never touch: what does the tool do when there genuinely isn't a good prospect or conversation to surface?

Across the 792 opportunity-finding runs Revtive's agents made between July and September 2026, 52% came back with nothing that cleared the bar. Some of that is our own quality gates doing their job, and some is simply that most markets don't produce a fresh, relevant conversation every few hours. The point is that an empty result is often the correct result. A tool that always finds something, every time you run it, is padding, and you'll be the one reading the padding. So when you trial one, point it at a narrow segment where you know there's very little happening and see whether it admits that.

What it costs to run, not just to buy

The quote covers the licence. Three other costs tend to show up later.

Enrichment and data is usually billed per record the tool tries to resolve, including the attempts that fail, so a list of 5,000 costs whatever the number of attempts turns out to be, not 5,000.

Sending infrastructure means extra mailboxes and domains, plus the weeks it takes to warm them up. Anyone sending real volume spreads it across several domains, because putting it all on your main one is how you lose your main one.

Human review is the cost nobody quotes, and it's often the biggest. Someone has to read the drafts before they go out. If nobody does, you haven't bought an AI SDR, you've bought a faster way to be wrong.

Ask for all four numbers before comparing two vendors on the first one.

A two-week trial that tells you more than four demos

Run it on a separate domain, and keep the volume flat. Raising volume during a quality test tells you nothing about quality, and something you won't enjoy learning about deliverability.

The number to watch is how many drafts you'd send without editing. Our rule of thumb is that if it's under about a third, you've bought a drafting assistant, which may still be worth it, but price it as one. It's also the number no vendor publishes.

Where these tools earn their keep

In the reading, more than the sending. Someone researching an account properly spends most of their time checking that what a list claimed is still true. That's where the hour goes in prospecting, and it's the part a model can speed up most, because reading forty pages to find the two facts that matter is what models are good at. A tool that hands you fifteen well-sourced accounts with a reason for each has done the expensive part.

After that, the claims get looser. It's worth being precise about which stages of outbound are solved and which are still yours.

Who shouldn't buy an AI SDR tool?

If your messages don't work when you send them by hand, an AI SDR won't fix that. These tools multiply an existing process, so ten manual messages a week that get nothing become a hundred automated ones that get nothing, spread across four domains.

If you've got fewer than about fifty target accounts, the research is a morning a week and the tool is mostly overhead. Doing it yourself also means you keep the knowledge.

If nobody on the team can spare time to review drafts, it'll struggle. Review isn't optional, and a tool bought to save time that creates a daily reading task tends not to survive its first renewal.

And if the real problem is that nobody's heard of you, cold outreach has a low ceiling however good the research is. Showing up where your buyers already ask questions does more for the reply rate of a cold message than any tool.

Revtive sits in the research stage on purpose. It finds and drafts, it has to cite a source for every signal it reports, and it never sends anything without someone pressing the button. You can see what it does and doesn't do without booking a call.