Four jobs it does genuinely well: drafting, summarising, enriching and ranking. The three it is sold hardest for are the three where a confident wrong answer does real damage, and those are the ones that get abandoned by month four.
The useful split is not which model or which vendor. It is whether a human sees the output before it matters.
The four that work
Drafting. A first version of an email, a follow-up, a proposal section. The win is speed, not quality, and that is a real win. The rep edits it, which is exactly the right shape: the error gets caught.
Summarising. Call notes, a thread, a long account history. This is the most undersold use and the one with the least downside, because a summary that misses something is a smaller problem than a call nobody wrote up at all, which is the realistic alternative.
Enriching. Filling in a company size, a sector, a technology, a recent announcement, from public sources. Verifiable, and wrong in obvious ways rather than subtle ones.
Ranking. Putting a list in a sensible order. The caveat matters here and it is in the next section.
The three that do not
| Sold for | What goes wrong |
|---|---|
| Fully automated outreach | Generated personalisation that is obviously generated reads worse than three honest sentences, and it burns the domain as well as the list |
| Forecasting | A confident number with no visible reasoning gets trusted until it is badly wrong, then the whole system is distrusted |
| Unexplainable lead scoring | A rep calls the 92, it is a student. They call the 88, it is a competitor. By month three they are working their own list again |
The scoring one is worth dwelling on because it is the most common. A model that is 70% accurate and arguable beats one that is 85% accurate and opaque, because the first gets used and improved and the second gets quietly abandoned without anyone announcing it. That argument is made properly in can you read your own lead scoring model.
The question to ask any vendor
What happens when it is wrong?
If a person reviews the output before it reaches a customer or a decision, you are fine and the accuracy claim matters less than the vendor thinks. If the output goes straight out, or straight into a number a leader acts on, then accuracy matters enormously and should be tested on your own data rather than the demo workspace.
Almost nobody asks this, and it separates the products worth buying from the ones that demo beautifully.
The thing nobody tells you about AI SDRs
The category is real and the honest version of it is narrower than the marketing. An AI SDR is good at the volume end: researching, drafting, sequencing, handling the first reply. It is poor at the moment where a human would notice that this particular company is not actually a fit and stop.
The teams getting value run it as a hybrid rather than a replacement, which is covered in AI SDR vs human reps and what actually works with AI SDRs in 2026.
Where this leaves the stack
Most teams do not need an AI tool. They need the tools they already have to be joined up, because the single biggest source of bad AI output in sales is incomplete records, and no model fixes a CRM that three people update from memory on a Friday.
Empiraa Signal takes the position that AI belongs inside the system holding the record rather than bolted beside it, with ANI built into the platform rather than sold separately. The readable scoring model is part of the same argument: a score a rep can argue with is a score they will use.
Related reading: AI SDR tools, what the category actually is, and can you read your own lead scoring model.
