An AI SDR is software that does the mechanical parts of outbound prospecting: build a list from a described ideal customer, enrich it, write a personalised sequence, send it, and handle the routine replies.
That is a real product doing real work. The question worth asking is not whether it works, it is which parts of the job it covers, because the human role it is priced against is broader than the software.
The job, split
| Task | Software | Reality |
|---|---|---|
| Define the ICP | No | You do this. Wrong here, everything downstream is wrong |
| Build the list | Yes | Match rates vary hugely by segment and geography |
| Enrich contacts | Yes | Genuinely good, and the clearest win |
| Write the sequence | Yes | Competent, rarely distinctive |
| Personalise the opener | Yes | Beats a template, loses to a rep who did the reading |
| Send at volume | Yes | The strongest capability by far |
| Maintain deliverability | Partly | Tooling helps; the offer decides |
| Handle a routine reply | Yes | Scheduling, basic questions |
| Handle an off-script reply | No | Escalates to a person |
| Decide the offer | No | Nothing in this category touches it |
Read down the No column and you have the answer to why results vary so widely between companies running identical software. Three things decide whether outbound works: who you target, what you offer them, and whether your domain survives. That is one human decision, one human decision, and a shared responsibility.
Where it genuinely beats a human
Volume and consistency, and it is not close. Software does not have a bad week, does not ramp for 3.9 months, and does not leave after 17.6. Against the structural problem identified in part one, that 22% of a rep's tenure is spent below full output, repeatedly — software genuinely does not have that failure mode.
It is also strictly better at enrichment and at the long tail. A rep prioritising 200 accounts will do proper research on the top 20 and send templates to the rest. Software applies the same mediocre-but-consistent personalisation to all 200, which is worse than the top 20 and considerably better than the other 180.
Where it does not
The reply that goes sideways. A prospect who responds with something the sequence did not anticipate, a competitor mention, a restructure, a question about a specific integration — is the moment the deal is won or lost, and every product in this category escalates it. That escalation quality is the thing to test in a trial, and almost nobody does, because trials get evaluated on copy quality instead.
How to compare the cost honestly
Part one established a human at roughly $960 to $1,100 per held meeting. To compare, compute the software the same way: total annual contract, plus data and enrichment credits, plus sending infrastructure, plus the fraction of a human still needed to handle escalations, divided by meetings held rather than booked.
Done properly the software usually wins, sometimes by a lot. It just wins by less than the marketing suggests, and the honest version is checkable.