AI SDRs get cancelled when they are asked to do the one thing that cannot be undone, sending to strangers under your name, with a list and infrastructure nobody checked. The research, enrichment and drafting parts hold up much better when a person reads the output.
I have no survey of cancelled accounts, so this post does not claim to explain why any particular customer left. It is where I would look first, based on how outbound breaks.
Where does autonomous sending fail?
It fails in four places, and the dashboard rarely separates them.
The list. The agent can only email who it was pointed at. In my work, most campaigns are lost at the list, not the copy, and an agent that scales a weak list scales the problem. See how I classify an ICP.
The infrastructure. Volume that ramps faster than reputation looks like a spike, and a spike looks like spam. My own Workspace suspension, written up in Google Workspace inboxes suspended, came from setting up and connecting too much too fast. An autonomous tool that pushes volume without regard for domain age makes that mistake by default unless someone constrains it.
The message. Models write fluent, generic text unless you force specifics. Without facts in the table, you get the openers everyone recognises. My rules for avoiding that are in AI first lines that don't sound like AI.
The replies. Someone has to read what comes back, decide what counts as interested and act. That is the part that makes a campaign pay off, and it is the part least likely to be automated well.
What tends to be missing when a tool gets dropped?
I would run this as a post-mortem checklist. For any cancelled or struggling AI SDR, answer these in order:
- Was there a written ICP, with tiers, before anything sent?
- Were addresses verified immediately before sending?
- Were domains old enough and volume per inbox within a sane range?
- Did a person read a sample of the messages before each batch?
- Did a person read the replies the same day?
- Was the success measure positive replies and meetings held, or just emails sent?
If the answer to several is no, the tool was not the only problem. It is the same order I use in why cold email lands in spam, and cold email itself is fine when those basics are right, as I argued in is cold email dead.
What still works?
Anything where the agent prepares and a person decides.
| Task | Agent role | Human role |
|---|---|---|
| Account research | Summarise the company, find the fact | Check the fact |
| Fit scoring | Score and explain | Read the top tier |
| Inbound triage | Sort, summarise, route | Answer the hot ones |
| First lines | Draft from table facts | Approve before send |
| Reply sorting | Label, flag | Handle the interested ones |
The inbound case is the clearest. In my test of an inbound scoring table, the AI sorted five of six leads correctly and failed on the one with no company to look up, which a one-line rule fixed. I would still not auto-reply to hot leads. Details are in qualify inbound leads with AI. The same split applies to outbound, laid out in my checkpoint setup for Claude Code.
What about Clay's Inbound SDR Agent?
Clay's Sculpt recap, published after the 8 October 2026 event, describes the Inbound SDR Agent as previewed and pre-beta, with a waitlist for companies that want to become pre-beta partners. The same post says Workflows is in beta for all users and the agent-first builder is rolling out.
That is all anyone outside the program can say right now. I have not used it and I am not judging it. If it has a human review step and a limit on what it can do alone, my objections shrink. If it sends autonomously, run the checklist above on it before trusting it.
Where this advice stops working
My view comes from outbound where deliverability and brand risk are the constraints. In a setting with a warm inbound base, a small risk of a bad message, and a good review process, a more autonomous setup may be fine. I also have not measured autonomous tools against human-run campaigns, so I cannot say which wins on meetings held.
The practical rule: automate the preparation first, and give an agent the send button only after its drafts have held up under your own reading for a good while.
I'm Malik, and you can see how I run campaigns in the outbound numbers nobody shows you.