The AI SDR category has got substantially better in two years. Copy quality is no longer the bottleneck, reply handling works for most straightforward cases, and the tools genuinely compress work that used to take a person a day.
The failure modes have barely moved, though. Talk to teams that tried one and stopped, and the same seven things come up. Most of them are not really about the AI.
1. It was a deliverability failure wearing an AI costume
The most common story by a distance. The tool worked, the copy was fine, the sequences ran, and almost everything landed in spam.
An AI SDR sends more mail than a human does, from more mailboxes, faster. That amplifies whatever your infrastructure was already doing. If authentication was slightly wrong or warmup was rushed, a human sending 200 a day would surface the problem slowly. A machine sending 3,000 surfaces it at scale before anyone reads a report.
The diagnostic is straightforward: before concluding the AI failed, check placement. If you were not running inbox placement tests, you do not actually know whether the copy was ever read.
2. Nobody was watching the metric that mattered
Related but distinct. AI SDR dashboards report sends, opens, replies and meetings. None of those tells you where mail is landing, and open rates in particular have been unreliable since privacy protection became default on Apple Mail.
Teams watch reply rate, see it decline, and start changing copy. The copy was not the problem. Placement was, and it degraded four weeks earlier.
Set up placement monitoring on a schedule independent of whatever the AI tool reports.
3. The personalisation stopped being personal
Every team using similar models with similar prompts produces convergent output. The LinkedIn-bio opener that worked in 2024 now reads as automated, because it is, and because recipients have seen it two hundred times.
This is a moving target rather than a solvable problem. What works is personalising on something the model cannot infer from a profile: a recent post, a job change, a hiring signal, something that happened this week. The signal matters more than the sentence built from it.
What does not work is more adjectives. Turning up the personalisation dial on the same inputs produces more elaborate versions of the same email.
4. It does not know when to stop
A human SDR working a segment notices after forty sends that nobody is replying, and says something. An AI system completes the run.
Few tools have a genuine stop condition on segment performance. They have volume limits and bounce thresholds, which are not the same thing. The result is that a poorly chosen segment gets fully worked before anyone reviews it, spending both sender reputation and a list you cannot re-contact for months.
Worth asking any vendor directly: what causes this to pause on its own?
5. Ambiguous replies get handled consistently rather than correctly
“Interesting, but we just signed with someone else.” “Not right now.” “Can you send more information?”
Each of those could be a dead end or a real opportunity, and a human reads tone, timing and account context to decide. Models tend to treat each category the same way every time, and which way depends on how the system was prompted rather than on the situation.
That consistency is genuinely useful for clear positives and clear negatives. It is a liability in the middle band, which is also where most of the interesting pipeline lives.
The fix is not a better model. It is a handoff rule that routes ambiguity to a person.
6. It answers questions it should have escalated
A prospect asks about SOC 2 scope, data residency, or whether you can support a specific integration. The correct response is to stop and get a person.
Systems that answer everything will answer these, sometimes plausibly and wrongly. A confidently incorrect answer about security posture to a prospect evaluating you is worse than a slow one, and you often never find out it happened.
Ask a vendor what the system does when it does not know. Answers about escalation are good. Answers about confidence are not.
7. Nobody owned it
The quietest failure. The tool was bought, configured over a fortnight, and then ran. Nobody reviewed the segments, refreshed the list, iterated the copy or watched deliverability, because the pitch implied those things were handled.
An AI SDR reduces the work. It does not reduce it to zero, and the residual work is exactly the work that decays if unowned: list quality, messaging relevance, placement.
Teams that get results assign a person to it. Usually one person covering roughly the output of three traditional SDRs, which is a real productivity gain and a different claim from autonomy.
What to check before blaming the tool
In order:
- Placement. Run an inbox placement test. If you are in Promotions or Spam, nothing else on this list matters.
- Authentication. SPF, DKIM, DMARC and MX on every sending domain, verified rather than assumed.
- Warmup. Were the mailboxes warmed for three to four weeks, or rushed to hit a launch date?
- List quality. What is the bounce rate? Above 3% is damaging the domains as well as wasting sends.
- Segment performance. Is one segment dragging the average down while another works?
- Reply handling. How many replies went unanswered for more than a day?
Most “the AI SDR did not work” conclusions do not survive that list.
The short version
The category’s real failures are infrastructure failures, ownership failures and handoff failures, in roughly that order. The models are good enough that copy quality is rarely what breaks.
So before you evaluate the AI, evaluate the pipes. Check placement, check authentication, check warmup. Then decide what the system is allowed to handle alone and what it must escalate, and give it a stop condition that is about performance rather than volume.
And give it an owner. The tool does less work than the pitch implies, and the work it leaves behind is the work that matters most.
Next: AI SDR versus hiring an SDR, priced, or the deliverability guide.
Frequently asked questions
Do AI SDRs actually work?
For high-volume, broad-ICP motions with a product that explains itself in an email, yes. They fail most often not on copy quality but on deliverability, on the absence of a stop condition, and on nobody owning the programme after setup.
Why did my AI SDR campaign fail?
Check inbox placement first. AI SDRs send more mail from more mailboxes faster than a human, which amplifies any existing authentication or warmup problem before anyone reads a report.
What can an AI SDR not do?
Judge ambiguous replies, escalate questions outside its brief, notice a segment is not responding, or tell you the messaging has gone stale. Those require a person, which is why teams run roughly one person per three SDRs’ worth of output.
How much oversight does an AI SDR need?
A few hours a week at modest volume, growing to a meaningful part of a role at scale. Budget for the person; the software cost is the smaller line.