AI SDR: volume up, quality meh

localpack

We jumped on the AI SDR bandwagon about 3 months ago. Stack: Clay + an AI writer for personalisation. Results? Mixed bag.

The good: Volume is insane. We're hitting 10x the prospects we used to. The AI actually writes decent first lines when you feed it solid data. And it catches stuff human SDRs miss - job changes, funding news, that sort of thing.

The bad: Still needs a ton of human oversight. The AI will confidently write total nonsense. Like referencing a company achievement that never happened. Embarrassing when a prospect calls you out. Bounce rate is rough too because email data quality varies wildly depending on the source.

We're still tweaking the setup. Clay's powerful but the learning curve is steep. Tried pulling contacts from Apollo - emails bounced at an ugly rate. Kind of defeats the purpose of all that automation if half your messages never land.

Anyone else running AI BDR setups? What's your stack? How are you handling the data quality problem?

metricsmuse

Oh, this is exactly it. The hallucination and bounce issues feel like two different monsters, but they both come down to one thing: crap inputs. You can polish the AI writer all you want, but if the data enrichment layer is weak, you're just dressing up garbage. Fix the sourcing first, and suddenly the personalisation actually works without you having to babysit it.

socialbutterfly

That matches what I’m hearing from a lot of teams honestly. The bottleneck is shifting from “writing outbound” to data quality, enrichment accuracy, and keeping personalization grounded in reality instead of confident hallucinations.

Volume is basically solved now. The harder problem is maintaining trust signals so your outbound doesn’t start feeling like infinite AI sludge after the first sentence.

ethan-pr

The pattern I keep seeing is that AI SDRs help with throughput before they help with judgment.

They are useful for:

finding obvious account/contact changes
summarizing why an account might be relevant
drafting first-pass personalization
routing replies and enriching records
keeping follow-up from falling through the cracks

But the failure mode is exactly what you described: bad data plus confident writing. If the source says the wrong person, old role, fake trigger, or weak signal, the AI just turns that into a smoother bad email.

I would put guardrails around three things:

Data confidence: do not let one source decide the reason for outreach.
Human approval: review samples before a new segment/campaign goes live.
Success metric: track qualified replies and accepted meetings, not just volume or booked calls.

The best setup is probably not "AI replaces the SDR." It is more like "AI does the research/drafting/admin, and a human owns the judgment layer until the workflow is proven."

If the system cannot explain why this account, why this person, and why now in plain English, it is not ready to send automatically.

davidsearch

Haven't seen any reliable CPA improvements from AI SDRs ourselves. The novelty wears off once you look at actual conversion data - no sustained lift in qualified lead volume.

paperclick

Yeah, same here - AI SDR works when data is clean but falls apart fast when enrichment gets noisy. Someone still ends up doing the last mile checking before anything ships.

pixelperfect

i had a client who replaced a team of 30 SDRs with an off-the-shelf AI tool. The volume of emails shot up, but conversion was abysmal. When we pulled the numbers from their human team, 65% of booked demos came through the phone - something the AI just doesn't touch. And 90% of those demos had at least one phone outreach in the sequence. Average deal size was $58k. The AI was basically burning budget on noise while the real pipeline engine - human voice contact - sat silent.

tiktokguru

We've run Qualified and Zendesk through the wringer ourselves. Here's the short version:

🛠️ Qualified is brilliant for the initial outreach bit you're talking about - it nails that first touch. But as soon as a prospect asks anything remotely complex or technical, it falls over. Too rigid for deeper conversations.

🔧 Zendesk handles the technical side better, but our onboarding took a solid six months to get right. Not a quick win by any stretch.

If you can keep building your own system, I'd stick with that path. The flexibility and cost savings are real - and honestly, with Claude Code, building and iterating on something custom isn't as hard as most people think. That's essentially my day job now.

wordweaver

Yep, that's the classic AI SDR rollercoaster. The model itself is rarely the bottleneck - it's everything else around it that's broken.

From what I've seen, you pretty much nailed the top five: dodgy source data, half-arsed enrichment, no "why this company, why now" logic, zero human QA on claims, and a handoff process that might as well be a black hole.

The "personalisation" that's actually just AI making stuff up? That's the silent killer. Looks clever to the intern, but the prospect catches it immediately. I'd slap a hard rule on it: if the AI can't show you the exact source for a claim, that claim doesn't go anywhere near the email.

The setups that actually work aren't "AI SDR replaces SDR". They're more like: AI finds and preps candidate accounts, enrichment checks the contact and the trigger, a human sanity-checks a small sample for truth, then the sequence fires only if the reason to reach out is genuinely specific. Replies still get routed to a human fast.

Volume is only helpful once the data layer is boring and reliable. Otherwise you're just scaling up embarrassing mistakes at speed.

pixelperfect

Honestly, that's been my experience too. AI SDR tools only work when you keep a human in the loop for research and the first draft, then lock it all behind email verification plus a small daily sample check. Otherwise it'll scale bad data faster than you can say "pipeline."

For warmer outbound, I've had good results with SocListener - it pulls people already talking about the problem in online communities, so the data feels much more relevant than blasting through Apollo lists.

communityfirst

Oh, this is exactly what we saw too when we tested a handful of these tools. funnily enough, some of them come with a message rater built in. We started checking how it scored the emails that had actually booked meetings for our SDRs, and consistently those got a C or a D. That told me the calibration was completely off - the system didn't know what good looked like.

in the end, we just couldn't get the AI SDR to work the way we wanted. So we built something in-house instead, spending ages figuring out what pain the prospect is actually experiencing that we can solve, and what cues really highlight that pain. That way we move away from the generic funding announcements or surface-level signals that most of these tools rely on.

What I've noticed is that the early spike in meetings people see is mostly just volume - you hit more people, so you catch the ones already in market that your human SDRs might've missed. But the actual response rate is pretty atrocious.

I had a good look at what Jason Lemkin from SaaStr has been sharing about how much time they spend reviewing every email before it goes out. I did the same exercise and it hit me: these systems generate emails that sound fine on the surface, but they're hollow. I wouldn't answer them either.

pixelpusher

enrichment is mostly a vanity metric in the SDR world. I'm in senior roles too-never once acted on an email because someone knew my job title or company size. The only data point that ever made me pause? a personal trigger: a funding round announcement, a public hire, a product launch. without that, you're just blasting noise with better formatting. what's your conversion rate actually look like after enrichment?

digitalnomad

Oh, absolutely - this is the exact problem we're starting to see play out with a few clients too. 🚩

Everyone's feeding off the same data lake now: same intent signals, same firmographic triggers, same AI models. So every outreach ends up reading like a copy-paste job with just the logo swapped out. prospects are getting wise to it fast.

what I've found actually works is getting creative about where you pull signals from:

🔍 Proprietary / niche data sources - community forums, review sites, or even your own product usage data
💰 Paid-for exclusives - buying access to datasets your competitors haven't discovered yet
🧠 Internal behavioural cues - things like support tickets or feature requests that hint at immediate need

The real edge isn't the AI anymore - it's the unique signal set you're feeding it. if everyone's using ChatGPT on the same inputs, you just get identical output with different fonts.

metricsmuse

Honestly, it really depends on the industry and your market, but the core idea is pretty simple: you only reach out when you've got a genuine reason to believe your product could actually help them right now. Like, exactly what you said-job postings, company news, recent announcements about their direction or focus. That's what makes it relevant, good timing, and gives you a real excuse to start the conversation. Anything else just feels like noise.

gmbpro

I've seen a few AI SDR setups. Two consistent tracking issues:

Broken conversion attribution
Inflated intent scores from non-human traffic
What metrics are you using to validate?

pixelpusher

Makes sense, thanks! Though I'd push back on the assumption that most AI SDRs are actually moving the needle. Every implementation I've seen so far is just a glorified spam cannon with a chat interface. Teams celebrate meeting call quotas while pipeline quality tanks. If your definition of "working" is booking meetings that don't close, yeah, they're brilliant. But if you're measuring real revenue influence, most of these tools are still a net negative because they flood reps with low-intent leads. Happy to be proven wrong, but show me a case where an AI SDR actually out-performed a decent human outbound rep on conversion to closed-won, not just activity metrics.