We’ve now deployed 20+ AI agents at SaaStr, sent 60,000+ AI SDR emails, and generate 50%+ of our pipeline and revenue from AI sales agents. Here’s what we wish someone had told us before we started.
The first thing I’ll say about AI SDRs: the tool doesn’t actually matter that much.
I know that’s not what you want to hear when you’ve been demoing 10 vendors. But it’s the truth. The tool is maybe 20% of the outcome. The other 80% is how you train it, who owns it, how you segment, and whether you’re willing to put in the ramp time.
We figured this out the hard way. In the early days, we got generic results because we gave it generic inputs. Once we treated it like a new hire that needed real onboarding—real context, real examples, real correction—results followed.
Here’s what I’d tell myself before we started.
1. Your AI SDR Is Only as Good as Your Human SDR Playbook
Before you deploy any AI, document what your best human SDR actually does.
Not what the playbook says they should do. What they actually do when they’re hitting quota.
Which sequences? Which subject lines? Which opening lines? Which objection responses? Which accounts do they skip and why? What’s their actual follow-up cadence?
If you can’t answer those questions, you’re not ready to deploy AI. All you’ll do is automate your confusion at scale.
We cloned Amelia’s best-performing outbound sequences—the ones that drove 12%+ positive response rates on warm audiences. Replicated the tone, the structure, the specificity, even the exact way she handles “not the right time” replies. That’s what made our AI work. Not the vendor. The underlying playbook.
The rule: If humans haven’t proven something works, AI won’t magically make it work. AI scales what’s already working. It doesn’t discover what works.
2. The Work Humans Refuse to Do Is Your Biggest Opportunity
For six years, we tried to get human SDRs to re-engage return attendees about tickets. We offered incentives. We built processes. We monitored activity. Nothing worked. Humans think it’s beneath them, or awkward, or a waste of their time on “small” prospects.
We deployed an AI agent on that exact segment. It now generates 15% of our event ticket revenue.
Found money. Six years of leaving it on the table.
The pattern repeats everywhere: small accounts nobody wants to work. Follow-up sequences that humans “forget.” Re-engagement of churned customers. Inbound leads that come in at 11 PM. None of this is glamorous. All of it drives revenue. AI does it without complaint at 3 AM on a Saturday.
The question to ask your team: What are the things we know we should be doing but consistently don’t? That’s your first AI SDR deployment.
3. Segment Ruthlessly — 800 to 1,000 Contacts Max Per Campaign
One of the fastest ways to kill your AI SDR results is to blast your entire database.
We run 800-1,000 contacts per campaign, max. We keep it that tight because:
- Messaging that works for a 2024 SaaStr Annual attendee doesn’t work for a website visitor who downloaded one ebook
- You need to be able to read and respond to the output. If you send 10,000 emails and 400 people reply, you’re overwhelmed
- Personalization degrades fast at scale. At 500 contacts you can be specific. At 5,000 you’re sending slop
Our best campaigns targeted SaaStr Europa attendees (warm, recent, specific context). Our response rates were 11-12%. When we got lazy and broadened targeting, response rates dropped by half.
The discipline: Resist the temptation to maximize volume. Maximize relevance. The math works out better.
4. “Pretty Good” at Scale Beats Brilliant and Inconsistent
Human SDRs have good days and bad days. When your top rep is locked in, their emails are exceptional. When they’re tired, distracted, or just going through the motions, they’re mediocre.
AI is consistent. Every email is about the same quality. If you train it well, that quality is “pretty good to good.” Not exceptional. Not the single best email your best SDR ever wrote. But reliably solid, every time, at 10x-40x the volume.
Our AI SDR sends 3,221 emails per month from a single platform. Our human SDRs used to send 75-285 per month each. That’s an 11-40x increase in volume at consistent quality.
Stop trying to make it produce your best human’s best day. Train it to produce your average human’s average day—then scale it.
5. You Need Two Humans to Make This Work
People think they can just buy an AI SDR tool and let it run. That’s wrong. You need two specific humans to make this work:
Person 1: A forward-deployed engineer or implementation specialist from the vendor. Not a sales rep. Not a customer success manager who checks in once a month. Someone who is in the platform with you, troubleshooting training issues, pulling data, and helping you understand what’s actually happening under the hood. The best vendors have these people. Demand them.
Person 2: A GTM-obsessed internal owner. Someone who reads every early message, obsesses over response rates, updates training based on what’s working, and treats the AI agent like their most important new hire. At SaaStr, this person is Amelia—our Chief AI Officer who spends real time daily on our AI agents.
If you don’t have both, you will get mediocre results, blame the technology, and conclude AI SDRs don’t work. They work. You just can’t deploy them into a vacuum.
6. Read Every Message in the First 30 Days
We skipped this step early on. It cost us.
In the first month, your AI will get facts wrong, use awkward phrasing, misread company context, or reply to objections in ways that make you cringe. You won’t know unless you read everything.
We now have a strict rule: For the first 30 days of any new agent, read every outbound message and every inbound reply. Every one.
After 30 days of building trust—confirming the AI handles edge cases well, that the training is solid, that responses are on-brand—we move to spot-checking and flag-based alerts for unusual patterns.
This isn’t optional. If you’re not reading the output, you’re not running an AI SDR. You’re sending unsupervised emails to your best prospects and hoping for the best.
7. Ramp Time Is Real — Budget 30+ Days Minimum Per Agent To Get It Off The Ground
The “set it and forget it” pitch is a lie.
Every AI SDR tool will tell you you can be up and running in a day or two. That’s technically true. Being good takes two weeks minimum, per agent.
That two weeks goes toward:
- Training the agent on your ICP with real examples
- Testing messaging and iterating based on early replies
- Building escalation rules (what goes to a human, what the AI handles)
- Calibrating tone, length, and follow-up cadence
- Figuring out which segments actually respond
Budget the time. Don’t judge the technology in week one. The companies that conclude “AI SDRs don’t work” are almost always the ones who gave up after seven days.
8. Chat First, Then Voice — Don’t Over-Complicate the Stack
Everyone asks about chat vs. voice vs. video AI SDRs. Our data is pretty clear: about 85% of prospects prefer chat. 15% prefer voice.
Chat is dramatically easier to implement, easier to train, easier to monitor, and easier to fix when something goes wrong. Start there.
Voice is the natural second step once you have chat working well. We did our voice clone on 11 Labs in five minutes. Feels much harder than it is.
Video AI is two orders of magnitude more work. If you have a very high ASP and video trust signals are meaningful in your sales cycle, explore it. Otherwise, not worth it yet.
The mistake is trying to build across chat, voice, and video before you’ve proven value in any single channel. Pick chat, make it work, then layer in voice when you’re ready.
9. Don’t Bet Your Rollout on Someone Who Might Leave
This one hurt us to learn, and we’ve seen it hurt others more.
GTM turnover is high. If you’re building an AI SDR strategy that’s heavily dependent on one person—especially if you’re cloning their voice, training agents on their messaging style, building sequences around their relationships—and that person leaves, you’re in trouble.
We’ve watched founders do full AI SDR implementations centered on a CRO or VP of Sales who was gone three months later. All that training, all that voice cloning, all those sequences—orphaned.
Build your AI agents to be institutional knowledge, not individual-dependent. Document the logic. Make sure multiple people understand the training. And if you’re going to deeply clone a specific human, make sure they have real equity and real reasons to stay.
10. Fix the Fundamentals First — AI Scales What Exists
I’ll end with the hardest truth.
If your outbound doesn’t work with humans, AI won’t fix it. If your messaging is off, AI scales bad messaging. If your ICP is wrong, AI targets the wrong people more efficiently. If your offer isn’t compelling, AI delivers an uncompelling offer to 40x more people.
The companies that get the worst results from AI SDRs are the ones who deployed it because outbound “wasn’t working” and hoped this would fix it. It won’t. It amplifies what’s already there—good or bad.
Fix the fundamentals first:
- Proven messaging (evidence that humans get responses)
- Clear ICP (you can describe in one sentence who you’re targeting and why)
- Compelling offer (a real reason to engage, not just “let’s hop on a call”)
- Working sequences (you know what follow-up cadence produces replies)
Then deploy AI to scale those fundamentals.
If you do that—you’ve done the hard work, you own the playbook, you have two humans running the program, you’re reading everything early, and you’re segmenting ruthlessly—the results are real. We’re generating pipeline at a fraction of the cost, around the clock, from segments we never would have touched with human SDRs.
But it takes real work to get there. Anyone telling you otherwise is selling you something.
SaaStr has deployed 20+ AI agents across GTM, customer success, and content. See all our tools and specific use cases at saastr.ai/agents.
Learn how to WIN in the age of AI + B2B … and how not to get left behind. Join us and 10,000+ of the best in B2B + AI at 2026 SaaStr AI Annual, May 12-14 in SF Bay!!
More here.
