I Burned $23,000 on Lead Gen Tools Before I Understood What an AI Sales Assistant Actually Does

March 12, 2024 — 0.4%

I still remember the number. It was a Tuesday morning and I was staring at our outbound dashboard — the reply rate across our sending domain had dropped to 0.4%. Six months earlier, we were running at 6.8%.

For context: we started outbound properly in early 2023. Six SDRs, a list of about 50,000 contacts we'd bought from a data broker, and a lot of optimism. The first week was great. 9% reply rate on day one. In hindsight we were hitting fresh inboxes with a fresh domain — that combination works for about three weeks and then it doesn't.

By June we'd slid to 3%. We blamed saturation. By September, 1.5%. We blamed the market. By December I sat down and counted what we'd actually spent across 2023 on data, verification, enrichment, and automation: about $23,000, spread across four tools, none of which solved the problem.

Money wasn't the painful part. The painful part was that we kept applying patches to a wound we hadn't diagnosed. We figured out what the wound was in March 2024 — and honestly, it wasn't the tools.

Lesson one: verified email ≠ deliverable email

Our first purchase (Feb 2023) was an email verification tool. Roughly $300/month, priced by credits.

The logic felt airtight. Run our lists through the validator, cut the bad addresses, watch inbox placement climb. That's the pitch. We believed it.

It did exactly what it promised — it removed obviously invalid addresses. About 12% of our list fell off. Our reply rate barely moved.

Here's why. A verifier tells you whether an address resolves. It can't tell you whether the recipient has opened a marketing email in eight months, or whether they're a spam trap that looks identical to a real inbox. Catch-all domains make it worse — the validator shrugs and says "maybe." We sent to the maybes anyway.

It's tempting to think deliverability is a filter problem. But it isn't. It's an alignment problem — the list was full of people who had never been a fit for us in the first place, and no amount of cleaning was going to fix that.

That's the simplification error in a nutshell. We took a messy, multi-variable problem and reduced it to "validate the list." It was a satisfying thing to do. It just didn't work.

Lesson two: LinkedIn automation killed our sending domain

By August 2023 we were feeling desperate and bought a LinkedIn automation tool — around $800/month.

Three weeks in, our primary sending domain got blacklisted by two major providers. Not because our cold email was bad. Because the connection-request activity got flagged, the domain got associated with it, and everything we sent from that domain got dragged down with it.

People think the reply rate fell because the channel was saturated and we needed a new one. Actually, it's the other way around — the reply rate fell because we were pushing the same generic message through every channel we could find, and moving to LinkedIn just gave the problem a new address.

That mistake cost us about $4,200 in domain rehabilitation, lost sending time, and two weeks of SDR productivity.

Lesson three: we bought an AI SDR as a replacement, not a multiplier

Early 2024, we started seriously testing AI sales assistants. If you've read anything about what an AI sales assistant is and when a B2B team should use one, half the articles tell you it replaces headcount. We bought into that framing.

Our thesis was simple: the assistant sends the emails, we handle the replies, one person runs the work of three. Wrong. Emphatically wrong, and we found out fast.

What it actually did well: drafting first-touch emails, sending cadences, booking meetings on autopilot for obvious replies. All fine. What it couldn't do: tell us which accounts were worth touching at all, or decide when a reply was a polite no versus a real opportunity, or know when to stop.

By April we had hundreds of AI-drafted emails going out every week and a human bottleneck at the exact same spot as before — the reply triage. Worse, some of the drafts were going to people who had just been laid off at their company. Nothing in the workflow caught that. A human would have.

Looking back, I should have realized that we were buying sending capacity without ever defining a trigger. An AI sales assistant multiplies whatever you already have. If what you have is a list and a template, you get more of that. If what you have is a clear ICP, a signal-based trigger, and a human making the final call, you get something genuinely useful.

What we finally got right

Reconstruction started in June 2024. We spent two weeks doing nothing but writing down who we actually wanted to sell to and what would have to be true about a company for us to reach out.

That document was 400 words. It was more useful than the $23,000 we'd spent on tools.

Then we went looking for something that combined signal detection with human-in-the-loop sending. We evaluated a few platforms in the agent-native prospecting space — okkigo's okki-go prospecting agent was one of them — and the okki-go features that mattered to us weren't the flashy ones. It was the boring stuff. Waterfall enrichment across multiple providers so we weren't relying on a single data source. Intent signals pulled from hiring posts, funding news, and tech-stack changes, all merged into one score. And a gating rule: if the enrichment said an account was a bad fit for our current offer, the system wouldn't send. Period.

That last part sounds trivial. It isn't. Most outbound tools will send whatever you tell them to send, whenever you tell them to. Having a sales intelligence platform that occasionally tells us "don't touch this account yet" was the thing that finally made our outbound credible.

The reply rate didn't jump back to 9%. Nobody gets that twice. But it stabilized in the 4-6% range on a much smaller send volume, with way fewer spam complaints and a domain we're no longer terrified of checking in the mornings.

When should a B2B team actually use an AI sales assistant?

I get this question a lot now, and I've got a much shorter answer than I did a year ago.

Don't buy one until you can answer three things in writing:

1. Can you describe your ICP in one sentence without using the word "or"? If it takes you three clauses, the AI won't know who to target either.
2. Is there a specific, observable signal that tells you a company is in-market right now? Hiring for a specific role. Just closed a round. Migrated off a competitor. If your trigger is "they're in our TAM," that's not a trigger, that's a list.
3. Who makes the final call on whether a message goes out? If the answer is "the tool," don't buy the tool yet.

If you can answer all three, an AI sales assistant will make you meaningfully faster. If you can't, it'll just make you faster at something that wasn't working in the first place.

That's not a dig at the technology. It's a description of what happened to us — twice, because we apparently needed the second one to stick.

The tools are good now. Some of them, anyway. The question worth asking isn't whether the tool works. It's whether the thing you're pointing it at is real.