We Fired Our SDR Pipeline. An Agent Runs Prospecting Now.

Six months ago we had the same problem every B2B company has.

Find companies that look like your best customers.

Figure out who to email.

Write something that doesn't read like a mail-merge.

Follow up three times before anyone gives up.

That's the entire SDR job description. Not the impressive part — the grind part. The part where a smart person spends four hours a day doing pattern-matching and typing, and calls it "sales."

We didn't hire a fourth researcher to keep up. We built an agent that does it instead.

What Actually Sat on an SDR's Desk

Strip the job title away and a prospecting SDR's week is four repeatable steps: pull a list of companies matching an ideal customer profile, research each one enough to know if they're real, write a first-touch email that references something true and specific, queue the follow-ups.

None of that requires judgment that changes case to case. It requires consistency — which is exactly what humans are bad at across a 40-hour week and software is good at continuously. The cost isn't the salary. It's the variance. A good Tuesday produces a sharp, specific email. A tired Friday produces "Hi {{firstName}}, hope you're well." Same person, same job, wildly different output — and the tired version is the one that actually goes out most weeks.

Here's Exactly What Runs Our Top of Funnel Now

Not a black box. Not "AI handles it." Six connected n8n workflows, each with one job:

1. Sourcing. Pulls candidate leads against our ICP on a schedule and drops them into a queue. No one is building lists by hand.

2. Research and qualification. Claude reads real signals for each lead — company activity, role, public context — and makes a keep/reject call. Most leads get rejected here, before a draft is ever written. This is the step that used to be a person's entire morning.

3. Drafting. For every lead that clears the research gate, Claude writes the first email — a specific paragraph, referencing something specific to that company, not a template with a name swapped in.

4. Human-in-the-loop gate. Nothing reaches a real inbox without a person approving it first. This isn't a compliance checkbox — it's the actual control that makes the rest of the system safe to run unattended. The agent proposes. A human decides.

5. Campaign push and follow-up. Approved emails go out through the sequencing platform. Follow-ups are scheduled automatically — no one is manually remembering to touch a lead again in four days.

6. Sync. Results flow back into a shared pipeline in real time, so a lead that replied yesterday doesn't get a generic follow-up today from someone who didn't see the reply.

Six stages, each doing exactly one job, each replaceable on its own if it needs to get smarter without touching the other five.

The Number That Actually Matters

The system is engineered to sustain real outbound volume — the kind of daily send rate that would otherwise require multiple full-time researchers just to keep the top of the funnel fed — without adding headcount for that function. We're not going to hand you a "10x'd our pipeline" number we can't defend. What we'll defend is this: the throughput this replaces used to require people. Now it requires review time, which is a fraction of the labor.

What This Doesn't Replace

The person who gets on the call and closes. The relationship. The judgment about which deal to push and which to walk away from. None of that is automated, and none of it should be — that's the actual sales job, the part that was never repetitive in the first place.

What disappears is the research-and-first-draft labor that used to consume most of an SDR's day before a single reply ever came back. That's the part that was always administrative, wearing a sales title.

Why This Isn't Actually a Sales Story

The pattern underneath — research a target, decide if it qualifies, draft a first pass, put a human in the loop before anything commits — isn't sales-specific. It's the same shape as a paralegal screening incoming matters, a bookkeeper classifying transactions before they hit the ledger, a support rep triaging tickets before they're answered. We built this architecture once. The vertical is a configuration detail, not a rebuild.

Where DIY Versions of This Fall Apart

Every team that tries to DIY this makes the same three mistakes, in the same order.

Mistake one: skipping the qualification stage. Teams plug a list straight into an email-writing prompt and wonder why the output reads generic. Qualification is the stage doing the real work — it's what stops the agent from writing a beautiful email to a lead that was never going to buy. Skip it and you've built an expensive template generator.

Mistake two: no real approval gate. "Human review" that's actually a person glancing at a dashboard once a day isn't a gate, it's a rubber stamp. The gate only works if it's genuinely in the send path — nothing goes out without it, not "nothing goes out without it most of the time."

Mistake three: one giant workflow instead of six small ones. Teams build this as a single sprawling automation because it feels faster to ship. It works for about a month, then the first time research logic needs a fix, changing it breaks drafting, which breaks the send step. Six small workflows with one job each cost more to design up front and save every week after.

The Realistic Rollout Timeline

Nobody should turn this on for an entire pipeline on day one.

Weeks 1-2: Connect sourcing to a real ICP definition and a small test segment, not the full target list. Tune research and qualification until reject/accept decisions actually match what a good SDR would decide.

Weeks 3-4: Turn on drafting for the accepted segment, with every draft going through the human gate. This is where most real tuning happens — prompt adjustments, tone corrections, catching patterns the model gets wrong.

Month two onward: Expand segment size as approval rate and edit rate stabilize. The system doesn't get "finished" — the segment just widens as trust in the output grows.

This is the kind of build a focused AI automation agency ships for revenue teams — not a generic vendor bolting on a chatbot, one that actually wires the research-to-approval chain end to end.

Frequently Asked Questions

Does this replace our CRM or email tool?

No — it sits on top of them. The agent researches and drafts; your existing CRM and sending platform still own the record and the send.

What stops the agent from sending something wrong?

The human-in-the-loop gate. Every draft is queued for approval before it reaches a real inbox. The agent's job is to make that approval fast, not to remove it.

How long does something like this take to build?

Weeks for a well-scoped first version. The six-stage architecture is reusable — the ICP logic and prompt tuning specific to your business is the part that takes iteration.

Does this only work for outbound sales?

No. Source, research and draft, human approval, execute and sync — that shape is what we reuse across every research-heavy, repetitive-decision workflow we build.

What happens when the agent gets a lead wrong?

It gets rejected at qualification or corrected at the human gate — the same as a new SDR would get corrected in their first month. The difference: the correction feeds back into how the next batch is qualified, not just remembered by one person.

Do we need an internal engineering team to maintain this?

No — the six stages are built to be maintained by whoever owns workflow tooling, not a dedicated engineering hire. That's part of why it runs on n8n rather than custom code.