Jev vs. Your SDR Queue: How We Cut 60,000 Leads to 2,000 ICP Fits Before Outreach

Quick Overview

This playbook is for operators who own an SDR queue and keep watching expensive outreach burn on leads that never fit. The job is lead qualification before dial time: take a raw list, score it against a written ICP, and only then decide who gets a human. The mechanism is a cheap Jev classify pass that labels each row hold, route, or recommend, with a person reviewing the uncertain calls. The outcome is a short list of ICP fits instead of a week of wasted sequences.

One recent benchmark: 60% of SDRs were at quota. That is the lowest share in this study’s history. The median rep logged 112 activities a day and 4.1 quality conversations. Source: The Bridge Group, 351 B2B companies, published February 6, 2025. The study does not say what share of a raw lead list is a real ICP fit.

• Write ICP criteria once, then let Jev classify every row against that sheet.
• Route clear fits to SDRs; hold clear misses; park uncertain rows for a human.
• Never invent fit scores from vibes. Use fields you can audit.
• Treat the 60,000 to 2,000 cut as a queue design problem, not a model demo.
• Pair this sorter with a hire-or-build decision for the agent that owns the job.

Tuesday morning, an SDR manager opened a CSV export that should have been a gift and felt like a liability. Sixty thousand leads sat in the sheet. Firmographics looked fine at a glance. The outbound calendar for the week was already full. Everyone knew the truth nobody wanted to say out loud: only a thin slice of those rows would ever match the ideal customer profile, and every bad dial burned calendar, reputation, and sequence capacity.

Eric laid out the same shape in plain language on X. Say you have 60,000 leads, but only 2,000 fit your ideal customer profile. Give Jev the criteria to classify them, then decide which leads to hold, route, or recommend. A person reviews the uncertain calls before a bot sends anything that looks like outreach. That is the sorter. It is not a chatbot demo. It is a gate in front of the SDR queue.

Why ICP leads get expensive when you skip the sorter

Most teams treat the list as the asset. The real asset is the short list of ICP fits your SDRs can actually work. When every row looks equal, the queue fills with near-misses: wrong company size, wrong buying motion, wrong timing, wrong title. Sequences still fire. Meetings still book on paper. Pipeline still looks busy. Conversion tells a quieter story two weeks later.

The sorter flips the cost curve. Cheap classification happens first. Expensive human time happens second. That order matches how Eric talks about agent stacks elsewhere: chunk the work, tag it cheaply, and only then spend the scarce resource. For ICP leads, the scarce resource is SDR attention and domain reputation.

What “fit” must mean on the sheet

Fit is not a vibe score from 1 to 100 with no definition. Fit is a short checklist your team can disagree about in a review meeting. Industry match. Employee band. Tech stack signals you actually care about. Buying role vs influencer. Exclusion list for agencies, students, or competitors. If a criterion cannot be pointed to in the row or in a linked enrichment field, it does not belong in the first Jev pass.

The Jev classify loop: hold, route, recommend

Build the loop like an ops runbook, not like a prompt playground.

1. Freeze the ICP sheet. One owner. Versioned. No silent edits mid-batch.
2. Chunk the list. Send batches Jev can finish without burning the expensive model on every token.
3. Classify each row. Labels only: hold, route, recommend, or uncertain.
4. Human gate on uncertain. A person opens the row, checks the sticky fields, and upgrades or kills the label.
5. Hand route rows to the SDR queue. Recommend rows may need a manager glance before they enter a high-touch sequence.

Eric’s wording matters here. The person reviews the uncertain calls before a bot acts. That single sentence is the difference between a useful sorter and an autopilot that invents confidence.

Sample label contract

Label	Meaning	Next owner
hold	Clear miss vs ICP	Archive or nurture, not SDR queue
route	Clear ICP fit	SDR sequence owner
recommend	Strong fit with a caveat	Manager or senior AE glance
uncertain	Missing field or conflict	Human review desk

Operator scene: the uncertain pile is the real work

In a September working session on a Single Grain stack, the pattern that slowed the team was not the clear misses. It was the middle band: titles that looked senior but sat at companies outside the employee band, or domains that matched industry keywords while the product motion did not. Jev could flag those rows as uncertain in seconds. A person still had to open three fields and decide. That desk time is the cost of honesty. Skipping it just moves the cost into wasted dials.

If your enrichment is thin, the uncertain pile grows. That is a data problem, not a model failure. Fix enrichment or shrink the ICP criteria until the classify pass is auditable. Do not paper over gaps with a higher temperature and a fake certainty score.

Where this sits next to hiring or building an agent

Sorting ICP leads is a pipeline job with a clear input, a clear output, and a kill switch. That is the same framing we use when deciding whether to hire or build an agent that owns a job instead of demoing a chatbot. The broader hire-or-build guide lives here: how to hire or build AI marketing agents that own pipeline jobs. Use that page when you are choosing ownership. Use this page when you are designing the sorter in front of the SDR queue.

For kill switches and SLAs once the job is live, keep SLAs and kill switches for managed marketing agents next to this runbook. If cost per classified batch starts to matter as volume grows, pair it with how to price managed marketing agents.

What not to automate yet

• Changing ICP criteria without a named owner and a version note.
• Sending outreach from the recommend or uncertain buckets without a human.
• Letting the model invent firmographic fields that were blank in the source row.

A practical weekly cadence

Run the sorter on a fixed cadence so SDRs stop living in export chaos.

1. Monday: freeze ICP sheet and pull the raw export.
2. Monday afternoon: Jev classify pass on the full batch.
3. Tuesday: human desk clears the uncertain pile before noon.
4. Tuesday afternoon: route and recommend rows enter the SDR tools.
5. Friday: review false routes and false holds; update the ICP sheet only with written changes.

Measure the sorter with boring metrics: share of rows labeled uncertain, share of routed rows that earn a real conversation, and time from export to first dial on a routed lead. If uncertain stays high for two weeks, stop blaming the model and fix the criteria or the enrichment.

Failure modes to log every Friday

Pick five wins and five misses. For each miss, name whether the criteria, the gate, or the handoff failed. Update one rule only. Pair that teardown with the same order Eric uses on agent stacks in his Jev + Grokbot walkthrough: chunk first, spend expensive attention second.

When the uncertain pile stays fat, treat it as an enrichment problem. Eric’s note on deciding which leads deserve the next action is the same discipline: Jev classifies, humans still own the sticky calls.

Pick five wins and five misses from the week. For each miss, name whether the criteria, the gate, or the handoff failed. Update one rule only. Broad rewrites of the whole stack every Friday are how institutional memory dies and how teams lose trust in the agent job.

Keep the Friday note in the same place as the prompt. If the note lives in a random Slack thread, next week starts from folklore again. The point of a managed job is that a new owner can read the runbook and run the same loop without a private briefing.

Refuse remains a first-class outcome. Not every row, bookmark, call, render, or flagged sentence deserves a clever next action. Teams that cannot refuse end up automating noise and then blaming the model for being unreliable when the real failure was missing judgment.

Owner checklist

• Input contract frozen before the run.
• Uncertain items have a named human.
• Refuse remains valid.

Week-over-week, publish a one-line health note: volume processed, uncertain share, and refused share. Those three numbers tell you if the job is alive.

Owners should be able to pause the job in one step. If pause requires an engineering ticket, you do not have a kill switch yet.

New teammates should shadow one full cycle before they edit prompts. Prompt edits without cycle context recreate silent failures.

Keep screenshots out of the source of truth. The runbook and the audit columns are the source of truth.

If a stakeholder wants a demo instead of a job, show the audit column. Demos without audits are how chatbot theater returns.

Write the refuse path in plain language. If nobody can name what gets refused, the system will accept everything under time pressure.

Store prompt versions next to ICP or brief versions. Mixing versions mid-batch creates labels nobody can defend later.

Celebrate clean uncertain piles. A shrinking uncertain share usually means better inputs, not a smarter model.

When two systems conflict, log the conflict. Silent resolution teaches the agent to hide disagreement.

Schedule a monthly teardown with five wins and five misses. Change one rule. Only one.

If the job cannot survive a vacation week, it is still a personal workflow, not a managed agent job.

Want the sorter in front of your SDR queue?

Single Grain builds agent jobs that classify before humans spend calendar. If you are sitting on a pile of ICP leads that still looks like raw volume, talk to Single Grain about a Jev-gated sorter your SDRs can trust.