I have done manual SDR work. I know what it feels like to spend an hour researching a single prospect — reading their LinkedIn, their company blog, their press releases — and then write a four-line email that either gets ignored or bounced. At scale, this is unsustainable. Most reps do not spend the hour. They send a template with a merge field and a pain point that does not quite fit. The prospect can tell. The reply rate reflects it.
I spent a weekend building an alternative. The system is not perfect, but it produces research-backed, personalised outreach at a speed no human SDR team could match, and the reply rates are meaningfully better than template campaigns.
The Problem with Manual SDR Work
The math is simple and brutal. A good SDR can research and write 15-20 personalised emails per day. An enterprise sales cycle needs a pipeline of 80-100 active prospects to reliably hit quota. At 20 emails per day, building that pipeline from scratch takes 4-5 days of pure outreach work — before you have had a single conversation. And that is a good SDR working efficiently.
Most SDR teams respond to this math by sacrificing personalisation. Templates go out. Open rates stay high because subject lines are optimised. Reply rates suffer because the email clearly does not reflect any knowledge of the prospect's actual situation. You get a lot of "not interested" responses from people who might have been interested if you had known what to say.
The Stack
The system uses three tools in sequence:
- Apollo.io for prospect sourcing. Apollo has company technographic data, intent signals, and contact-level information that you can filter against your ICP. I set up a filter for my target persona — IT decision-makers at mid-market companies using legacy ITSM tools — and pull 50 leads per run.
- LangGraph for the pipeline orchestration. LangGraph is a Python framework for building agent workflows where each step can branch, loop, or call other agents. It is the right tool for a multi-step process where each step needs to be independently configurable.
- Claude claude-haiku-4-5 for the research synthesis and email drafting. Haiku is fast and cheap enough to run against every lead in the batch without the costs compounding into something unsustainable.
Before vs. After
The Scoring Model
The ICP scoring step is the most important part of the pipeline and the part most people skip. The score runs on a 0-10 scale across four dimensions: company fit (size, industry, growth stage), technology fit (are they using tools my solution integrates with or replaces?), role fit (is this person actually a decision-maker?), and timing fit (are there signals — hiring, expansion, recent press — that suggest they are in-market?).
Leads that score below 6 go to a cold nurture sequence. Leads that score 6-7 get a templated email with light personalisation. Leads that score 8+ get the full Claude-researched email. This tiering is what drives the reply rate improvement on Tier A — you are not wasting personalisation budget on cold leads who are never going to engage.
The scoring model is also how you improve the pipeline over time. Every reply — positive or negative — is feedback that updates the model's calibration. After three months, the ICP scores are meaningfully better calibrated than they were at launch.
What Still Needs Humans
The system is not hands-off. There is a human review queue that every email passes through before sending. Fifteen minutes per batch reviewing 50 emails catches the edge cases — prospects the system scored incorrectly, email drafts where the personalisation feels off, contacts who have already been reached previously. The automation handles the research and drafting; the human handles the judgement calls.
The pipeline also cannot replicate relationship-based outreach. If you have a warm introduction, a shared connection, or a mutual customer reference, that context matters more than any AI-researched email. The system is for cold outreach at scale. It does not replace the relationship network — it frees up time to build it.