most sourcing automation makes the problem worse, because it automates the wrong stage.
the bottleneck in sourcing was never finding more companies. you can find ten thousand this afternoon. the bottleneck is that reviewing them costs attention you do not have, so a pipeline that surfaces more is a pipeline that gets ignored faster.
so the useful frame is not "automate sourcing". it is: which stages can a machine do better than you, and which one must stay yours.
the five stages
1. detection, automate fully. something changed: an investor started following a young account, a founder's title went to stealth, a domain got registered. no human should be watching for this and no human can watch at the required rate.
2. enrichment, automate fully. who is this person, where did they work, does the career support the claim, has anyone else noticed. mechanical, and an agent with typed tools does it faster and more consistently than a person with tabs.
3. filtering, automate, and this is where the value is. score against a written thesis. not "interesting companies" but the actual sentence: pre-seed, technical founder, applied autonomy or robotics, north america, no announced round.
this stage is the whole thing. of 2,549 people we flag building in stealth, 896 survive a career check. the other 64% are not a smaller opportunity, they are noise, and an agent handed all 2,549 will rank the noise with complete confidence.
4. drafting, automate with constraints. an agent writes a better first message than a template, because it can reference the specific prior work. it needs explicit rules, which is the next section.
5. the decision to reach out, keep. this is not modesty about ai. it is that the agent has no access to the thing that decides it: whether you want to own this, whether it fits the fund, whether you have conviction. ask an agent for that and it will give you a confident answer built from nothing.
the four failure modes
optimising for volume. more sources means more reaching you, not more acted on. every unfiltered source raises review cost. the metric worth watching is how many surfaced companies you actually contacted, and adding sources usually makes that number worse.
no written thesis. an agent cannot filter against a vibe. if the thesis is not a sentence you could hand a stranger, the agent will substitute plausibility for fit and you will get a list of good-looking companies you would never back.
letting it explain how it found them. an agent drafting outreach will happily write "I noticed several investors recently started following you". that is technically true, public information, and reads as surveillance. state the rule explicitly: never explain the method, lead with the founder's prior work.
believing the confident answer. the agent will state a number no tool returned. the defence is typed tools and one question: which call produced that? a system built on scraped text cannot answer it.
the loop, concretely
connect the agent to a signal source, then run this on a schedule:
surface accounts with 3 or more new tracked-investor follows this week, and people flagged building in stealth with a strong verdict. enrich each with founders and funding. drop anything with an announced round or a background outside the thesis below. rank what remains, and for the top three give me the evidence and a drafted first message that does not reference how they were found.
thesis: pre-seed, technical founders, applied autonomy or robotics, north america.
what comes back is three names with reasons. what you do with them is still yours.
what the machine is genuinely better at
worth being clear, because the honest case is strong enough:
- it never gets bored on row 400
- it applies the same bar on friday as on monday
- it checks every career, not the ones that look interesting
- it reads the whole feed, which no person does
what it is worse at is everything downstream of "is this worth owning", and no amount of tooling moves that line.
related
open source deal flow
companies and founders surfaced on frontrun. curated, every week.
