A multi-stage European fund I work with now runs a real slice of its deal-flow loop through a single Slack message. A partner drops in a request, and a short moment later the answer comes back — a new pitch scored against the fund's own portfolio, with the reasoning attached. No new sourcing platform. No CRM migration. Nothing the team had to learn.
Let me be honest about the status before I sell you anything: this is a partner-only rollout, live and running as of May 2026, not a wrapped case study with a headline number. It's a build log. Here's what's running today, here's how it's wired, and here's what's still in progress.
I'm Michael. I run Black Matter VC — a solo, operator-led AI studio and consultancy that builds and operates these systems for funds. I don't sell sourcing software. I wire signal into the tools funds already pay for.
IWhat's running today: one Slack call instead of a manual analyst pass
The old version of this was a person. A new pitch call would finish, someone would pull the transcript, read it end to end, hold it up against everything the fund already knew, and write up a view. Useful work — and a slow, manual one.
What runs now: a partner asks in Slack, a custom scoring skill reads the new call transcript, scores it against the fund's portfolio signal, and returns the read — with cited reasoning — back into the same Slack thread. The first end-to-end runs already replace that manual analyst pass with a single call. It's capped to a small group of partners for now, on purpose, and I'll come back to why.
IIThe problem: more pitch transcripts than partner hours
The framing came straight from the fund, and it's the honest one: they had more pitch transcripts than partner hours, and a portfolio that's the only honest comparison set. The signal to judge a new deal against wasn't sitting in some third-party database — it was already in the fund's own history of calls, notes, and companies.
We'd tried the obvious thing first: a single-thread chat agent. It was fine for a demo and brittle the moment two people used it in parallel. One partner's request would step on another's. That's the gap between a thing that looks good in a screenshot and a thing a partnership can actually lean on during a busy week.
IIIWhy we didn't sell them another sourcing platform
The constraint was set on day one: the partnership had no appetite for a new tool. This is the same fund we'd already built a workflow layer we built for a multi-stage European fund for — a two-way sync between their CRM and Notion, plus a network-search bot — precisely because everything had to slot into the stack they already paid for. That earlier build has run in production for years.
That matters, because it runs against the grain of what you'll find if you search AI deal sourcing. The results are almost entirely platform-sales pages — buy this discovery engine, add these seats, migrate your data over here. Reasonable products. But most funds don't have a discovery problem so much as a judgment-at-speed problem: too much inbound, not enough hours to score it against what they already know. You don't fix that by adding a tenth login.
The top funds seem to agree. A Bain Capital principal told Affinity that what used to be "a 5–10 minute, multi-screen, multi-app workflow after every meeting is now one 30 second natural language request" — the model handles the prose and the tags and posts it straight into the CRM the team already uses. That's the pattern. Wire the intelligence into the existing surface; don't bolt on another surface.
IVInside the architecture: planner, sub-agent, and a scoring skill
Here's the actual shape of the build, in plain terms.
A Claude Opus planner sits at the front. When a partner sends a request in Slack, the planner acknowledges it instantly — no dead air while something churns in the background. Anything that's going to take real time, like scraping and research, it hands off to a Claude Sonnet sub-agent that goes and does the slow work.
The scoring itself is a custom skill. It reads each new call transcript and scores it against the fund's portfolio signal, pulling from the fund's existing transcript store rather than a new database we stood up. No data got relocated. The comparison set is the fund's own history, which is exactly the point — a generic model can tell you a company looks interesting; the fund's own portfolio can tell you whether it's interesting to them.
aWhy hot-path reads bypass the planner
One detail separates a production system from a demo here. Not every request needs the planner. A simple lookup — a fast read of something the system already has — goes direct to the API instead of routing through the planner first.
Why bother? Because when three partners hit the bridge at once, you don't want a quick lookup stuck in line behind someone's long research job. Sending the easy reads straight down a fast path keeps the whole thing responsive under real, parallel load. It's the unglamorous plumbing that decides whether people keep using a tool after week one.
VWiring signal into systems the fund already runs
This build didn't come out of nowhere. It's the same pattern behind the productized tools I already run across client funds — the pieces that make "no new CRM" a real promise rather than a slogan.
The Pitch Deck Scanner is the clearest example: it auto-routes every incoming pitch deck into a fund's CRM. It detects decks arriving by email — PDFs, DocSend links, even password-protected DocSend chains — extracts the company details, researches the founder, scores the deck against the fund's written investment thesis with cited reasoning, and writes structured rows straight into Affinity, Attio, HubSpot, or any CRM reachable through Zapier. No new system, no human babysitting the download step.
The Affinity Chat Bot does the same trick from the other direction: it lets a team query the CRM in plain English from Slack — companies, deals, co-investors, notes, network data — without anyone learning Affinity's UI, deployable either as a Slack-native bot for the team or a Custom GPT for an individual partner. And LinkedIn Network Search can answer "do we know anyone at X?" in about five seconds by indexing the whole team's combined connections and scoring the warmest intro path — kept in the fund's own infrastructure, not a third party's.
Two-way CRM sync, Slack-native delivery, scoring against a written thesis. Those are the building blocks. The transcript-scoring bridge is just the newest thing assembled from them.
VIWhat this replaces, in practice
The honest way to describe the change is job by job — what a person did before, and what runs now.
None of this replaces the partner's judgment. It replaces the shuffling around the judgment — the reading, the cross-referencing, the copy-paste into the CRM — so the scarce resource (partner attention) lands on the decision instead of the prep.
VIIWhy sourcing speed is the real battleground right now
The pressure behind all of this is real, and it's measurable.
In a 2026 survey of nearly 300 private-capital dealmakers, Affinity found that 85% now use AI to automate daily tasks — up from 76% a year earlier — and 82% use it specifically to research companies for deal sourcing, an 18-point jump from 64% the year before. Affinity frames that as "a change from experimentation to standard practice."
The money explains the urgency. AI and machine-learning deals took 65.6% of all US venture funding in 2025, up from 47.2% the year before, per the PitchBook-NVCA Venture Monitor. When capital concentrates that hard into one category, funds compete less on whether they can write the check and more on whether they saw the company first.
And most still miss most of it. Research from Sutton Place Strategies, cited by Zenit Data, puts the median private-equity firm's captured share of relevant deal flow at only about 18% — meaning more than 80% of potential opportunities never reach the investment team when sourcing leans on shared intermediary channels and reactive timing instead of continuous signal.
aHow well-funded platforms are attacking the same problem
The best-resourced funds aren't just wiring signal into an existing CRM — some are buying or building entire discovery engines. It's worth knowing the buy-side options before you decide to wire your own.
These are strong tools, and Evertrace in particular shows the direction of travel: even the discovery platforms now sync into Slack and the CRM rather than asking you to live somewhere new.
VIIIThe honest verdict: when to buy a sourcing platform vs. wire your own signal
Here's the decision I'd actually give a fund, without the hype.
Buy when your problem is discovery — finding companies you don't already know about, out in the public universe of stealth and pre-seed founders. That's a genuinely hard data problem, and platforms like Harmonic, SignalFire, and Evertrace have spent years and a lot of capital solving it. You will not out-scrape a company indexing 30 or 80 million organizations, and you shouldn't try.
Wire your own when your edge is in data you already hold: pitch transcripts, a portfolio that's your only honest comparison set, a written thesis, and a CRM full of history. That signal is proprietary to you, it never appears in anyone's public index, and scoring inbound flow against it is exactly the kind of thing a small, well-plumbed agent setup does well — inside the systems you already run.
Most funds want both, in that order: a discovery platform for reach, and a thin intelligence layer that scores whatever lands against the fund's own judgment. The mistake is treating "add a sourcing platform" as the whole answer when half the value is in the data already sitting in your CRM.
And separate what's proven from what's still cooking. The wire-into-existing-systems pattern — CRM sync, Slack delivery, thesis scoring — is proven and running in production across funds, some of it for years. The transcript-scoring agent bridge is newer, and I'm treating it as a live build, not a finished result.
IXWhat's next
What's shipped: the planner-and-sub-agent bridge, the transcript-scoring skill, hot-path reads for responsiveness, and the first end-to-end runs replacing a manual analyst pass — all live for a capped group of partners.
What's in progress: full-team rollout, which is gated on sub-agent isolation. That single-thread brittleness I mentioned earlier is the whole reason. Before the entire partnership hits this in parallel, each partner's long-running job needs to be properly isolated so nobody's request steps on anyone else's. That's the work between "partner-only" and "everyone." When it's done, I'll say so plainly.
XFAQ
aDo we have to replace our CRM to do this?
No — that's the entire point. Everything wires into the CRM you already run. The Pitch Deck Scanner writes structured rows straight into Affinity, Attio, HubSpot, or any CRM reachable through Zapier, and the Affinity Chat Bot puts a plain-English layer over your existing Affinity data from Slack or a Custom GPT. Your data stays where it is; the intelligence comes to it.
bHow long does wiring AI deal-sourcing signal into Slack and our CRM take?
It depends on scope, but the productized pieces set a useful benchmark. LinkedIn Network Search, for example, deploys in roughly a week per fund — LinkedIn export, indexing, and Slack bot setup included. A bespoke agent bridge like the transcript-scoring one takes longer and gets rolled out in stages, which is why it starts partner-only. I run these on a build-and-operate model, so it's not a one-time integration you're then left to maintain alone.
cWhat does this cost compared to buying a sourcing platform?
I work on a flat $10K/month build-and-operate retainer — that covers building the system and keeping it running, not just a hand-off. Sourcing platforms typically price per seat on an annual subscription, so their cost scales with the size of your team and the tier you're on. Directionally, the retainer is a fixed monthly line whether one partner or the whole team uses it; a platform is a per-seat spend that grows as you add people. They also solve different problems — one discovers companies you don't know, the other scores what you already see against your own thesis — so for a lot of funds it's not either/or.
— Michael Rouveure
