How a Tiny Team Ships Like a Big One · The Product Track
Reading an industry's public exhaust
The buying signals enterprises pay for are lying in public - filings, earnings calls, job postings, regulatory actions - in volumes no human team can read. So we stopped using humans.
During one customer's evaluation of our platform, the system flagged a prospect with a high buying-intent score - about 80% - for a product line the customer sold. The sales team looked it up and got quiet for a second: they already had an RFP in flight from that exact company. The model had never seen the RFP. It had read the same public record the prospect's own actions were part of, and arrived at the same place.
That's the moment this line of the product clicked for people, so it's the right place to start the post. Not because the system is psychic - because the signals were public the whole time. Somebody just had to read everything.
The problem is volume, not access
The events that make an enterprise ready to buy something rarely happen in private. A regulator issues a consent order, and suddenly a bank needs better monitoring. An earnings call admits a modernization program is behind. A company posts eleven job listings for a skill it never hired before. A merger closes, and two incompatible systems now share one roof - we watched an analysis flag a strong product fit for one customer driven by exactly that: a multi-bank merger plus an existing product relationship.
None of this is secret. All of it is public record: filings, earnings transcripts, press releases, job postings, regulatory actions, news. The reason sales teams still discover these things late - through word of mouth, or a lucky Google - is that the reading load is inhuman. A single rep covering global banks would need to continuously monitor thousands of documents a quarter, across a hundred accounts, forever. So historically, "signal monitoring" meant a spreadsheet, a news alert, and guilt.
What we built
Continuous readers instead of quarterly research sprints. Agents monitor those public source categories for the accounts our customers care about, extract candidate signals, and - this is the part that took the longest - decide which ones actually mean something for a specific seller with a specific product.
The hard-won lesson is that extraction is the easy half. Any competent pipeline can find "Company X announced Y." The value lives in the judgment layer: is this signal new, is it material for what you sell, does it converge with other independent evidence, and is it something the seller can act on this quarter? A signal that just restates a public fact anyone's chatbot could fetch is noise wearing a suit. Our filter's bias is convergence - independent signals pointing at the same conclusion beat any single loud one.
From there, the output isn't a feed. Feeds die unread. Each surfaced signal arrives attached to an account, a use case, and a reason - "this account, this quarter, because these things happened" - in the places sellers already look.
Real numbers
- An 80% buying-intent score on a prospect that turned out to have an RFP already in flight - validated against pipeline the model had never seen.
- A merger-driven product-fit flag confirmed by the customer's own account team.
- Public sources monitored continuously: filings, earnings transcripts, press releases, job postings, regulatory actions, news.
- Hundreds of accounts watched at once for a single sales team - the coverage a rep can't get back by working weekends.
Where the humans sit
The system proposes; sellers judge. Every surfaced signal shows its evidence - which documents, which dates - so a rep can disagree with the conclusion in thirty seconds instead of taking it on faith. Every claim in the recommended opening and the summary carries a citation, and hovering one opens the evidence behind it: the document, its date, the sentence the model actually read, and a link out to the original.
And customer feedback flows back the way everything does here: "this signal was wrong" is an input we tune on, not a support ticket we apologize for.
What stays out of this post is the taxonomy itself - which signals we track, how they're weighted, what makes a score cross a threshold. That judgment layer is the product.
Steal this
If you're building anything in this space, start with the one source category whose semantics are unambiguous: regulatory actions. A consent order means what it says, it's dated, it names the institution, and its implications for what that institution must now buy are legible. Prove your precision there, where ground truth is checkable, before you touch the murkier categories - job postings and news will happily bury you in maybes. And output reasons, never feeds.
Next week the build track covers how 100% of our code came to be written by AI. Then back here: the agents doing all this reading learned to manage themselves - including knowing when to stop.
This post is part of the product track of How a Tiny Team Ships Like a Big One, a series on how six builders run a production AI company. Building at Aithon - if this is how you want to work, talk to us.