I build the pipelines that sit underneath sourcing teams, so when a firm asks me to look at the AI deal sourcing tool they're evaluating, I see the part of the stack the salesperson skipped. Usually it's a chat interface, an embeddings index over the vendor's existing company table, and a system prompt. That's a product. It isn't what the marketing implies.
The honest split is whether the model is touching the part where ranking decisions get made, or decorating the part where they already got made by a vendor's category taxonomy six months ago.
What the SERP is actually selling
Search the term and the front page is wall-to-wall vendor copy. AI that "reads like an analyst." AI that "thinks like an M&A analyst." AI that "scores and classifies deals instantly, based on your fund's thesis." The pitches are interchangeable.
Under the hood, almost all of it is the same architecture. A fixed company dataset the vendor has been selling for years. An embedding model that turns a company description into a vector, plus a retrieval step returning nearest neighbors to whatever you typed. Sometimes an LLM in front translating "midwest HVAC roll-up candidates" into the filter the platform already supported. None of this is fake AI. It just isn't doing the work the buyer thinks they're paying for.
The marketing implies the model is finding companies competitors can't see, ranking on signals nobody else is pulling, and learning from outreach outcomes. On most page-one platforms it's doing none of those three. It's helping you write the same filter your associate would have written, against the same dataset twenty other firms are querying this week.
What the model layer has to do
Three layers have to be inside the system before the label stops being a paint job.
Signal discovery. A feature store the firm controls, populated by extractors pulling raw primary data (permit feeds, license renewals, role-change cohorts, web stack diffs, hiring posts, court records), and a training routine that fits weights on those features against the firm's actual outreach outcomes. The model's job is to find which features track conversion, not apply a ranking the vendor handed it. A scorer that ranks against the eight columns a vendor exposes isn't discovering anything. It's the same SQL with a softmax on top.
Scoring iteration. The score has to change as labels come back. Meeting taken, declined, "not for another year," bounce, no response after four touches: each is a label, and a real system retrains on them on a schedule. That means a label store, a retraining job (nightly or weekly), and a versioned model so you can tell whether last quarter's ranking beat this quarter's. If the model ranking your list in January is the same one ranking it in July, no learning is happening.
Automated enrichment. The system has to reach into primary sources and produce structured fields that didn't exist before. A founder's tenure parsed from LinkedIn, handling the eight ways the date renders. A license renewal pulled from a state portal with no API. A CMS migration inferred from a sitemap diff. Plain extraction work, with an LLM for the messy pieces and a deterministic parser for the rest. Enrichment that renames fields already in the vendor's schema isn't enrichment. It's a relabel.
Take any product calling itself AI deal sourcing and ask which of the three it actually does. Most do one, some do zero, and the few that do all three aren't the loudest names on the SERP.
The wrapper tell
The cleanest test is to ask the vendor a question their dataset can't answer.
"Show me independently-owned HVAC operators in the Midwest where the founder is over fifty-five, hiring has stopped for nine months, and a CMS migration happened in the last two quarters."
A wrapper returns a near-empty result or quietly drops two of the three filters. A real system returns a ranked list with the founder-age signal from LinkedIn, the hiring trend from a job-board feed, the CMS signal from a sitemap crawl, each row showing which extractors fired. If the vendor can't show evidence per row, the model isn't pulling it. It's guessing on the company name.
What scoring iteration looks like when it works
The loop is unglamorous. Extractors per source. A feature store the firm controls. A gradient-boosted scorer or something similarly boring. A label-ingestion endpoint reading outreach outcomes from the CRM. A retraining job on a schedule. None of those components are research; they're plumbing, and the work is in keeping them honest.
What changes when the loop runs for a few months is which features survive. The signals a sourcing team walks in convinced about — size band, recent senior hires, geography — are not always the ones that end up carrying weight. Sometimes the intuition holds. Sometimes the coefficient comes back small or negative and the human has to update. That update only happens because the model was fit against outcomes the CRM logged, not against the team's prior about what should matter.
The features that tend to outperform the vendor schema aren't exotic. They are the ones you can only build by reaching into primary sources the vendor doesn't index — permit and license feeds, hiring-cadence series from raw job boards, sitemap and web-stack diffs, referral-graph proximity computed from the firm's own contact history. Whether any specific feature lands is an empirical question for each firm and each thesis, which is the point. The vendor cannot answer it for you.
The artifact at the end isn't the initial list. It's the loop: extractors feeding the feature store, the CRM feeding the label store, the retrainer fitting on a schedule, the firm owning every piece. No chat box required.
Why this matters at the renewal
A wrapper on someone else's data has a structural ceiling. The vendor owns the dataset and decides which signals you can filter against, and the model is shared across every customer, so your top decile is also every competing buyer's top decile on the same platform. When the contract ends, none of the labels you generated stay with the firm. Not the meetings, not the declines, not the disqualifications. The vendor keeps the most valuable asset you produced: the outcome data that would have trained your next model.
The owned version compounds the other direction. Each campaign leaves the firm with more verified contacts, more outcome labels, more signal-weight calibration. Year two doesn't start from zero.
So who's actually building the model layer
Honest read from the data side, looking at what these tools do on a test set instead of in a demo. Platforms that index company data and put a fast retrieval layer on top (Grata, Sourcescrub, Cyndx) are real products, and the embeddings work. The "AI" part is search relevance. Useful. Not a scoring system. Reuben and the newer entrants pitching agentic workflows are orchestration over the same data, with an LLM doing what an associate used to do with filters. The vendors actually doing the model layer publish per-firm performance, expose a label-ingestion API, and let you bring your own extractors. There aren't many, and they tend not to be the loudest names on the SERP.
If the AI in your sourcing stack isn't touching signal discovery, isn't retraining on your outcomes, and isn't pulling structured fields from primary sources, the label is doing the work the product isn't. Don't pay for a model when what's shipping is a search box.
