Scoring B2B leads with an LLM, and where it fell over

I built a lead scoring system for my own outreach because the scores I was getting out of the box were not usable. It is called the ANTA Lead Intelligence Agent, it runs in production, and I use it every week on my own pipeline before I ever sold it to anyone.

That last part is the reason this post is worth reading. I am not describing a client engagement from a safe distance. I am describing the thing I depend on, which means I can tell you precisely where it falls over.

The problem with off-the-shelf lead scores

Generic scoring products weight what they know about everyone: headcount, industry code, revenue band, technologies detected, funding events. Those are real signals. They are also the same signals every other vendor's customer is scoring on, which means acting on them puts you in the same queue as everyone else, contacting the same companies in the same week.

Worse, for my business they were weighting the wrong things. A generic scorer loves a recent funding round. But a company that just raised is often in the middle of hiring a team to do internally exactly what I would be proposing to do for them. Meanwhile a bootstrapped 20 person company that just posted a job listing for an accounts payable admin whose whole day is processing documents is, for me, a far better prospect. No generic scorer will ever rank it that way, because that judgement is specific to what I sell.

That gap is the entire thesis of the tool. The firmographics are table stakes and a data vendor can supply them. The judgement about what those facts mean for my offer is the part that has to be encoded, and that turns out to be a language problem rather than a numeric one.

What actually replaced it

The rough shape, in order:

1. Source the raw list. Companies and contacts pulled from a data provider, filtered on the coarse firmographic criteria that a database is genuinely good at: country, headcount band, industry.
2. Gather the unstructured context. The parts that carry the real signal and do not fit in a column: job postings, site copy, what the company says it does versus what its hiring implies it is struggling with.
3. Score against a written profile, not a formula. This is the actual change. Rather than assign point values to fields, the model reads the context against a written definition of a good fit, including an explicit anti-profile of who to reject, and returns a score with a stated reason.
4. Generate the outreach off the reason, not the company. The email is drafted from the specific evidence that produced the score, which is what makes it read as researched rather than merged.
5. Sequence and hand off. Ordering, timing, and the record of what went where.

The critical design decision is number three, and specifically that the model must return why. A bare number is unauditable. A number attached to "this company posted a role whose described duties are document intake and manual data entry" is a claim I can check in ten seconds, and reject when it is wrong.

Where it fell over

Three failure modes, all of which showed up on real accounts.

It was agreeable. Early on, the scores clustered high. Ask a model to evaluate fit and it will find a way to argue for fit, because that is the shape of the request. Two things helped: writing an explicit anti-profile, an actual list of who is not a customer and why, and forcing the output into a fixed structure with a required disqualifying-evidence field, so the model has to look for reasons against rather than only reasons for. The scores still skew generous. I read the reasoning, not the number.

It hallucinated plausible context. Given a thin company page, it would generate a confident description of what the company does that was mostly inference. On a scoring pass that is tolerable. On an outreach draft it is lethal, because a cold email containing a specific and wrong claim about someone's business is worse than no email. The fix was structural rather than a prompt tweak: the draft may only reference evidence that was actually passed into it, and anything not in the source context does not get to appear in the email.

The unstructured signal is often just missing. The job posting angle works beautifully when there is a job posting. Plenty of good-fit companies are not hiring this month, and for those the tool has little more to work with than the generic scorer did. I have not solved this. It means the system is good at ranking the companies that are broadcasting a signal and roughly average at everything else, which is a real limitation and not a temporary one.

What I cannot claim

I am going to be exact here, because this is where most posts like this start inventing numbers.

I have not proven this converts better. I have sent a small number of sequences off it, in the dozens, not the thousands, and I do not have a statistically meaningful reply rate to report. Anyone quoting a precise uplift percentage off a sample that size is quoting noise.

What I can say is narrower and true: the scoring surfaces companies that the firmographic filter ranked as unremarkable, and the reasoning it attaches is specific enough to check. The sharpest email I have sent came from a vertical I had not originally targeted at all, which the system flagged off a job posting. That is a real outcome, and it is an anecdote, and I am labelling it as one.

The other honest caveat: it currently runs as a single-user beta on a free hosting tier. It is not multi-tenant, so there is no self-serve login to hand you. If you want to see it, I will walk you through it live rather than send you a link that does not work.

What transfers to your version

If you build something like this, three things carry over regardless of industry.

Write the profile before writing the prompt. Most of the quality lives in how precisely you can describe a good customer, including the anti-profile. That document is the asset. The code around it is comparatively easy.

Require the reasoning as a first-class output. Not for interpretability theatre. Because it is the only way you will catch the model being confidently wrong, and it is what turns a score into something a salesperson can open an email with.

Never let the outreach draft exceed the evidence. Constrain it structurally so it cannot reference facts that were not supplied. Prompting it to be accurate is not the same as making it unable to invent.

If you want the architecture for a version pointed at your own ICP, the Scope it live tool will produce one, or read what a fixed-price pilot includes if you would rather know the commercial shape first.