What AI prospecting actually costs per qualified lead (our ~$1 run, line by line)
Nobody publishes what AI prospecting costs per qualified lead in tokens, searches and page fetches, so here is ours, measured on live runs. Where the money goes stage by stage, the line item that dominates, and what we got wrong the first time.
Aman Jha··5 min readCredits. Every AI prospecting tool prices in credits, every pricing page has a slider, and nobody will tell you what AI prospecting actually costs per qualified lead once you translate a credit into tokens, searches and page fetches.
I get why. It's messy. So here are our numbers with the messy parts left in, and this time they're measured, not estimated. The first version of this post said $4.80 a run. That was a model on paper. When we ran the real pipeline end to end and metered every call, it came in at about a fifth of that, and the money went somewhere I didn't expect.
A run (20 researched companies) costs us about $1: we measured $0.77 to $1.16 across four live runs. The models are about $0.21 of that. The rest is web search. Finding a named buyer at each company is optional and adds about $0.90.
AI prospecting cost by stage, with numbers
Assumptions first. Every model call uses one small model (gpt-5.4-nano, $0.20 per million input tokens and $1.25 output). Page reading is self-hosted (Firecrawl plus SearXNG), so the marginal crawl cost is the server. Search is Exa at list price: $7 per thousand searches plus $1 per thousand pages of text, capped at 110 searches a run. Prices as of October 2026; re-check before you copy them.
| Stage | Volume | What it spends on | Cost |
|---|---|---|---|
| Analyze your site and infer the ICP | ~10 pages, 3 calls | model | ~$0.01 |
| Discover | 100–230 candidates | search | ~$0.30 |
| Screen | 100–230 homepages | rules + a cheap classifier | under $0.01 |
| Qualify | 50 companies, ~4 pages each | model + company size lookups | ~$0.40 |
| News and timing check | 20–30 companies | search | ~$0.25 |
| Write the cards | 20 calls | model | ~$0.05 |
| Total | ≈ $1.00 | ||
| Buyer lookup (optional) | 20 people + email checks | people search | ~$0.90 |
Two things jump out.
First, the model bill is small and boring. About 140 calls, half a million tokens in, $0.21 a run. Search is three quarters of the cost. If you're optimising, count your searches, not your tokens.
Second, screening is nearly free, and that's a design decision, not luck. Most candidates are decided by rules (wrong country, a careers page instead of a homepage, a known competitor) or by a cheap yes/no classifier. Only what's left is read properly.
What we got wrong the first time
The $4.80 model assumed a mid-tier model for qualifying and a model call for every homepage screened. Both turned out to be the wrong place to spend:
- A small model qualifies well if you give it the right job. We don't ask it "is this a good prospect?" We ask narrow questions against quoted text ("does this page say they run their own stock?") and let deterministic rules do the scoring. Narrow questions don't need a big model.
- Screening doesn't need a model at all for most candidates. Rules handle the obvious ones; a cheap classifier handles the rest.
- Search was the cost we hadn't modelled. The paper version assumed free self-hosted search. In practice the sources that find good companies (lookalikes, people search, news) are paid, so we cap searches per run and cache results for a week.
The honest lesson: our estimate was off by 5x, and in the direction that flatters nobody. We'd priced for the expensive thing (tokens) and ignored the one that actually costs money (searches).
What it costs per lead
- Per researched company on the list: about $0.05 (a $1 run, 20 companies).
- With a named buyer: about $0.10 per company. Not every company gets one; on our recent lists about two in three did.
- With a verified email: the buyer lookup checks the address with a free SMTP check. Catch-all domains can't be confirmed that way; a paid verifier closes some of that gap.
The time is the real cost now: 17 to 43 minutes a run, most of it reading pages politely.
How that compares
Clay is the interesting comparison. Claygent, their per-row web research agent, runs about 7 credits per company. On the $185 Launch plan that's somewhere around a dollar per researched company before you count the enrichment credits around it, and you build the workflow yourself. Our number for a company that's actually been read and checked, with a buyer, is about a dime. The difference is that we only do one thing.
The number that tells us to stop
Every cost model needs a kill signal. Ours:
If getting 20 good companies takes more than two runs' worth of searches, the problem isn't cost. It's that discovery is looking in the wrong place.
We hit exactly this on one run: the search phrases described the seller's own product category, so discovery found the seller's competitors and only 6 of 170 candidates survived. The fix wasn't to spend more. It was to search for what the buyer builds, not what the seller sells, which is the whole argument in how to build a target account list from your website: start from places where membership already proves something, so verification is confirming, not discovering.
What I'd tell you if you were building this
Four things, from the mistakes:
- Meter every call before you trust a cost model. Ours was 5x off, and we'd have priced the product on it.
- Count searches, not tokens. With small models, search is the bill.
- Ask models narrow questions and let rules score. It's cheaper, and it's the only way to show your evidence.
- Write down the kill signal before launch. Once there's revenue you'll rationalise anything.
And if you'd rather not build it: send us your website and we'll run it for you.