7 steps · 鹦鹉AI
AI lead generation with 鹦鹉AI: a verified B2B lead list from one prompt
Give your agent one prompt and get back a lead list with a named person, a verified work email and a scored reason to write, for every company that fits. 7 steps through 鹦鹉AI: jev judges each company on the fields the list already returned, before the paid person, email and news lookups, and scores the opener at the end. The price is printed before the agent spends it, and the numbers on this page come from running it, not from a rate card.
One prepaid balance · no per-tool signup · no card
Try it
One prompt runs the whole thing
Set your agent up, once
set up data capabilities — https://parrot.amorist.cn/llms.txt
Then ask for it
Using 鹦鹉AI, build me a lead list: 50 US software companies with 51 to 200 staff that raised a Series A. Before you find anyone, have jev judge each company against my ICP (B2B software that sells to sales or marketing teams) from the list fields alone, and drop anything under 50%. For the rest find the VP or Head of Marketing, find their work email with the cheapest tool that only bills on a hit, verify it, and pull the latest news. Have jev pick the event to lead with and score how usable it is as a first line. Show me the total price before each step, and give me a CSV at the end, strongest opener first.
One list in, one CSV out
The agent carries the domain from step to step. You never paste anything twice.
Judge before you spend
A jev verdict costs a fraction of the person step. Name the threshold and the agent only pays for rows that pass.
Only bill on a hit
Most email finders' rate cards charge nothing for a miss. Say so and the agent picks one.
Ask for the price before each step
Every step is metered per call, so the agent can show the bill before it spends.
The steps
What 鹦鹉 AI calls, and what each call costs
Prices come from the catalog's own rate book, read when this page loads, on one prepaid balance.
| # | Step | What the agent asks | Price |
|---|---|---|---|
| 1 | Build the company listApollo bills per page, not per company, so one page of 50 is one charge. | 50 US software companies, 51 to 200 staff, latest round Series A | 18.2 credits/call |
| 2 | Judge the fit with jevApollo's list page carries the domain, NAICS and SIC codes, printed revenue and headcount growth for nothing extra. jev reads those and returns a probability, so a company that fails never reaches the paid lookups. | does this company fit the ICP closely enough to pay for a person and an email; keep it at 50% or more | not in the catalog (runs on your own key) |
| 3 | Find the personFindymail's rate card bills per contact returned. LeadMagic's role finder is the fallback, and it found a person at every company Findymail missed. | the VP or Head of Marketing, or Head of Growth, at each company | 13.86 credits/result |
| 4 | Find the work emailTomba is the cheapest per-success finder in the catalog and was out of capacity on 鹦鹉AI's key for this run, so Hunter served, and Kitt ran on Hunter's misses. A miss is free at all three. | their work email, cheapest tool that only bills on a hit | 6.23 credits/success |
| 5 | Verify itLeadMagic bills a quarter credit per definitive verdict and nothing for an unknown. | drop anything not deliverable, keep the unknowns separate | 4.38 credits/success |
| 6 | Find an openerPredictLeads bills per call for classified events. Akta is cheaper per call but was out of credit on the first run, so both runs used PredictLeads. | the three most recent news events about each company | 28.01 credits/call |
| 7 | Score the opener with jevThe agent drafts nothing here. jev ranks the events PredictLeads returned and scores the best one, so the CSV sorts by opener strength and a weak one is flagged before anyone writes. | which of the events to lead with, and how usable it is as the first line of a cold email, on a four-level scale | not in the catalog (runs on your own key) |
At the rates above, 50 rows where every call hits comes to 2642.20 credits. The receipt below is what it actually cost.
The receipt
What it actually cost, in credits
Run on 2026-09-23, 50 companies in.
- Companies matched
- 958 on Apollo; the first page of 50 taken, one charge of 18.20 credits
- Rows with a usable domain
- 48 of 50
- Passed the jev gate
- 27 of 48 at 50% or more; 21 dropped before any paid lookup; 14 would have passed at 60%
- A named marketing lead found
- 27 of 27 (20 by Findymail, 7 by LeadMagic's role finder)
- Work email found
- 21 of 27 (18 by Hunter, 3 by Kitt on Hunter's misses; Tomba out of capacity, see below)
- Verified deliverable
- 20 of 21; 1 unknown; 0 invalid
- A news event in the last year
- 19 of 21 (PredictLeads)
- Opener scored by jev
- 19 of 19; 4 scored decent or better on a 0 to 3 scale, mean 1.79
- jev, both steps
- 67 verdicts on 55,009 input tokens, 1.61 credits at list price; the team's own key, so not metered
- Wall clock, four calls in parallel
- about 5 minutes; the gate took 90 seconds for 48 companies
- Total metered
- 1631.01 credits for 50 companies, or 84 credits per deliverable lead; the 21 dropped rows would have cost about 1253 credits more at this run's rate
Every number above is what 鹦鹉AI's ledger settled on 2026-09-23 for this run, not a rate-card estimate. The gate is the change from the first run of this workflow on 2026-08-26, which spent 2534 credits on the same filter with no gate and delivered 27 leads, 91 credits each. This run spent 1631.01 credits and delivered 20, 84 credits each, with 21 of the 48 companies never reaching a paid lookup. Whether those 21 held good leads is the one thing the run cannot say, because nothing was spent on them; at this run's 59.51 credits per passed row they would have added about 1253 credits to the bill.
What jev had to read was thin. Apollo's list page carries the domain, NAICS and SIC codes, printed revenue, founding year and headcount growth, and nothing about what a company sells or to whom. On that, jev's probabilities ran from 15% to 67% and never higher, so the 50% threshold is doing real work and 60% would have kept 14. The drops read sensibly: a consumer social network, an education publisher, a consumer lender, a robotics maker and a mobile-games studio all went. A third question, which title buys lead data at this company, answered Head of Marketing for all 48, which is the prior and not a judgement: a question the state cannot answer returns the prior, and the page no longer asks it. Enriching each company first would sharpen the gate; it would also cost more per row than the verdict it feeds, which is the trade this workflow exists to show.
Where the money went after the gate: 371 credits to Findymail for 27 calls, of which 7 were misses settled at the list rate because Findymail does not report the charge in its response, and 245 credits to LeadMagic's role finder for the 7 it then found. Tomba, the cheapest per-hit email finder, answered all 27 calls with a capacity error on 鹦鹉AI's own key at no charge, so Hunter served at 308 credits for 18 hits and 9 free misses, and Kitt found 3 of those 9 for 10.50 credits. Verification was 91 credits, with the one unknown free. News was the largest single tool bill again at 588 credits, PredictLeads at 28.01 credits a call. jev's own bill for 67 verdicts was 1.61 credits at its list price on the team's own key.
The opener scores are the honest part of the tail. Of the 19 companies with a news event, jev rated 4 as a decent or strong first line, most of them a funding round or an acquisition; the rest were product launches and partnership notices it scored as generic congratulation. The CSV is sorted by that score, so the four worth writing to first are at the top and nobody has to read 19 events to find them. The catch-all field came back empty on every verified row, so the CSV carries no catch-all verdict for this run.
Tested
What lead generation with 鹦鹉AI costs: recorded data costs
Two measured numbers, and what they do not cover. Both are the metered API and data calls; the agent subscription is separate.
| What was measured | Result |
|---|---|
| This workflow, 23 Sep 2026, 50 companies in | 84 credits per verified contact (1631.01 credits metered, 20 contacts) |
| The same filter without the jev gate, 26 Aug 2026 | 91 credits per lead (2534 credits, 27 leads) |
The email step on its own, 292 people on 16 Sep 2026, cost per correct work email:
| Tool | Per correct email | Exact match |
|---|---|---|
| 鹦鹉AI | 3.92 credits | 90.4% |
| Clay | 27.65 credits | 89.7% |
| Freckle | 29.89 credits | 90.1% |
| Deepline | 64.68 credits | 86.6% |
Not compared: the whole flow in another tool. The bench covers the email step only, and each column is an aggregator rather than a single source. The method and every row are in the work email finding bench.
Questions
Before you start
Can 鹦鹉AI do lead generation?
Yes. 鹦鹉AI runs the prompt on this page: it builds the list, has jev judge each company, finds the person and their work email, verifies it and writes the CSV.
How much does lead generation with 鹦鹉AI cost?
The data is the bill, not the model: the agent subscription is separate. The 23 Sep run came to 1631.01 credits metered for 50 companies, 84 credits per verified contact; news cost 588 credits and finding the people 616 credits including the fallback. The receipt on this page prints the real total for a 50-company run, next to the first run of the same filter without the gate. Per-call rates are the tool's own. A miss on a per-success step is free at the tool's rate card; the receipt shows where that held and where it did not.
Do I need a jev key?
No. jev ran on the team's own TypeSafe key in this run and was not metered by 鹦鹉AI. Without one, leave out the two judging steps: the first run of this workflow had no gate and cost 2534 credits for 27 leads, 91 credits each, against 84 credits with it.
What does jev decide, and what does it cost?
Two things. Before the paid lookups it reads each company's list fields and returns the probability that it fits your ICP; the agent keeps the rows at or above the threshold you name. After the news step it picks which event to lead with and scores how usable it is as a first line. jev is priced on input tokens only, and the two steps together cost a fraction of a cent for the whole run. You need a jev key of your own, or the agent can make the same judgements itself behind the same interface, slower and dearer, until you have one.
Does 鹦鹉AI pick the tools?
No. 鹦鹉AI shows the agent every tool available for each step, with its price; the agent picks, or you tell it which one. There is no automatic failover. jev is the same kind of choice: it is the agent's judge, not 鹦鹉AI's router.
What comes back at the end?
A CSV with company, domain, person, title, email, which tool found it, the verifier's verdict, the catch-all flag, the event to lead with, jev's opener score and jev's fit probability, strongest opener first. The one from the run on this page is linked above with the person, title and email columns removed, because these are real people and a title at a named company is enough to identify one; the row-level outcomes are what the numbers on this page come from. Your own run returns every column.