AI lead qualification: qualify companies at scale with Jev and Claude Code

What is AI lead qualification, and what do most guides miss?

AI lead qualification is the use of a model to decide whether a company or person fits your ideal customer profile (ICP), so that sales time and sequencer volume go only to leads worth contacting. The usual definition covers frameworks such as BANT and scoring by behaviour. That is lead qualification for inbound leads, where someone has already raised a hand.

This guide covers the other case: outbound, where you start with a list of 5,000 or 500,000 accounts nobody has spoken to, and a model must decide which ones are in your market. The current top Google results for "ai lead qualification" (checked 25 October 2026) are mostly vendor posts. They explain frameworks and benefits and promise accuracy, but they give no way to measure that accuracy on your own list and no cost per lead. That gap is what this guide fills.

The method is mine, from my own test, and it has one rule: nothing runs at scale until you have measured how often the model is right. I am David Walker-Dobson, founder of Tentt, and I build these systems for B2B teams. If you want the wider context first, start with our GTM engineering overview.

What is Jev, and why use it for true/false questions?

Jev is the first model from TypeSafe AI Inc., a San Francisco startup. SiliconAngle's launch report (16 September 2026) says it was founded in 2024, led by CEO Diogo Almeida, who previously worked on RLHF, InstructGPT, ChatGPT and GPT-4 at OpenAI, with co-founders Erik Gafni and Sasha Sheng, and that it raised a $40 million seed round led by DCVC. TypeSafe calls Jev a "System One Model": it does not write text. You send it some information (the "state") and a question, and it returns a typed answer with probabilities.

Two question types matter here. A choice question picks one option from a list and returns the choice, a probability for every option and a confidence value. A Noul question asks whether a statement is true and returns one number from 0 to 1, which TypeSafe's docs describe as "the probability that the answer is yes". Near 1 is a strong yes, near 0 a strong no, and near 0.5 means unsure. Noul answers carry no separate confidence field. Several questions can go in one request and are evaluated in parallel against the same state.

Pricing, as of 25 October 2026: TypeSafe's site lists $42 per billion input tokens, which is $0.042 per million, and LiteLLM's docs show the same input rate with no output charge. The access page says Jev is "in early access", with keys created in TypeSafe's console. TypeSafe's speed and cost multipliers (such as "193.6x faster") are its own marketing claims, and I could not verify them independently. TypeSafe's confidence page also does not claim the numbers are calibrated for your data and tells you to test them yourself. That is exactly what the method below does.

How do you qualify leads at scale with Jev and Claude Code, step by step?

The stack is four parts: a data API for the account list, Jev for the true/false answers, Crawl4AI to scrape each website, and Claude Code to run the checking. In my test I used Blitz API for the list, stored it in Supabase, and ran Claude Code on a server. Here is the process.

1. Pull a wide account list from a data API. Use broad keywords and cast a net on purpose. In my test I searched for manufacturing-related companies and kept everything that came back, then tested on about 1,000 accounts at first. You are not trying to be precise here. The qualification step does that.
2. Turn every qualification into a true/false question. "Is this company in the manufacturing industry?" works. "Tell me about this company" does not. The same shape covers ICP fit ("Does this company sell to financial services firms?") and job titles ("Is this person a head of revenue operations or equivalent?").
3. Write the control question first. This is your first attempt at the definition, written in plain words with a short criteria line saying what counts as yes and no. It is your baseline. Every later variant is measured against it.
4. Run the control through Jev, then check the answers. Give Jev each company's website text as the state, collect the yes/no results, and note that roughly half of my 1,000 accounts came back as a fit. Then have Claude Code sub-agents do what a person clicking through each site would do: open the scraped page (Crawl4AI turns it into Markdown), decide independently whether the statement is true, and record agree or disagree.
5. Write variants and repeat. Change one thing at a time, such as a stricter definition, an explicit exclusion list or a different cut-off on the probability. Run each variant over the same accounts and score it against the same sub-agent review, the way you would split-test cold email copy.
6. Read the accuracy report and scale the winner. The report says, for each variant, how many yes answers were truly yes and how often it was right overall. My bar was 95%. My best variant reached 96% on manufacturing, so that is the one I scaled. If nothing reaches the bar, tighten the question and loop again rather than lowering the bar.

A sub-agent in Claude Code is a Markdown file with a name, a description and an optional tool list, and per Anthropic's docs "each subagent runs in its own context window", which suits this job: one agent per batch of accounts, each reviewing in isolation and returning only a summary. Claude can run several at once, and the docs say a session stops spawning more once 20 are running unless you raise the limit.

Crawl4AI is open source under the Apache 2.0 licence, installs with pip and outputs clean Markdown, which is the format you want to feed both Jev and the reviewers. Run it locally or on your own server and the scraping itself costs nothing beyond compute.

What do good control and variant questions look like?

A good question names one thing, defines yes and no in a sentence, and says what to do with edge cases. Here are examples for the three common qualifications. The control is the first one you write. The variants change the definition, not the topic.

Notice what the variants do. They add an exclusion, they replace a topic label with a testable observation on the website, and they say how to treat the confusing neighbours: distributors next to manufacturers, agencies next to B2B companies. If a control scores badly, adding exclusions for those neighbours is the first change to test.

What does an accuracy report look like?

The report is a table with one row per variant. The numbers below are illustrative, to show the shape. They are not results from my test, apart from the fact that my best variant reached 96%.

Read it in two directions. Precision tells you how clean the list is, and the sub-agents only reviewing yes answers is how I measured it. The last column catches the failure precision cannot see. V4 looks best on precision, yet it drops 18 of every 100 true fits, so you would be scaling a list that quietly shrinks your market. Pick the variant that clears the bar and loses the fewest fits.

The sample size sets how far you can trust a headline number. With 100 reviewed accounts, a measured 96% has a 95% margin of about plus or minus 3.8 points. With 200 reviewed it is about 2.7 points, and with 400 about 1.9. If the bar is 95% and you reviewed 100, a 96% result is not yet proof.

How much does it cost compared with Claygent or Clay AI columns?

Clay's pricing page (checked 25 October 2026) says Claygent is on every plan, Data Credits start at $0.05 each and Actions cost less than $0.01 each. Per Clay's AI pricing, a web research run on a fixed-price model costs from 0.5 credits (GPT-5 Nano) to 1 credit (Helium) or 3 (Argon), plus one Action. Our Claygent guide has the full table. At the starting credit price, that is about $25 to $150 per 1,000 rows before Actions, so 100,000 rows is $2,500 to $15,000 and a million rows is $25,000 to $150,000.

Jev's cost is per input token. Suppose each company's scraped Markdown is trimmed to 2,000 tokens. At $0.042 per million, 1,000 companies is 2 million tokens, or about $0.08, and a million companies is 2 billion tokens, about $84. The assumption that matters is your page length. Even at 10,000 tokens a page, a million companies is about $420 for the Jev calls.

My own estimate for processing 1 million companies end to end is about $720, based on the stack I used: Blitz API at roughly $500 a month, Claude Code on a plan of about $90 to $120 a month, Crawl4AI at $0, and Jev's per-call cost. I also said a lower figure in my video, so treat $720 as my estimate, not a quote. Blitz's pricing page lists self-serve plans at $99 a month (200,000 records) and $499 a month (5,000,000 records), with custom plans from $999, so check the record cap against your list size, and Jev's price is early-access pricing that may change.

The comparison is not like for like. Claygent searches the web and navigates, while Jev answers a question about text you give it. The measurement step also costs sub-agent time, which comes out of your Claude Code plan. But for a yes/no decision over a large list, the gap is two to three orders of magnitude.

What are the pitfalls when you test AI qualification?

• Too small a sample. Fifty reviewed accounts cannot distinguish 90% from 98%. Review at least 200 before you trust a 95% bar.
• Ambiguous definitions. If two people on your team would disagree on whether a distributor is a manufacturer, no model can score well. Settle the definition in writing before you test it.
• Only measuring false positives. Reviewing the yes answers shows precision. Also sample 100 of the no answers and count true fits among them, or you will not see what you are throwing away.
• The reviewer is a model too. Sub-agents can be wrong, so spot-check 20 of their verdicts yourself, and make them quote the line on the page that decided it.
• Thin websites. A one-page site gives Jev and the reviewers little to read. Add an "unsure" lane for probabilities near 0.5 and send those rows to a second source instead of forcing a yes or no.
• Drift. A variant that hits 96% today can slip when the list changes or when Jev ships a new version (TypeSafe's quickstart calls the jev-latest alias, and each response reports the version that served it, such as jev-1.13.0, so log that version with every run). Re-run a 100-row check each month and before every large batch.

When should you stay on Clay instead?

Stay on Clay when the answer is not on the company's website. Hiring signals, recent funding, a technology in the stack or a news event need web search, and that is what Claygent is for. Stay on Clay too when your team lives in tables, you are qualifying a few thousand rows a month, or nobody on the team is comfortable running scripts. At that volume the Clay bill is small and the setup time is the bigger cost.

The two also combine well. Use the Jev method to cut a wide list down to the accounts that truly fit, then spend Claygent credits only on those, which is the cheap-filter-first approach from our waterfall enrichment guide. For the stack around it see Clay and Claude Code. If you want this built into your own stack, our Clay agency service does it.

See how we build qualification systems

Talk to us about your list