Leads with a reason to write: hedwigai ask, Apollo and treg on one brief

Here is one real lead from the test, exactly as ask built it. Only the person's name and address are withheld.

Look at what that gives the agent drafting the email. The first line writes itself, and it is true. It is going to the person the company itself named. And the agent knows the address is real, not because a vendor said "verified", but because the company's own mail server accepted it while refusing made-up addresses at the same domain.

Compare a lead for the same kind of event from the other two. Apollo returns a database row: a funding stage dated to the first of the month, a person from its directory, and an email labelled "verified". treg returns a news headline and a directory's guess at who runs the company. Neither tells the agent which parts to trust. In this test, that is where they lost their leads: half of Apollo's "raises" were not raises (mostly loans and acquisitions), 5 of the 11 chief executives treg named were the wrong person, and 17 of Apollo's 26 "verified" addresses were at servers that accept anything.

On raises, ask named the right person 19 times in 20. And every one of its 26 email labels, "confirmed" or "cannot be confirmed", matched the independent check.

It is tempting to judge a lead tool by how full its rows are: does every row have a name, a title, an email, a company size. That is the wrong way to look at it. A lead is useful only if three separate things are all true, and they fail independently.

1. The reason is true and fresh. The raise happened, it was equity and not a loan, and it happened inside the window.
2. The person is right. They hold the role that decides, at this company, today.
3. The address reaches them. A mail server accepts it.

Because they fail independently, the chance a lead works is the product, not the average. Suppose a tool gets each link right four times in five. That sounds good. But 0.8 × 0.8 × 0.8 is about 0.5: half its leads are broken somewhere. A tool that is 0.95 on two links and 0.5 on the third does no better than one that is 0.7 on all three. The weakest link sets the result.

So the number that matters is cost per usable lead: what you paid, divided by the leads where all three links held. A cheap tool that breaks a link half the time is not cheap.

Every system got the same brief: 40 leads at US companies, in two halves of 20.

• Just raised. A private company that sold equity in the last 30 days (on or after 12 September 2026), at least $1M, not an investment fund. The lead is the chief executive or a founder.
• Hiring engineers. A company that posted software or engineering roles in the last 30 days. The lead is whoever runs engineering: the CTO, a VP or head of engineering, or a founder at a small company.

Each lead had to come back with a company, a person, their title, a work email, the reason in one sentence with its date, and the page that shows it.

We built each system's leads the way an agent using it would: with its own search, its own people data and its own email step, and nothing borrowed from the others.

All 112 leads were mixed together, stripped of which system made them, and split among three checkers. Each checker was an AI agent allowed only web search and public pages, with written rules for each link:

• Reason. A public page (a filing, a press release, a job post) shows the event inside the window. A loan, an acquisition, a job that belongs to a different company, or a post older than the window fails.
• Person. A public page shows this person holds a deciding role at this company now. A recruiter, a person at a company with a similar name, or a "former" CTO fails.
• Email. Every address, from every system, went through the same mailbox checker (MillionVerifier). Each comes out one of two ways. Confirmed real: the company's mail server accepted this address and refuses made-up ones, so a mailbox is there. It counts as one. Can't be checked: the server accepts any address at all, so nothing outside can tell whether this person is behind it. It counts as half.

One rule we tightened after reading the checkers' notes. A Form D filed in the window can describe money that arrived months before; one filed in October had a first sale in July 2024. We counted a raise as fresh only if its first sale was on or after 28 August, the 15-day filing deadline before the window opened. This cost ask six leads.

Four of treg's leads were set aside, not failed. Their "company" was a fragment of a news headline ("Why we raised", "Sequoia Leads") that our own script cut badly. An agent reading those headlines would not make that mistake, so it would be unfair to count it against treg.

On the funding half, ask picked the right person 19 times in 20. Apollo did it 14 times in 20, treg 6 in 16.

The difference is where the name comes from. A Form D lists the executive officers, and the company signs it under penalty of law. A sales database or a routed people search has to guess who runs the company from a directory, and directories go stale. Several of the "chief executives" the other two returned had left the job, ran a different company with a similar name, or were the founding CEO of a company that has since named someone else. When the record of the event also names the people, the second link of the chain comes for free.

Apollo labelled every one of its 26 addresses "verified". When the same mailbox checker looked, 9 were accepted by the server. The other 17 were at companies whose servers accept any address at all, so the label could not have meant what it says.

ask labelled 11 of its 26 addresses "confirmed", and the checker agreed on all 11. For the other 15 it said, in the answer, that the company's server accepts everything and nothing more can be known. The checker agreed on all 15 of those too.

This matters more than it looks, and it is why the strict count is closer. An agent that knows which addresses are certain can write differently to the uncertain ones, or ask a person before sending. An agent told everything is verified will send to all of them, and some will bounce, and bounces are how a sending domain loses its reputation.

Apollo's funding filter returned well-known companies with real news behind them, and that is its strength. But 10 of its 20 "raises" were not raises. Several were loans or credit lines (a robotaxi company's $5B, a logistics company's $400M), two were companies being bought, and one was a research initiative paid for by a government and partners. They are all "funding events" in a database. None is a reason to congratulate someone on their new round.

ask made the same kind of mistake in its own way: six Form D filings whose money had arrived months earlier, and one offering of debt. The filing says both things plainly, in its first-sale date and its security type. We were simply not reading them. We are now: since this test, funding recent leaves out loans and late filings unless asked for them.

Hiring was the hardest half for everyone. Apollo's job filter mostly returned staffing agencies and very large employers, and the people it found there were recruiters or held the title in another part of the company: 2 right people in 20. treg's job feed spent its budget on 25 postings that came from 12 companies, and several postings belonged to a different company than the board that showed them: 2 right people in 12.

ask read the companies' own job boards, which have no staffing agencies on them, and got 19 true reasons and 13 right people in 20. Its mistakes were name collisions. A job board called "Mercury" or "Omni" says nothing about which Mercury, and ask twice found the right title at the wrong company. Short names are the failure to watch for.

Apollo charged per company looked up and per email revealed, whether or not the lead survived: $1.90 for 40 leads. treg's job feed charged $0.04 a posting, so most of its $1.40 went on finding the hiring companies.

ask's reasons come from public records that cost nothing to read. Its money goes on searches at about a tenth of a cent each, mailbox checks at a fifth of a cent, and a paid finder only when the cheap steps fail. That is how 40 leads cost $0.48, at the price a customer pays. On our separate test of email finding alone, 120 people whose addresses are published, ask got 97 right and 16 wrong at about $0.006 per right address.

ask is a tool on hedwigai's MCP server, the standard way AI assistants plug in outside tools. In Claude Code:

claude mcp add --transport http --scope user hedwigai https://mcp.hedwigai.com

Then ask for leads in plain words, for example:

Find 20 US software companies that raised money in the last 30 days. For each, the CEO, their work email, and the filing that shows the raise. Say which emails are confirmed.

New accounts start with $1 of credit, which at these prices is enough for a first list.