In this article
457 runs, with the failures counted
A software services firm we work with has run 457 company research jobs. 402 finished, 54 failed, and 1 was canceled.
That is a 12% failure rate, published because the runs record it themselves. A failed run says what broke instead of shipping half a brief. It keeps what it found so far, and charges only what it used.
The 173 calls, broken down
A finished run averages 10.1 minutes and 43 cents, metered call by call. Receipts started on 8 October, partway through. 284 finished runs carry them; earlier runs came before. Reads about people are counted in their own post. Across those 284 runs, the calls split like this.
| Call kind | Per run (average) | What it is |
|---|---|---|
| Web and data reads | 68 | pages, news and public records fetched |
| Model judgments | 62 | fixed questions answered against the evidence |
| Tool steps | 43 | parsing, scoring and assembly |
Each call keeps its receipt, and the receipt does two jobs. The 43 cents can be checked, and each finding can name the exact place it was seen.
4.7 findings per run, each with a source
Runs average 4.7 findings, each tied to where it was seen. A finding that cannot name a source does not ship.
The brief opens with a verdict that can say no. The runs do not yet store that verdict as a count, so we publish no yes-to-no split. When the count exists, it will be here.
The quality bar behind it all is a gold brief written by hand, matched claim by claim.
The run that lost $1.09
On 19 October, a run was pointed at a company with no website we could find. It used its full 24 minutes, the time cap per run. When it timed out, it lost $1.09 of finished work.
Two rules came out of that day. The website is found first, before anything else is paid for. And a run keeps its partial results and receipts, whatever fails.
The failure is in the journal with the rest of our build record.
The seat sums a buyer should do
Our price is public: $1,000 a seat a month, with research for 50 companies a seat. Mach isn't open yet. That puts the seat's ceiling at $20 per researched company, next to the 43-cent meter above.
Do the same sums for any tool you test. Put the price per included unit next to what the unit really costs, if the seller will show it.
Three numbers to ask any research tool for
Failure rate, minutes per finished run and cost per run, from the tool's own logs. Ours today: 12%, 10.1 and $0.43, for this one firm.
Then read one brief against reality and count what it missed. Research on people is where the message gets written. Research on the company decides whether anyone should write at all.
Common questions
Why do runs fail at all?
Mostly the outside world: a company with no website to find, a timeout, a source that will not answer. The run records the cause and charges only what it used.
Does the meter include the time my team spends reading briefs?
No. It counts only the machine's calls. Your team's reading time is a separate cost, so add it when you compare tools.
Where the numbers come from
Totals from our research runs' own stored records. One firm we work closely with; exported 20 October 2026.
