The run stops, nothing is written into the room, and no message tells you it stopped. The record is kept on our side; a reason for the stop usually is not. Nothing retries automatically and no alert reaches you. Say so in the conversation where it happened, and a person picks it up.
What actually happens, in order
- The run stops part-way through its chain of steps. Nothing is written into the room.
- No message is produced to tell you it stopped. The conversation goes quiet, and quiet looks identical to a job still working.
- The run itself is kept. That it existed, when it ran, which account it belonged to and what state it ended in are all recorded and can be looked up on our side.
- A reason usually is not kept with it. Our records will say a run ended other than completed; in the ordinary case they do not say why.
- Nothing else happens on its own. No alert goes out, no retry starts, no queue picks the job back up.
Step five is the one most people assume is false. It is not. If a run of yours was cut off and nobody wrote to you about it, that is the normal behaviour of the system today, not an oversight in your particular case.
Slow is not broken
Before treating silence as a failure, check it against how long this work actually takes. Most jobs land within the hour: median run times sit at roughly 11 to 48 minutes depending on which product you used. The slow end is genuinely slow — 90th-percentile times run from about 44 minutes up to about 142 minutes, and the tail is longer than that. One ordinary chat run took 24 hours before it finished.
So an hour of quiet is usually a job still working. A day of quiet is not.
The failures that do not look like failures
Some runs finish and are still wrong for you. These are the three worth knowing about.
A search outside our coverage returns nothing rather than an error. Territory search works over registered US states; asked to work outside them, it does nothing and says nothing. The automatic coverage check that would warn you runs for one product out of fifteen and has fired 25 times across 12,542 runs. Ask whether a place is covered before you run it, not after.
Two properties in one single-subject job collide. Put two addresses into a job built for one, and the facts of the second can be dropped while the output still looks complete. First writer wins. We have seen this on 9 runs. Run them separately.
Thin evidence produces a narrower answer, by design. Below three comparable sales, an opinion of value comes back as a range rather than a point. Where the facts behind a block were not found, the block comes back empty; an empty block means we did not have the fact, not that the fact is zero. In one teardown of our own run, the run scored its own confidence down and named the specific documents that would raise it, rather than presenting a clean number on thin ground. That is the intended behaviour, not a malfunction — a confident answer over two comparables would be the dishonest version.
What we do about it without being asked
Narrowly, and by hand.
In September 2026 we went back through runs that had been cut off on our side, worked out which person each one belonged to, and wrote to each of them naming the cause as ours rather than theirs. One property in that group had been run seven times, cut off every time, and nobody here had said a word about it.
That was a catch-up on our own silence, done one person at a time. It is not a standing guarantee, and it is not automation. There is no monitor today that notices your run stopped and tells you. If something of yours broke after that sweep, the honest position is that we may not know unless you tell us.
What we cannot tell you
We can confirm that a run happened, when it ran, and how it ended. We usually cannot tell you why it ended that way, because the reason is not stored. Anyone who gives you a confident root cause for a cut-off run is reconstructing it, and should say so.
There is nothing to download about a failed run — no export, no file, no link. What exists is the record on our side and a person who can read it.
People have been leaving instead of complaining
This is the part we would rather not print.
There is a complaint button in the product. Across the whole history of the product, exactly two reports have ever come through it, on 14 and 15 September 2026, and both are ours: one was a test that the button worked, the other was our own founder reporting that the first one produced no answer. Not one client has ever used it. That is not a quiet user base; it is a channel nobody reaches.
The complaints that do exist are inside conversations, and the people who were failed worst left none at all. They ran something, got nothing, and did not come back. Until this month, nobody who rated one of our answers at the bottom of the scale had ever been asked why, in either direction. The first replies to those ratings were written by hand in September 2026.
We are stating this because the alternative is to read our low complaint volume as satisfaction, and it is not.
What we are asking you to do
Say it in the conversation where it happened. That is the fastest route to a person. The complaint button exists on one screen and almost nobody finds it; a sentence in the room does more.
Rate the answer, low if it deserves it, and add one line of why. Numbers, comparables, format, or simply that it did not answer the question you asked — one line is enough, and it is the part that tells us what to fix.
Do not run it a third time. If a job was cut off and re-running it broke it again, the repetition is itself the signal. Tell us instead of trying again — a run that fails twice the same way will usually fail a third time.
Tell us if the job is still needed. A cut-off run from last month is worth reopening only if the property is still in front of you. If it is, say so; if it is not, saying that is useful too.