Most claims about how fast AI values property arrive without a number behind them. This is a teardown of one run on our own system, with the log counts attached, so that the cost of a source-backed answer can be read rather than asserted.
The run is run 25007, executed on 14 September 2026 on our own internal account. Nothing in it belongs to a client, so no third-party address, asset or name appears here, and nobody's permission was needed to publish it.
What was asked for
The room carried the title its user gave it: Fast BOV at Scale, with the subtitle Price Every Address with Source-Backed Values. In other words: put a number on every address, and show where each number came from.
What the system predicted before it started
Before the run began, our own preflight estimate told the user to expect 45 minutes.
What actually happened
- Time from launch to finished result: 69 minutes (18:41:58 to 19:51:21).
- Rows written to the work log: 513.
- Of those, calls to models: 276.
- Final state: done.
Two things there are worth saying out loud rather than smoothing over.
The time estimate was low by half again: 45 minutes predicted against 69 minutes spent. The preflight estimate of what the run would consume was wrong in the other direction and by more: it overstated the resources the run actually used by roughly four times. Both are misses, both are ours, and naming them is more useful than averaging them away.
How typical this run is
This was not a record. Across completed runs on the product with ten or more logged actions:
- 4,620 such runs have completed.
- Median duration: 26.0 minutes.
- 90th percentile: 55.1 minutes.
- Runs that finished in under two minutes: 14, or 0.3%.
There is a second number that explains a good deal of loose talk in this market, including our own. Of 20,982 runs with any logged action at all, 14,560 consist of exactly one action. Those are short questions, not reports, and they are what drags a median taken across every run down to under a minute. So when anyone says this kind of work "takes about two minutes" (and we have written that sentence ourselves), that figure is the median of short questions presented as the time to produce a report. It is not the same thing.
One note on method, because it decides what a speed number means: whenever a duration is quoted, the population it was taken over is part of the claim.
What this means if you are not buying anything from us
When a vendor offers an instant portfolio valuation "with a source under every number," there is one question that separates the claim from the work: how many steps does the system take, and how many of them are trips to a source?
Five hundred and thirteen logged steps, of which two hundred and seventy-six reach a model, over sixty-nine minutes. That is what source-backed cost on this run. A number that arrives in two minutes either did not go to its sources, or went to exactly one.
That test applies to any tool of this kind. It applies to ours.
What this teardown does not claim
It does not claim 69 minutes is good. Our own preflight estimate said 45, and being off by half again is a defect in the estimate, in the run, or in both. We have not established which.
It does not claim the run's output was correct. This is a measurement of cost in steps and minutes, not an audit of the opinion of value the run produced.
It does not claim that one run settles anything about a second run. It is one measured case, published with its counts, offered as a floor for what this kind of work costs rather than as a promise about yours.
How to check these numbers
Every figure above comes from two tables in our production database. The run-level numbers (timings, estimate, state) come from conversations_room_runs, row id 25007. The step counts come from conversations_room_run_action_costs, which holds 513 rows for this run. The distribution figures cover all runs in state done with ten or more cost rows; the single-action figure covers every run with at least one cost row.
The run was our own, on our own account. That is why it could be published in this detail at all.