Case study · Submittal review
How one engineer got submittal checks down to 20 minutes
A mechanical engineer was checking vendor submittals against his own equipment schedules by hand, a column at a time. 20 days after the first commit he had a system that does the comparison and hands back only what does not match, running on his own machine.
- 15–20
- minutes a check takes now, from 30–60 by hand, on his estimate
- $700–$3,300
- a year, on the same estimate
- 20
- days, first commit to last
The challenge
An equipment schedule is the document he hands a contractor. One row per unit, one sheet per equipment type, and a column for every number that matters. A vendor bids against it and sends back a submittal describing what they actually intend to supply. Somebody then has to hold one document against the other.
He put that check at 30 to 60 minutes a submittal, 5 submittals a month, across 6 to 20 columns each. Roughly 10% of it was wasted effort. The expensive failures are a number that came back outside what he specified and nobody caught, and a unit the vendor never quoted at all, which is invisible if you only read their document.
The solution
The vendor PDF goes in directly. The system reads the values off it and compares them against the schedule, so no number is retyped on its way into the comparison. Tolerances carry a direction as well as a size, because a vendor may exceed a specified airflow freely while the same overage on pressure drop is a problem.
It reports what is missing, not just what is wrong. Units he specified that the vendor never mentioned come back as their own finding. That is the discrepancy no amount of careful reading catches, because an absence does not appear on a page.
It refuses 3 comparisons rather than guess at them. It will not match a cfm reading to a gpm column, report on a partly readable table, or let an empty result look like a clean one. Being confidently wrong is worse than stopping, because a check that quietly compares the wrong things still produces a clean report, and a clean report is the thing somebody acts on.
The results
On his own figures the check now takes 15 to 20 minutes instead of 30 to 60, which is 10 to 45 minutes back on every submittal and up to 3.7 hours a month. At around $75 an hour for his own time that is worth $700 to $3,300 a year.
What is left is mostly him ruling on parameters the system could not match on its own, and that should fall as it goes: a mapping he confirms once is remembered for later submittals against that schedule. The export is a real workbook with the title row, the merged header and the footnotes underneath, so it is the document his firm already hands over rather than an approximation of it.
If this sounds familiar
How I keep AI-assisted work from breaking things, worked out on my own product first, is its own case study.
A half-hour check that runs a few times a month quietly costs a few hours every month, and occasionally costs a wrong number reaching a contractor. Here those hours came back as $700 to $3,300 a year of one engineer’s own time, for 20 days of work.
Build figures are a frozen snapshot as of 2026-08-20: 256 commits and 3,331 automated tests, across 2026-07-29 to 2026-08-18. The timings, the 10% and the hourly rate all come from the engineer using it, in conversation, and the rate is a band rather than his exact figure.