Enforce what can be exact.
Parsing, arithmetic, pricing logic, schemas, validation, and state stay in code.
Work / Applied AI / Commercial estimating
The Review Desk that surfaced
$221,915.70
across $3.44M of bids.
Commercial bids were being assembled from several independently maintained sources with no systematic way to catch disagreements between them before the quote went out. I built a reconciliation system that cross-checks those inputs, attaches evidence to every finding, and keeps every commercial decision with the estimator.
Identifying details anonymized. Results and measurements unchanged.
- 34live projects in the measured set
- 6.4%verified catch rate across the observed set
- Zeroescapes that cost the company money over the observed set
01 / The problem
The constraint was attention, not judgment.
Estimators already knew how to resolve discrepancies. The expensive part was holding several changing sources in mind, under deadline, on every bid.
One recurring risk was duplicate scope across disciplines: independently maintained sources could assign overlapping responsibility for the same work even when each document looked reasonable on its own.
On one rebid, about $213K of obsolete scope came out while about $35K of drawn but unpriced scope went in as an allowance before customer delivery — roughly $248K of scope reconciled in opposite directions. The goal was not to make the quote lower or higher. It was to make the quote match the current basis.
- 01Internal quotes
- 02Manufacturer quotes
- 03Drawing takeoffs
- 04Bid breakouts
- 05Catalog pricing
- 06Service + labor rules
02 / The product decision
I automated the reconciliation, not the decision.
The first question was not how much estimating AI could automate. It was where human judgment was valuable and where human attention was being wasted.
- Spreadsheet-only lost.
The difficult part was not arithmetic. It was maintaining relationships across independently maintained documents that keep changing through the estimating process.
- Full autonomy lost.
A wrong commercial decision could directly affect a customer-facing quote. Removing the approval gate would make the system more autonomous and less useful.
- A bounded system won.
The system received a narrower responsibility: find evidence-backed disagreements before the bid goes out. The estimator decides what those disagreements mean.
The responsibility boundary
The boundary is the product.
Interpret where judgment creates leverage.
Models read drawings and unstructured material, identify possible discrepancies, and synthesize evidence.
Keep authority where consequence lives.
Every decision that can change a commercial quote remains human-owned.
03 / Then reality broke it
The first real packet contained three wrong lines.
One of them exposed a deeper problem than an ordinary model miss.
System finding
“5 WS-tagged wall stations.”
But WS was a scheduled wall-sconce fixture tag, not a wall station.
Reconstructed from the review packet and rerun evidence. Not a screenshot of the protected project.
I had corrected the same underlying fixture-tag trap months earlier. The correction existed in an interactive AI session. It had never become durable system logic. Later runs sometimes caught it and sometimes missed it.
WS was interpreted as wall stations.
A known correction lived in session history, not in the pipeline.
A fixture-tag misread gate was added after synthesis in plain code.
WS was treated as a fixture rather than a station.
A correction that lives only in conversation is not a durable fix.
04 / The ratchet
The past reviews became an evidence base.
That raised a larger question: how much expert correction history was still buried in prior sessions?
62historical job sessions mined for operator corrections
65operator corrections recovered from those sessions
9recurring failure classes across the correction set
Not every historical correction automatically became code. That would be the wrong lesson. The useful shift was treating corrections as an evidence base. A recurring class can be traced to its mechanism and, where appropriate, converted into a durable gate, test, procedure, or replay surface.
The product loop changed from documents → AI → findings to real work → human correction → mechanism → system change → rerun.
05 / What happened
$221,915.70
of human-confirmed missed scope and estimating errors surfaced across $3,442,234.02 of real commercial estimating work.
34 live projectsmeasured set
6.4%verified catch rate
Zeroescapes that cost the company money over the observed set
The system did not autonomously recover revenue. It did not replace the estimator. It did not make customer-facing commercial decisions. Its responsibility was narrower: make consequential discrepancies visible while a qualified human could still act on them.
06 / What carried forward
The hardest question in applied AI is rarely “can the model do this?”
The more useful questions are which parts of the workflow should be deterministic, where model interpretation creates leverage, where authority must remain human, and what happens when an expert says the system is wrong.
- Find the expensive judgment bottleneck.
- Decompose the workflow.
- Automate bounded pieces.
- Preserve human authority where consequences matter.
- Observe failures in real use.
- Turn recurring corrections into machinery.
- Run it again.
More work