How we measured scope decisions and operating margin at an engineering firm

We traced changes from first appearance to decision and billing.

We studied one live engineering program and twelve closed fixed-fee jobs. Eight team members recorded their work and waiting time, and we checked the entries against timesheets. With finance, we traced 214 owner-directed changes through correspondence, drawing comments, and project accounting.

The main operating measure was the share of changes reaching a documented price decision within ten business days. The baseline was 43%. Across three live programs six weeks after handover, it was 86%, using the same decision definition and record-tracing method. The baseline and follow-up covered different project populations. This compares operating performance over time but does not isolate the effect of the system from every other change.

A decision could be a priced notice or a documented waiver. The measure was not the share of changes billed.

The AI prepared evidence and drafts for a human decision.

We connected project accounting, email, meeting records, shared documents, and drawing comments for the pilot program. Project identifiers linked the records. Integration access was scoped to the selected program before expansion.

The change detector produced a candidate with source passages. When the record didn't support a flag, it could abstain. For a supported candidate, the notice drafter used the firm's template and routing rules. A project manager decided whether to price the change, absorb it with a reason, or escalate to a principal.

The decision and supporting passages returned to the project accounting record, which the firm required before an amendment could be billed. Finance could therefore trace a billed change back to the request and notice that produced it. No notice went to an owner without a person sending it. Processing ran in the firm's cloud environment, with its own model used to check citations.

A separate pre-seal check compared drawing comments with the firm's checklist and prior review findings. It supported the discipline lead. The licensed engineer retained responsibility for the review and seal.

The detector had to find real changes without flooding managers with false flags.

Two project managers and a principal built a reference set of 180 agreed examples under a written standard. They turned labeling disagreements into rules, then reached 91% agreement on a fresh sample, up from the initial 84%. The standard separated owner changes, clarifications, and internal coordination errors.

The test used held-out historical examples, earlier information only, and job separation so the example being evaluated did not supply its own reference answer. Recall measured the share of real changes found. The false-flag rate measured incorrect flags as a share of all flags.

The agreed thresholds were at least 85% recall and no more than 20% false flags. The observed results were 88% and 17%. These are separate measures, not an overall accuracy score. Human labeling agreement is a separate consistency measure and is not a mathematical ceiling on recall.

During expansion to requests for information, false flags rose to 31% for three days. We added a rule requiring owner confirmation before treating a contractor's question as an owner change. The rate fell to 17%. Because we'd also widened a mailbox scope that week, we changed our build process to add one channel at a time with a regression check between changes.

Firm margin and additional billing answer different questions.

The financial figures follow our founder's account and the results reported by the client. Firm operating margin on net revenue rose from 15.9% to 23.8%, against a goal of 20.5% for 2026. Work began in early 2025, and the later CFO report was relayed in October 2026. The exact financial reporting period wasn't supplied.

The CFO attributed a large part of the improvement to our work. The project records provide a separate view of additional billing: priced changes carried their source threads into the accounting system, so finance could trace the fees produced by the notices.

Additional services billed across three programs rose from ~$225K per quarter to $410K in the first quarter after rollout. The CFO attributed the $185K increase to notices from the new process.

Earlier budgets reduced the work needed to see a problem.

Managers received budget updates within three days, down from a delay of two to four weeks. Weekly budget review began using existing accounting data before the connected workflow was complete.

Budget re-keying fell from 36 to four staff-hours per week across twelve project managers. Those hours were redeployed, so the result is released capacity. It does not establish a cash saving.

On the pilot, redesign hours after late changes fell 22% against its own prior phase. The projects were still open, so this was a leading indicator rather than a completed-job rework rate. Median waiting time for a seal fell from four to two business days on pilot sets. Neither measure is separately converted into money.

The client proved it could make and test a change.

We gave the client's team operating instructions, extension points, and a regression gate that checked updates against the reference examples before production. Training moved from running the build to changing it and taking ownership, so the team practiced the work it would need to do after handover.

The innovation team independently shipped a channel for requests for information through that gate. Managers retained commercial decisions and licensed engineers retained professional decisions, while the innovation team could maintain and extend the workflow. Six weeks after handover, the decision rate was 86%.