Lesson objective
Measure a mapped construction workflow step by step, separating touch time from waiting time and recording first-pass acceptance and defect origin, so that a redesign targets the real constraint.
Plain-language explanation
Why this matters now
Module 1 set out why a baseline comes before redesign. This lesson applies it to the map you have just drawn. A single end-to-end number, such as "submittals take three weeks", tells you something is slow. It does not tell you where, and it is the "where" that decides whether an AI step can help at all.
In most construction workflows, the hands-on work is a small fraction of the elapsed time. A submittal may take three weeks from the subcontractor's email to the returned stamp, and fewer than two hours of that is anyone in the contractor's office actually working on it. The rest is waiting: in a queue, for a person, for the architect. An AI tool that halves the hands-on drafting time may change the elapsed time by almost nothing.
There is no industry benchmark to measure against. Research (preprint, not peer reviewed): the review by Vrcelj and Sandanayake (Preprints.org, 31 July 2026) of 34 studies on AI-enabled construction project controls found only one with clear evidence of implementation in a live project and concluded that the evidence "supports theory building rather than causal claims of performance improvement." Your own step-by-step baseline is the only comparison you will have.
Professional guidance (voluntary framework): NIST AI RMF 1.0 (January 2023) places MEASURE after MAP. It is voluntary and binds no contractor. The step-level baseline is that sequence in practice: the map says where to measure.
A question to sit with
An AI tool promises to cut the time your project engineers spend drafting submittal comments by half. Your submittals take twenty-one days from receipt to return. How many of those days would the tool actually touch?
What goes wrong
Only the end-to-end time is measured. The redesign aims at the step that feels slow, not the one that is.
Touch time and wait time are mixed. A project engineer who says a review "takes four days" usually means it waited three and a half days for them and took an hour. Automating the hour leaves the three and a half days.
The constraint is outside the company. If most of the elapsed time sits with the architect or the owner, faster internal drafting does not change the outcome, and pushing more items out faster can lengthen the outside queue.
Rework loops are not counted. A submittal returned for resubmittal goes round the loop again. If the baseline counts only first passes, it misses the most expensive path.
Defects are counted but not traced to their origin. A rejection caused by a wrong product from the subcontractor is not the same as one caused by the contractor missing a requirement. Only the second is something an internal AI step could change.
The practical answer
Measure each step on the map, not just the whole. For a sample of real items, record for each step the time it waited and the time someone actually worked on it.
Record first-pass acceptance. Of the items that went out, how many came back accepted without resubmittal?
In the same fictional month, 18 of the 72 submittals were returned for resubmittal, so 72 − 18 = 54 were accepted first time: 54 ÷ 72 = 75%. Each of the 18 went round the loop again, adding its own wait.
Trace every defect to where it started. Use a short origin list and apply it to every rejection and rework.
| Origin | Example | Could an internal AI step change it? |
|---|---|---|
| Subcontractor submitted wrong or incomplete data | Wrong product, missing test report | Only by flagging it before it goes out |
| Contractor missed a requirement in review | Missed a required sample | Yes, if the step is checked |
| Document changed after review | Late addendum or revised section | Only with revision control |
| Design question needing the design team | Conflict between drawings and specification | No: routed to the architect or engineer |
That is four origins, and only one is squarely inside the contractor's review. If most of your 18 rejections come from the first row, the redesign belongs at intake, not at review.
Record the measures on the map itself. Each step box carries its wait, touch time and defect contribution. The operations lead can then see at a glance where the time and the defects actually are.
Measure the exceptions. Record separately how many items took an exception path, such as a substitution or a deferred submittal, and how long they took. They are often few and slow.
Keep the same method afterwards. Whatever sample, clock and definitions you use now, use again after the change, and record who measured and when. AI Coalition Network recommendation: write the definitions of "received", "returned" and "accepted first time" on the map, so that nobody can change them between baseline and result.
Remember this
- Measure each step on the map, separating waiting time from hands-on time.
- If most of the elapsed time sits with the architect or owner, faster internal drafting changes little.
- First-pass acceptance and the rework loop are part of the baseline, not an afterthought.
- Trace every defect to its origin; only some origins are within an internal AI step's reach.
- Fix the definitions and the method on the map so the before and after compare like with like.
Construction example
Bellhaven Mechanical, a fictional mechanical subcontractor that also runs design-build jobs as a prime, plans an AI step to draft its internal review of vendor equipment submittals before they go to the general contractor. Operations lead Tobias Renn follows 40 submittals across four jobs for a month. Average elapsed time from vendor data to return is 18 days: 2 days waiting for intake, 3 days in the engineer's queue, 11 days with the general contractor and design team, and 2 days for return and ordering. Hands-on time in Bellhaven's office averages 80 minutes, of which 55 minutes is the engineer's review. Bellhaven's own steps account for 2 + 3 + 2 = 7 of the 18 days. Of the 40 submittals, 10 came back for resubmittal, so 30 of 40 were accepted first time, 75%. Tobias traces the 10 rejections. Six started with the vendor sending the wrong model or missing performance data. Two were requirements Bellhaven's engineer missed. One followed a revised specification section. One was a coordination conflict that belonged to the engineer of record. Tobias writes the sentence the redesign needs: the biggest internal lever is intake, where six of ten rejections began, not the 55-minute review. He changes the plan: the AI step will first compare vendor data against the specification requirements at intake, flagging gaps for a person to send back before review, and only later be considered for drafting review comments. He records the definitions of received, returned and accepted first time on the map, and the same 40-item sampling method for the result.
Responsible-use warning
Do not redesign a workflow from a single end-to-end time or from the step that feels slowest. Separate waiting from hands-on work, and find where the elapsed time and the defects actually are. Do not count only first passes: rework loops are often the most expensive path. Do not blame an internal step for rejections that began with a supplier or a design question, and do not let a faster internal step lower first-pass acceptance unnoticed.
Human-review requirement
The operation owner designs the sample, defines the measures on the map and keeps the baseline. Project engineers and project managers confirm the wait and touch times and the origin of each rejection from their own job records. Rejections that trace to a design question are recorded as belonging to the architect or engineer of record and are not treated as a target for an internal AI step. The accountable owner approves the redesign only once it names the measured step it targets and how a worse result would show.
Practical exercise
Use a fictional company and invented figures.
- Take the workflow map you drew earlier in this module, with its handoffs, review points and hard edges already marked, and copy it so the baseline figures can be written directly onto each step box.
- Choose a sample of at least 30 real or fictional items and a period that includes a busy week.
- For each step on the map, record the average waiting time and the average hands-on time.
- Total the elapsed days and the hands-on minutes. Show what share of elapsed time sits in your own steps and what share sits outside the company.
- Calculate first-pass acceptance. Show the arithmetic.
- Trace every rejection to one of the four origins in this lesson and count each.
- Write the step on the map where most of the time sits and the step where most of the defects begin.
- Write the definitions of received, returned and accepted first time on the map.
- Draft the one-sentence statement naming the step your AI proposal targets, the change you expect, and what would show it made first-pass acceptance worse.
- Record who measured, over which dates, and what you excluded, such as a job shut down for weather or a week when the architect's office was closed.
- Count how many items took an exception path, such as a substitution or a deferred submittal, and record their average elapsed time separately from the main path.
Knowledge check
-
Why must a baseline separate waiting time from hands-on time?
Answer guidance: Most elapsed time in a construction workflow is waiting in a queue or with an outside party. An AI step that shortens hands-on work may barely change the elapsed time unless it also changes the wait.
-
In the fictional example, 18 of 72 submittals were returned for resubmittal. What is the first-pass acceptance rate?
Answer guidance: 72 − 18 = 54 accepted first time, and 54 ÷ 72 = 75%. Each of the 18 rejected items went round the loop again and added its own waiting time.
-
Why trace each defect to its origin?
Answer guidance: Only some origins, such as a requirement the contractor missed in review, are within an internal AI step's reach. Supplier errors, late document changes and design questions need different fixes or belong to someone else.
-
In the fictional submittal baseline, how much of the 21 days sat in the contractor's own steps?
Answer guidance: Intake 1 day, review 4 days and return 2 days, so 7 of the 21 days, one third. The architect's 14 days were two thirds of the elapsed time.
Sources and further reading
- Preprints.org (Vrcelj and Sandanayake) — From Prediction to Governed Intervention: AI-Enabled Construction Project Controls for Productivity, Resilience and Net-Zero-Oriented Delivery (2026-07-31). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023-01-26). Accessed 2026-09-15.
- Construction Specifications Institute (CSI) — CSI Standards: MasterFormat, UniFormat, OmniClass, SectionFormat/PageFormat. Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.