Lesson objective

Define the baseline measures a contractor must record before changing an AI-assisted workflow, and build the business case from the consequence of unverified output rather than a claimed industry gain.

Plain-language explanation

Why this matters now

Contracting businesses often adopt a tool without writing down how the work performed before. Six months later someone says the tool saves an hour a day per project engineer, and nobody can say whether that is true, or whether resubmittals went up while the drafting time went down.

The pressure to skip the baseline is strong, because the tool arrives with a story attached. A vendor presentation, a conference talk or a trade article suggests the gain is already known. It is not. Research (preprint, not peer reviewed): in their review posted on Preprints.org on 31 July 2026 (DOI 10.20944/preprints202607.2397.v1), Vrcelj and Sandanayake assembled a 34-study corpus on AI-enabled construction project controls. "Only one provides clear evidence of implementation in a live project, while two others approach an identifiable project-control intervention pathway." They conclude that "[t]he evidence therefore supports theory building rather than causal claims of performance improvement." The corpus came from one seed database with cross-checks, so it is not exhaustive, and the paper has not been peer reviewed. As of mid-2026, though, the published work is overwhelmingly about how models score on tasks, not about what happens on real projects.

That changes how a contractor should build its case. There is no industry gain to borrow. There is your own baseline, measured honestly, and there is the consequence of an unverified output, which you can describe precisely.

A question to sit with

Your owner wants to know whether the AI-drafted RFI process "works". What would you need to have written down before it started to give an answer that a sceptical controller would accept, and what would you need to show besides the time saved?

What goes wrong

There is no baseline. Without a record of time, volume, rework and errors before the change, no later claim of improvement can be tested and no deterioration can be seen.

The baseline measures only what flatters. A baseline that records drafting time but not resubmittals, architect returns or field questions is set up to show success. Time saved is easy to produce by checking less.

A borrowed number becomes the target. A figure from a vendor, a single case study or an article is written into the plan as the expected result. The pilot then measures whether that number was hit instead of what happened, and every judgement leans toward the number.

Before and after are measured differently. A stopwatch baseline compared with a figure from a software dashboard measures two different things.

The case rests on the gain, so the controls get cut. If the business case depends on hours saved, every review step looks like a cost to be removed. If it depends on consequence, the review step is the point.

The practical answer

Build the case from consequence. For each AI-assisted operation, write what a wrong output costs when it escapes, who bears it, and how reversible it is. That is the business case for the controls, and it does not depend on any claimed gain.

Output If wrong and unchecked Who bears it Reversible?
Internal meeting summary Confusion, a repeated discussion The team Easily
Submittal requirements checklist Resubmittal, a subcontractor priced against the wrong list Company and subcontractor With cost and delay
RFI draft to the architect Wrong question, wrong answer built into the work Company, owner Sometimes, with rework
Change narrative to the owner A concession or a claim you did not intend Company Often not
Anything in a signed certification A false statement by the signer The named signer No

Module 3 turns this into a tier grid. Here it is enough to see that the argument for discipline gets stronger as reversibility falls, and that none of it requires a productivity claim.

Use the preprint's frame as a control design. Vrcelj and Sandanayake describe the missing piece as "governed decision translation": how an AI output is "interpreted, validated, challenged, authorized, assigned for implementation, documented, and reviewed." That is seven steps, and each can be a line in an operation.

Step Question for your operation
Interpreted Who reads the output and what does it mean for this job?
Validated Against which document and revision is it checked?
Challenged Who may disagree with it, and how?
Authorized Who has authority to act on it?
Assigned for implementation Who carries it out?
Documented Where is the record?
Reviewed When is the result looked at again?

Record the baseline before anything changes. For the workflow you plan to change, measure at least these, over a period that includes a normal busy week.

Measure How to record it Why it matters
Volume Outputs per month, by job Converts rates into counts later
Time Minutes per output, from a timed sample Any claim of saving needs a before
Rework Share of outputs corrected before release The first sign of quality change
Escaped errors Resubmittals, architect returns, field questions traced to the task What reached someone outside the company
Exceptions Items that could not be completed Kept separate from errors
Review coverage Share of outputs actually checked, and by whom Shows whether review is happening

Measure with the same method before and after. If the baseline timed twenty outputs with a stopwatch, the later measure does too. Record who measured, on which dates, and what was excluded, such as a week when two jobs were shut down for weather.

Map, then measure, then manage. Professional guidance (voluntary framework): NIST AI RMF 1.0 says that after instituting GOVERN outcomes "most users of the AI RMF would start with the MAP function" and then continue to MEASURE or MANAGE, iteratively. It is voluntary and binds no contractor. The baseline is the first MEASURE step, taken on the map Module 2 builds.

Remember this

  • No verified source shows a proven AI gain on construction work; the 34-study preprint supports theory building, not causal claims.
  • Build the case for controls from what an unverified output costs and how reversible it is.
  • Record volume, time, rework, escaped errors, exceptions and review coverage before anything changes.
  • Measure the same way before and after, and never set a target from someone else's number.
  • Governed decision translation gives seven steps a construction operation can make concrete.

Construction example

Larchmont Build Group is a fictional commercial contractor planning an AI tool that drafts RFIs from project engineers' notes and marked-up sheets. The vendor's slide promises large time savings, and the owner, Petra Hollis, asks operations lead Desmond Achebe for a business case. Desmond refuses to use the slide. Instead he records a four-week baseline that includes one busy week. The company issued 180 RFIs across its jobs. A timed sample of 20 took 500 minutes, 25 minutes each, so 180 × 25 = 4,500 minutes, 75 hours. Nine were corrected by a project manager before release, 5%. Three came back from architects as unclear. Four could not be drafted because the relevant sheet was missing from the job file. Project managers reviewed all 180. He writes the dates, who timed what, and that one warehouse job was excluded because it was shut down for two weeks. For the business case he writes the consequence table: an unclear RFI costs a week of response time on average on the company's own records; a wrong question can build the wrong answer into the work. He maps the seven governed decision steps onto the RFI process and finds that "challenged" has no owner, so he adds a line: any project manager may reject a draft and log why. Petra approves a pilot on three jobs, with the same stopwatch method, the same quality measures and no savings target. Controller Ivo Marsh checks the arithmetic and adds one line to the file: any result reported to the company's lender will state the sample, the period and what was excluded.

Responsible-use warning

Do not skip the baseline, and do not record only time. Without rework, escaped errors, exceptions and review coverage, a later comparison can only show success. Do not write a vendor figure, a single case study or any article's number into your plan as a target or a projection: no verified source shows a proven AI gain on construction work, and a target bends every later judgement toward itself. Build the case from consequence and your own measurement.

Human-review requirement

The accountable owner approves the baseline method and the decision to run a pilot. The operation owner takes and keeps the baseline, records who measured, when and what was excluded, and uses the same method afterwards. The project managers who own the outputs confirm the rework and escaped-error counts from the job records. The controller checks the arithmetic and that no borrowed figure has entered the plan as a target. Anyone presenting results to an owner or lender reports them with their sample size, period and limits.

Practical exercise

Use a fictional company and invented figures only.

  1. Choose one workflow you plan to change with AI, such as RFI drafting, submittal checklists or daily reports.
  2. Build a consequence table with at least four outputs from that workflow: what a wrong, unchecked output costs, who bears it, and whether it is reversible.
  3. Map the seven governed decision steps onto the workflow and mark any step with no named owner.
  4. Choose a four-week baseline period that includes a busy week. List the six baseline measures and how you would record each.
  5. Using invented figures, calculate the time per output, the monthly total and the rework rate. Show the arithmetic.
  6. List what you excluded from the baseline and why, and who took each measurement on which dates.
  7. Write which method you will use after the change and confirm it matches the baseline method.
  8. Find one claimed AI gain you have seen in a presentation or article. Write why it will not appear in your plan as a target.
  9. Draft the two-sentence business case, built from consequence, that you would give your owner.
  10. Write how you will report the pilot result to your owner: the sample size, the period, what was excluded, and both the time and the quality measures side by side.

Knowledge check

  1. Why should a contractor build the case for AI controls from consequence rather than from a claimed gain?

    Answer guidance: No verified source shows a proven AI productivity gain on construction work; the 34-study preprint supports theory building only. The cost and irreversibility of an unverified output is something the company can describe precisely, and it justifies the review step.

  2. What are the seven steps of governed decision translation named by Vrcelj and Sandanayake?

    Answer guidance: An AI output is interpreted, validated, challenged, authorized, assigned for implementation, documented and reviewed. Each step can become a named line in a construction operation.

  3. A timed sample of 20 RFIs took 500 minutes and the company issues 180 a month. What is the monthly baseline time?

    Answer guidance: 500 ÷ 20 = 25 minutes each, and 180 × 25 = 4,500 minutes, or 75 hours a month. It must be recorded with rework, escaped errors, exceptions and review coverage.

  4. Why must the before and after measurements use the same method?

    Answer guidance: A stopwatch baseline compared with a software dashboard figure measures different things, so any difference could come from the method rather than the change, and the comparison proves nothing.

Sources and further reading