Lesson objective
Triage a contracting task on consequence if wrong, data class, verifiability against a document and who signs, then route it to standard review, enhanced review, redesign or no AI before any tool is opened.
Plain-language explanation
Why this matters now
Level 1 explained how AI tools produce answers, where they fail on drawings, specifications and numbers, and which duties stay with named people on a job. Level 2 starts on an ordinary Tuesday in a project trailer: a submittal log that is behind, three subcontractor emails asking about scope, a superintendent's voice note from the slab pour, and an approved AI assistant already open on the laptop.
This programme will not motivate that work with a sector statistic, because there is no reliable one to give. No verified source records how many contractors use AI, and none shows a proven productivity gain from AI in construction.
Research (preprint, not peer reviewed): Vrcelj and Sandanayake, posted 31 July 2026, assembled 34 studies on AI-enabled construction project controls from one seed database with cross-checks. They found that "only one provides clear evidence of implementation in a live project," and concluded that the evidence "supports theory building rather than causal claims of performance improvement." The corpus is limited and is not an exhaustive review, but it is the clearest statement available that published work mostly measures models, not results on real jobs.
The reason for an operating model is therefore not a benchmark. It is consequence. A wrong quantity becomes a wrong bid. A wrong answer about scope becomes a back-charge dispute. A confident code answer drawn from the wrong edition becomes a failed inspection. Triage is how one person decides, task by task, whether AI belongs in the work at all.
A question to sit with
At 7:40 a.m. a project engineer has two items. One is a request to reformat last week's meeting minutes into the company template. The other is a subcontractor's email asking whether the ceiling grid bracing in the corridor is theirs or the framer's. Both could be pasted into the same assistant within a minute. What makes them different tasks?
What goes wrong
Choosing the tool before the task. The assistant is open, so the question goes in. Nobody asks whether the answer should come from AI, a document, or a named person.
Triaging on time pressure. "It saves twenty minutes" is a benefit, not a triage answer. Twenty minutes saved on a wrong scope answer can cost a week of change-order argument.
Deciding the data question after the paste. Once an owner's drawings, a subcontractor's pricing or an employee's details are inside a tool, the decision about whether they belonged there has already been made by default.
Assuming everything can be checked. A reformatted set of minutes can be checked line by line against the original. An answer about who owns the bracing can be checked only against the executed subcontract, the specification section and the drawings, and if nobody has them to hand, it cannot be checked at all.
Treating a code or design question as office work. A question about what the building code requires, or whether a detail will work, does not become administrative because it arrived in the project inbox.
Losing track of who signs. The draft moves from project engineer to project manager to superintendent, and each assumes one of the others owns it.
The practical answer
Ask four questions, in order, and write the answers down before any tool opens.
| # | Question | What you are looking for | Typical contractor answers |
|---|---|---|---|
| 1 | Consequence: if this is wrong and nobody notices, where does the error land? | The furthest place the output travels | An internal note; a subcontractor; the owner or homeowner; a bid, a pay application or a change order; the field, where work is built from it |
| 2 | Data class: what would the tool see? | The most sensitive item in the material | Public product data; company records; owner or design documents under contract terms; subcontractor pricing; personal information about customers or employees |
| 3 | Verifiability: can every part be checked against a document you hold? | How, and at what cost | Fully, against the supplied document; partly; only by redoing the work |
| 4 | Who signs: which qualified person owns the result? | A named person, not a job title | The project manager; the estimator; the qualifying individual; the design professional; the building official; the person who signs a payment document |
Question 1: follow the output to the end. A scope summary "for the file" is often read later by an estimator pricing a change, or by a foreman laying out work. Rate the task by where an uncaught error finishes, not where it starts. Anything that could be built from is a field consequence.
Question 2: settle the data class first. Module 13 works classification through in detail. At triage you need only the headline: what is the most sensitive thing this task would expose, and has the company approved a system for it?
Question 3: prove it with a document. Ask which document you would check the output against: the drawing at its current revision, the specification section with its addenda, the executed contract or subcontract, the code as your jurisdiction adopted it. If the honest answer is "none that I have," the output cannot be verified and the task needs redesign.
A code question is therefore verifiable only against the adopted text, and the decision on compliance belongs to the building official, not to the contractor or the tool.
Question 4: name the owner. Construction has no single licensed professional who owns everything. The owner changes with the output.
| Output | Owner |
|---|---|
| A bid, estimate or quote levelling | The estimator, with the person who approves the bid |
| A letter, RFI, change-order narrative or owner update | The project manager |
| Anything about workmanship or construction operations | The licensed contractor and its qualifying individual |
| Anything that changes design, drawings or specifications | The architect or engineer of record; the contractor drafts a question, never a design |
| Whether work complies with the adopted code | The building official |
| A document a named person signs, such as a payment application | That person, personally |
Now turn the four answers into a route.
| Route | When it applies | What happens next |
|---|---|---|
| A: proceed, standard review | Internal consequence; data cleared; fully verifiable; owner named | Run the Lesson 2 workflow; the preparer checks every item |
| B: proceed, enhanced review | Output reaches a subcontractor, owner, bid, payment document or the field; data cleared; fully verifiable; owner named | Run the workflow; the named owner checks every quantity, date, dimension, section number and quoted term |
| C: redesign, then triage again | Data not cleared, or part of the output checkable only by redoing it | Reduce the data, split the checkable part from the judgement, or get the source document first |
| D: no AI | No owner can be named, the output would be a design decision, a code determination, a safety sign-off or a certification | Do the task without AI and route the judgement to its owner |
Writing the answers down has a purpose beyond tidiness. Best practice: the NIST AI Risk Management Framework (AI RMF 1.0, January 2023; voluntary, and being revised) says in Appendix C that human roles in decision making and oversight "need to be clearly defined and differentiated." A one-line triage entry is where a small contractor makes that true for one task. The ABC 2025 AI Resource Guide, a voluntary trade association resource, makes the same point in plainer words: "human interaction is still an important factor."
Remember this
- Triage before any tool opens: consequence, data class, verifiability against a document, who signs.
- Rate consequence by where an uncaught error could land, and anything that could be built from is a field consequence.
- A code answer is verifiable only against the code as your jurisdiction adopted and amended it.
- Owners change by output: estimator, project manager, qualifying individual, design professional, building official, the person who signs.
- Design decisions, code determinations, safety sign-offs and certifications are Route D.
Construction example
Harrow Creek Construction LLC is a fictional commercial general contractor with an AI assistant the company has approved for project records and contract documents. On a Monday the project engineer, Priya Nandakumar, triages four tasks before opening it. Task one: convert last week's owner-architect-contractor meeting notes into the company minutes template. Consequence: an internal record the team will rely on. Data: project records, cleared. Verifiable: fully, against the original notes. Owner: the project manager, Dale Okafor. Route A. Task two: a fictional drywall subcontractor asks whether fire-caulking at head-of-wall joints is in its scope. Consequence: a scope position the subcontractor will price and build from. Verifiable only against the executed subcontract, Section 07 84 00 as issued with Addendum 3, and the partition schedule at Revision D. Priya has all three, so she routes it B: the assistant extracts what each document says, with quoted sentences, and Dale decides and answers. Task three: the superintendent asks whether the fictional City of Brennock requires a particular smoke-detector location in the tenant suites. That is a code question. Priya routes it D for the answer: the superintendent is sent to the adopted code with the city's amendments and to the plan reviewer, and the assistant is not asked. Task four: an estimator wants a unit price for rooftop unit curbs "from the AI." No document supports the number and the estimator owns bid pricing. Route D. The four triages take seven minutes and each one-line entry is written before work starts.
Responsible-use warning
Do not let an approved tool make triage feel unnecessary. Approval means the company cleared a system for certain data and uses. It does not mean every task suits it, that every output can be checked against a document, or that someone owns the result. The two costliest mistakes are asking a question whose answer could only be checked by redoing the work, and letting a code, design or safety question pass as office work because it arrived in the project inbox. Triage the task, not the tool, and write the route down first.
Human-review requirement
The person who will own the output confirms the triage before work starts. The estimator confirms it for bids and quote levelling, the project manager for letters, RFIs, change-order narratives and owner updates, and the qualifying individual for anything touching workmanship or construction operations. Design questions go to the architect or engineer of record and code compliance to the building official, never to the contractor or the tool. The owner checks that consequence was followed to the field, the data class is cleared, a real document exists to check against, and their own name is on the task. Anyone may move a task to a stricter route; AI never assigns the route.
Practical exercise
Use a fictional company, fictional projects and fictional documents only.
- List eight tasks from a typical week in your role, such as minutes, an RFI draft, a scope question, a quote comparison, a daily report, an owner update, a code question and a homeowner email.
- Replace every real name, address, price and project detail with a fictional one.
- For each task, answer the consequence question by following the output to its final destination, and mark any task whose output could be built from.
- Record the most sensitive item each task would expose and whether an approved system covers it.
- Name the document you would check each output against, including its revision, issue date or adopted edition.
- Mark each task fully, partly or only-by-redoing verifiable.
- Name the owner of each output from the owner table.
- Assign a route, A to D.
- For each Route C task, write what source document or data reduction would move it to Route A or B.
- For each Route D task, write which answer ruled AI out, and name the person the judgement goes to instead.
- Pick one Route B task and, for its owner, list every quantity, date, dimension and section number the owner would check.
- Ask a colleague to triage the same eight tasks blind, compare, and note which question caused each disagreement. Record the results in your workbook.
Knowledge check
-
What are the four triage questions, and when are they answered?
Answer guidance: Consequence if wrong, data class, verifiability against a document, and who signs. They are answered and written down before any tool is opened.
-
Why is a question about what the building code requires never fully verifiable against "the code" in general?
Answer guidance: A model code binds only through the jurisdiction's adopting instrument, in the edition and with the amendments it names. The check is against that adopted text, and compliance is the building official's decision.
-
What does California B&P section 7068.1 make the qualifying individual responsible for, and what does it not say?
Answer guidance: Exercising supervision and control of the licensee's construction operations, including technical and administrative decisions and checking workmanship. It says nothing about software, and no source says AI or remote monitoring satisfies the duty.
-
Why does this programme not open with a construction AI adoption rate or productivity figure?
Answer guidance: No verified source provides one. A 34-study preprint review found only one study with clear evidence of live-project implementation and said the evidence supports theory building, not causal performance claims.
Sources and further reading
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.
- 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.
- International Code Council (ICC) — Code Adoption. Accessed 2026-09-15.
- California Legislature — Business and Professions Code section 7068.1 (Responsibility of qualifying individual). 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.
- Associated Builders and Contractors (ABC), guide produced by GCM — 2025 AI Resource Guide for Construction Professionals. Accessed 2026-09-15.