Lesson objective

Distinguish rules-based automation, machine learning, computer vision and generative AI in construction work, and match the check you apply to the way each one fails.

Plain-language explanation

Why this matters now

Almost every product a contractor touches now carries an "AI" badge. The scheduling software says it predicts slippage. The camera on the gate says it spots a missing hard hat. The estimating package says it counts fixtures off a PDF. The chat window says it will write your change-order letter. The badge is the same on all four. The systems underneath are not, and they fail in four different ways.

The Associated Builders and Contractors 2025 AI Resource Guide for Construction Professionals is one of the few construction-specific resources that draws the distinctions at all. It separates artificial intelligence, machine learning, deep learning, generative AI and predictive AI, and it organises applications across the project lifecycle: preconstruction (predictive analytics, design optimisation, automated takeoffs, contract review), construction (autonomous equipment, project management, computer vision for safety, labour tracking) and building maintenance (energy management, predictive maintenance). Professional guidance: that guide is a trade association resource with a vendor-supplied component. It is voluntary, it is not a regulator's position, it creates no legal obligation, and it is not an industry standard of care.

Be equally careful about what anyone claims these systems deliver on a real job. Vrcelj and Sandanayake, in a preprint posted on Preprints.org on 31 July 2026 (DOI 10.20944/preprints202607.2397.v1), assembled a 34-study corpus on AI-enabled construction project controls from a ScienceDirect seed stream with Web of Science cross-checks. Of those 34 studies, "[o]nly one provides clear evidence of implementation in a live project, while two others approach an identifiable project-control intervention pathway." The authors conclude that "[t]he evidence therefore supports theory building rather than causal claims of performance improvement." That is a preprint, not peer reviewed, and a 34-study corpus from one seed database is not an exhaustive review. Read within those limits it still says something a contractor needs to hear: as of mid-2026 the published work is mostly about how models score on tasks, not about what happens on a job.

A question to sit with

Your scheduler flags a slip, your gate camera flags a missing hard hat, and an assistant drafts the letter you will send the owner about both. Which of those three could be wrong in a way you would not notice just by reading what it produced?

What goes wrong

Treating every "AI" badge as one thing. A company writes one line of policy — "check AI output" — and applies it identically to a fixed software rule and to a generated letter. The rule gets re-checked every week; the letter goes out unchecked.

Checking generated content the way you check a rule. A rule that was tested once behaves the same way the next time it meets the same facts. A generative system writes new text every run. Having checked last week's letter tells you very little about this week's.

Reading a confidence score as a verdict. Vision and prediction features usually attach a score. The score is the system's own estimate of how strongly the patterns it learned point to that answer. It is not proof that this worker is unharnessed, that this fixture is a fixture, or that this activity will slip, and it is weakest exactly where the situation is unusual.

Mistaking fluent writing for understanding. A well-written paragraph about why a foundation pour ran long can rest on a quantity nobody verified or a reason nobody supplied. Fluency is a property of the text, not of the analysis behind it.

Letting the product category decide who owns the result. No tool moves responsibility. Whoever would have owned the output without the tool still owns it with the tool.

The practical answer

Sort every AI feature you meet into one of four families before you decide how to check it. Most products combine several families, so sort the feature, not the product.

Rules-based automation Machine learning (including predictive) Computer vision Generative AI
How it works A person writes explicit if-then logic The system learns patterns from past examples A model trained on images or point clouds detects and classifies what it sees A model trained on very large amounts of text produces new text, tables or code
Typical construction use A workflow that routes any submittal over a set value to the project manager Predicting which activities are likely to slip; flagging unusual invoice amounts Detecting personal protective equipment, counting deliveries, comparing progress photographs to a model Drafting an RFI, summarising a specification section, explaining a delay
How it fails Consistently: the rule is wrong, or the situation falls outside it Quietly on unusual items, and it drifts when the work changes Misses and false alarms driven by light, angle, occlusion and camera position Plausibly: confident, well-formed content that is false
What a wrong output looks like The same wrong answer every time A reasonable-looking flag, usually with a score A box drawn round the wrong thing, or nothing drawn at all A fluent, complete-looking answer
How you check it Test the rule, review the exceptions, control changes to the rule Sample the flags, look hardest at unusual and low-score items, watch for drift Spot-check against what is actually on site, and record what the camera could not see Verify every quantity, citation and clause against the document you hold

Four families, four rows in each column. Notice that only one of them fails loudly.

Where each family tends to show up. The lifecycle framing in the ABC guide is a useful place to hang them.

Lifecycle stage (ABC guide) Applications the guide names Families usually involved
Preconstruction Predictive analytics, design optimisation, automated takeoffs, contract review Machine learning, generative AI, rules
Construction Autonomous equipment, project management, computer vision for safety, labour tracking Computer vision, machine learning, rules
Building maintenance Energy management, predictive maintenance Machine learning, rules

The middle column is the guide's; the right-hand column is this programme's reading of it, not ABC's text.

Three questions turn the table into a habit.

  1. Which family produced this output? If you cannot tell, ask the vendor or whoever administers the software, and write the answer down.
  2. How does that family fail? Match the check to the failure. A rule needs a test. A prediction needs sampling. A detection needs a look at the real thing. A generated draft needs line-by-line verification against the drawing, specification, adopted code or executed contract.
  3. Who owns the result? That follows the work, not the tool. In construction there is no single licensed professional who owns everything, so name the right one every time:
    • the licensed contractor and its qualifying individual for the contract, the workmanship and anything advertised under the licence;
    • the architect or engineer of record for design, drawings, specifications and anything that must be sealed;
    • the building official or other authority having jurisdiction for whether work complies with the code that jurisdiction has actually adopted;
    • the employer, and the duty-holder named in the applicable safety standard, for safety decisions;
    • the person who signs the payment application, the certified payroll or the lien waiver;
    • the owner or principal for the company's own records, marketing, customer data and money.

Licensing and the design line are state law and sometimes city or county law, and Module 4 works through them. Whatever this course says about "your licence" is a prompt to check your own state and municipality, never a statement of fact about it.

Apply it one feature at a time. For each AI feature in your work, write a single line: what the system does, what a person does, and who decides.

Feature System does Person does Who decides
Submittal routing rule Routes items over a threshold Reviews the weekly exception list Project manager
Slip prediction Ranks activities by risk with a score Walks the work and tests the top items Superintendent
PPE detection on the gate camera Flags frames it classifies as non-compliant Confirms on site before any conversation with a worker The employer's named safety duty-holder
Drafted owner letter Produces a draft Verifies every quantity, date and clause; edits; sends The licensed contractor, through whoever signs

Remember this

  • "AI" covers four families on a construction project — rules, machine learning, computer vision and generative AI — and each fails differently.
  • Match the check to the failure: test rules, sample predictions, verify detections on site, check generated content line by line.
  • A confidence score is the system's own estimate, not proof about this worker, this fixture or this activity.
  • The published construction evidence supports theory building, not causal claims of performance improvement; sort the feature before you believe the claim.
  • The tool never changes who owns the work — the contractor and its qualifying individual, the design professional, the safety duty-holder, or whoever signs.

Construction example

Kestrel Ridge Builders is a fictional twenty-two-person general contractor doing light commercial work. Before a monthly operations meeting, owner Renata Vasquez asks project engineer Dmitri Salas to list every feature in their software that is sold as AI and say how each one is checked. Dmitri finds four. The first is a workflow rule in the project management system that routes any submittal with a value above $25,000.00 to the project manager. It is rules-based automation. He tests it against the last two months and finds it missed a $31,400.00 curtain wall submittal because the value field had been left blank, so the rule saw nothing to compare. He fixes the rule to route blank-value submittals as well, and adds a weekly exception list. The second is a slip-prediction feature in the schedule that scores activities by risk. It is machine learning. It has ranked the roof deck as low risk on the strength of past projects, none of which were winter pours. Dmitri notes that it was not wrong about the pattern, it simply had no pattern for this situation, and he asks that any activity with weather-dependent work be walked regardless of score. The third is a gate camera that flags missing hard hats. It is computer vision. Dmitri reviews a day of flags and finds eleven, of which three are the same worker walking under a low sun toward the lens, hard hat on, brim shadowed out. He records that the camera is a prompt to go and look, never a finding about a person. The fourth is an assistant that drafted last week's letter to the owner about a delay. Its draft said the delay followed a "two-week material lead-time extension from the supplier." Nothing in the file says that; the purchase order and the supplier's emails show the delay came from a late design clarification. The sentence was generated because it read like a reason. Dmitri deletes it, writes the real cause from the RFI log, and records in the meeting note the family, the check and the owner for all four features.

Responsible-use warning

Do not let a high score or a polished paragraph shorten your review. A score is the system's own estimate from patterns it has already seen, and it is least reliable on the unusual job condition that matters most. A camera flag is a reason to go and look, not a finding about a worker. A fluent generated explanation can contain a cause nobody supplied and no document supports. None of these is evidence. Before you accept an AI output, name the family that produced it and apply the matching check: test the rule, sample the prediction, verify the detection on site, or check every quantity, citation and clause against the document itself.

Human-review requirement

The owner of the work decides whether each AI feature may be used and how it is checked, and in construction that owner changes with the output. The licensed contractor and its qualifying individual own the contract, the workmanship and anything sent out under the licence. The architect or engineer of record owns design, drawings and specifications. The employer, through the duty-holder named in the applicable safety standard, owns safety decisions. Whoever signs the payment application or the certified payroll owns that signature. The reviewer confirms that the family of each feature was identified, that the check matched its failure mode, and that no output was accepted on the strength of a score or a well-written sentence.

Practical exercise

Map the AI features in one workflow you already run. Use fictional or anonymised material only, and put no project, customer or employee data into any tool for this exercise.

  1. Choose one recurring task, such as weekly schedule updates, submittal routing, jobsite photo review, or owner correspondence.
  2. List every feature in the software you use for that task that is described as AI, smart, automatic, predictive or suggested. Include features already switched on inside software you licensed for another reason.
  3. For each feature, decide its family: rules-based automation, machine learning, computer vision or generative AI. If you cannot tell, write "unknown" and the name of the person you will ask.
  4. For each feature, write one sentence on how that family fails.
  5. For each feature, write the check that matches the failure: a rule test, a sample of flags, a look at the real thing on site, or line-by-line verification against the drawing, specification, adopted code or executed contract.
  6. Complete a four-column table: feature, what the system does, what a person does, who decides.
  7. Name the owner of each output from the list in this lesson, and say which one applies and why.
  8. Circle any feature where the current check does not match the family, and write one change you would take to your owner or project executive.
  9. Keep the table with your workbook notes. You will reuse it when you build the risk matrix later in this Level.

Knowledge check

  1. Why should a submittal routing rule and a generated owner letter not be checked in the same way?

    Answer guidance: The rule is explicit logic that fails consistently, so it is tested once and its exceptions reviewed. A generative system writes new content every run and fails plausibly, so every quantity, date, citation and clause in each output must be verified against the document itself.

  2. What does a confidence score on a slip prediction or a PPE detection tell you, and what does it not tell you?

    Answer guidance: It is the system's own estimate of how strongly its learned patterns point to that answer. It does not prove anything about this activity or this worker, and it is least reliable on unusual conditions that do not resemble the data the system learned from.

  3. What did Vrcelj and Sandanayake report about the 34-study corpus, and what limits apply to it?

    Answer guidance: Only one study provided clear evidence of implementation in a live project and two others approached an identifiable intervention pathway, so the evidence supports theory building rather than causal claims of performance improvement. It is a preprint, not peer reviewed, and a 34-study corpus from one seed database is not an exhaustive review.

  4. What must you say about the ABC 2025 AI Resource Guide whenever you rely on it?

    Answer guidance: That it is a trade association resource with a vendor-supplied component: voluntary, not a regulator's position, creating no legal obligation, and not an industry standard of care.

Sources and further reading