Lesson objective
Identify whether a construction AI product retrieves documents, calls tools or acts as an agent, and set the human control that matches what it is permitted to read and change.
Plain-language explanation
Why this matters now
Very few construction AI products are a bare language model. Most wrap the model in extra machinery: a search across the project documents, a connection to the schedule or the accounting system, a calculator, access to a shared inbox. That machinery is what makes the product useful, and it is also what changes the risk. A model that can only write text can mislead you. A system that can read a subcontractor's email and then issue a purchase order can act on the mistake before anyone reads it.
The move toward systems that act is already being measured, if only outside construction. The Thomson Reuters Institute 2026 AI in Professional Services Report (published 9 February 2026; more than 1,500 respondents from legal, tax and accounting, corporate functions and government in 27 countries) reports that 15% said their organisations had adopted some type of agentic AI tool and 53% were actively planning or considering it. Label it properly when you repeat it: it is vendor research from the research arm of a company that sells AI products, it combines all professional-services respondents rather than any one profession, and the fieldwork dates and sampling method were not stated on the pages reviewed for this course. It is not a construction figure and must never be presented as one.
What construction does have is a clear statement of the missing piece. Vrcelj and Sandanayake, in their 31 July 2026 preprint on AI-enabled construction project controls, frame the gap as "governed decision translation": how an AI output is "interpreted, validated, challenged, authorized, assigned for implementation, documented, and reviewed." That sentence is a control design, and the rest of this lesson builds one from it.
A question to sit with
An assistant in your project management system reads every email that lands in the shared project inbox. Who, other than your own staff, can now put words in front of it?
What goes wrong
Assuming retrieval makes the answer true. A product that searches your project documents first does sound safer, and often is. But it can fetch a superseded drawing, the proposal instead of the executed contract, or nothing relevant at all — and the model can still misstate the passage it was handed.
Not knowing what the tool can change. A feature described as a helper may be able to issue an RFI, send an email, update a schedule or change a vendor record. If nobody has written down what it can change, nobody is controlling it.
Trusting content that came from outside your company. Supplier emails, subcontractor schedules, owner-furnished PDFs and drawings from a shared portal are written by other people. When a system reads them, their wording enters the same stream as your instruction.
Relying on a product claim instead of a permission. No product claim removes the risk that content it reads influences what it does. Controls reduce likelihood and impact; they do not bring either to zero.
Keeping no record of what the system did. Without a log of inputs, outputs and actions, nobody can reconstruct why a purchase order went out or why a schedule changed — and a claim two years later is exactly when someone will ask.
The practical answer
Learn to see three layers inside any product. Each adds usefulness and adds a question.
| Layer | What it does, in plain terms | Construction illustration | The new question it raises |
|---|---|---|---|
| Retrieval | Before answering, the system searches a set of documents and puts relevant passages in front of the model | An assistant that searches the project specification and submittal log | Did it fetch the current, executed, issued-for-construction document, and does the answer match that passage? |
| Tools | The model can ask other software to do something — run a calculation, query the schedule, read a file — and gets the result back | An assistant that pulls the current schedule before drafting a two-week look-ahead | What can each tool read, and what can it change? |
| Agents | The system plans a sequence of steps toward a goal and calls tools along the way, with less step-by-step direction from a person | A system that reads incoming subcontractor emails, matches them to commitments and prepares change requests | What may it do without a person approving, and how is it stopped? |
The permission ladder. Put every AI feature on a rung before anybody uses it. The rung depends on what the system may do, not on how clever the model is.
| Rung | What the system may do | Construction example | Minimum control before use |
|---|---|---|---|
| 1. Read and answer | Read supplied or retrieved material and produce text for a person | Summarise a specification section | A person verifies the answer against the section itself |
| 2. Draft | Prepare a document that is not saved to any record | Draft a two-week look-ahead from the current schedule | A person checks every date and quantity and edits before use |
| 3. Propose | Create a pending item that waits in a queue | Prepare an RFI or a change request for approval | A named person approves each item, and the approver is never the system |
| 4. Act with approval | Carry out an action once a person confirms it | Issue a submittal transmittal after sign-off | Approval recorded, action logged, permissions limited to that action |
| 5. Act alone | Change records or communicate without per-item approval | Issue purchase orders or send owner correspondence automatically | Not appropriate for the work taught in this Level; see Module 12 |
Five rungs, five rows.
Controls that do not depend on the model behaving. The strongest safeguards are the ordinary ones: a permission the AI account simply does not have; a purchase order limit enforced by the accounting system rather than by a sentence in a prompt; a vendor bank-detail change locked to a separate verified call-back process; correspondence to the owner that cannot leave without a named signature. A control written into a prompt is a request. A control written into permissions is a control.
Turn "governed decision translation" into a checklist. The preprint's seven verbs make a serviceable one for any AI output that could change a project, and it is a preprint, not peer reviewed, so treat the framing as useful language rather than an authority. For each output, name a person against each step.
| Step (preprint's wording) | On your project, that means | Named person |
|---|---|---|
| Interpreted | Someone states what the output actually claims | |
| Validated | Someone checks it against the drawing, specification, adopted code or executed contract | |
| Challenged | Someone with standing can say no, and does so on the record | |
| Authorized | Someone with authority approves the action | |
| Assigned for implementation | Someone is named to carry it out | |
| Documented | Prompt, output, tool, date and decision are saved | |
| Reviewed | Someone looks back later at whether it was right |
NIST AI 600-1 (26 July 2024) names Human-AI Configuration and Information Security among the 12 generative AI risks it lists, and describes Human-AI Configuration as including humans "experiencing algorithmic aversion, automation bias, over-reliance, or emotional entanglement" with these systems. That is a voluntary, cross-sectoral profile; nothing in it is written for contractors.
Questions to ask the vendor or whoever administers the software, and record the answers.
- What does the feature retrieve from, and who can add or change documents in that source?
- Which tools can it call, and for each one, what can it read and what can it change?
- What content written outside our company does it read — email, portal uploads, supplier documents, web pages?
- Which actions require a person to approve, and can a user switch that off?
- What is logged — inputs, outputs, tool calls — where, and for how long?
- How is it stopped, and who exactly can stop it?
Remember this
- Most products add retrieval, tools or agent behaviour around the model, and each layer adds a question.
- Retrieval can fetch a superseded drawing or the wrong version of a contract, and the model can still misstate the right one.
- Content written outside your company enters the same stream as your instruction; limit what the system can do after it reads it.
- Put every feature on the permission ladder, and prefer controls the model cannot talk its way past.
- Name a person against every step of interpretation, validation, challenge, authorisation, implementation, documentation and review.
Construction example
Kestrel Ridge Builders is piloting an assistant inside its project management system on an invented job, the Alder Point Clinic fit-out. Project executive Renata Vasquez asks contract administrator Joyce Oyelaran to find out what the assistant actually does before it goes live. Joyce works through the vendor questions and writes the answers down. The assistant retrieves from two places: the project document library, and a shared folder of subcontractor schedules that any of the six subcontractors can upload to through the portal. It calls three tools. One reads the shared project inbox, one reads the schedule, and one creates commitment change requests. Joyce asks what "creates" means and learns the default setting saves change requests as issued, not pending. That puts the feature on rung 5 of the permission ladder: it acts alone. It also reads outside content, because subcontractors and suppliers write the emails and the schedules. Joyce writes up two concerns. First, anything written inside an emailed schedule reaches the same model that can issue commitment changes. Second, the subcontractor folder has no version discipline, so a superseded schedule can be retrieved and quoted as current. With Renata she agrees three changes before the pilot. The change-request tool is set to save as pending, which moves the feature to rung 3, and only Renata and the project manager may approve. Vendor bank-detail changes are taken out of the assistant altogether and stay in the existing call-back process. The subcontractor folder is restricted so that only the project engineer files an uploaded schedule, and every file carries a transmittal date. Joyce confirms the vendor logs inbox inputs, outputs and tool calls for the pilot period, and names Renata as the person who can switch the assistant off. In the first three weeks the assistant prepares 46 pending change requests; the project manager rejects 7, each for a mismatch the log lets her trace back to the email that produced it.
Responsible-use warning
Do not approve an AI feature from its description alone. "Helps with change requests" can mean anything from summarising an email to issuing a commitment change against a subcontract. Find out what it retrieves from, which tools it can call, what each tool can change, and what outside content it reads. A system that reads supplier and subcontractor documents and can also change records combines the two conditions that make a steered output expensive. Never accept a vendor assurance in place of a permission: set what the account cannot do, and check what it produced.
Human-review requirement
The owner of the process approves the rung each AI feature sits on and the controls that go with it. For anything that leaves the company under the licence — correspondence, a transmittal, a commitment change — the licensed contractor, through the person who signs, confirms that a named individual approves each item and that the approval is enforced by the software rather than by the prompt. Where an output touches design, the architect or engineer of record remains the owner. The reviewer checks that the vendor answers were recorded, that outside content reaching a system with change permissions was identified, that logs exist, and that a named person can stop the system.
Practical exercise
Place one AI feature on the permission ladder. Use a product your company uses or is considering, and fictional transactions only.
- Name the product and the single feature you are assessing. Do not try to assess a whole product at once.
- Record whether the feature uses retrieval, tools, agent behaviour, or none of these, and note where you found the answer.
- List every source it retrieves from, and who can add or change documents in each. Mark any source without version or transmittal discipline.
- List every tool it can call. For each, write what it can read and what it can change.
- List every kind of content written outside your company that it reads: supplier email, portal uploads, owner-furnished documents, subcontractor schedules.
- Place the feature on a rung from 1 to 5 and write one sentence explaining the placement.
- Write the minimum control for that rung, plus one safeguard that does not depend on the model behaving, such as a permission the account lacks or a limit enforced by the accounting system.
- Complete the seven-step table from this lesson for one output the feature produces, naming a person against interpretation, validation, challenge, authorisation, implementation, documentation and review.
- Record what is logged, where, and for how long, and name the person who can stop the feature.
- Decide whether the feature may be used now, used only after a change, or not used, and record who made that decision and when.
Knowledge check
-
Name the three layers commonly built around a language model in a construction product, and one question each raises.
Answer guidance: Retrieval (did it fetch the current, executed or issued-for-construction document, and does the answer match it?), tools (what can each tool read and change?) and agents (what may it do without a person approving, and how is it stopped?).
-
Why does the rung on the permission ladder depend on permissions rather than on how capable the model is?
Answer guidance: Because the harm comes from what the system can change or send after it is wrong or is steered. A very capable model with read-only access cannot issue a purchase order; a weak one with issuing rights can.
-
What do Vrcelj and Sandanayake mean by "governed decision translation", and what is the status of that source?
Answer guidance: How an AI output is interpreted, validated, challenged, authorized, assigned for implementation, documented and reviewed. It comes from a preprint posted 31 July 2026 that has not been peer reviewed, so it is useful framing rather than an authority.
-
Why is a control written into a prompt weaker than one written into permissions?
Answer guidance: A prompt asks the model to behave, and the model is predicting text; a permission is enforced by other software regardless of what the model produces or what content it has read.
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.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- Thomson Reuters Institute — 2026 AI in Professional Services Report (2026-02-09). Accessed 2026-09-15.
- Associated Builders and Contractors (ABC), guide produced by GCM — 2025 AI Resource Guide for Construction Professionals. Accessed 2026-09-15.