Guide
AI for construction project management is chosen one task at a time: which kind of AI should prepare the first draft, which checks must stay in software, and who approves the result.
The gap between interest and use is large. In Mastt’s State of AI in Construction Project Management 2026 survey of 108 project management professionals (March to June 2026), 66.7% had used ChatGPT for work in the previous three months, 3.7% the AI inside Procore or Autodesk, and 9.3% used AI agents. Reporting was the area most respondents wanted help with (84.3%). The sections below map the common project management tasks to four options and give the tests for when a custom workflow beats a tool.
Hugo Sequier runs State Method, which sells the assessments, pilots and builds these guides discuss. Apply the same tests to it as to any other provider.
01 / Options
Each option has a different owner, a different view of the project record and a different failure mode. A firm can use more than one.
ChatGPT, Microsoft Copilot, Claude or Gemini, used for drafting, summarising and rewording. It only sees what someone pastes or uploads, keeps no project audit trail and must follow the firm’s rules on confidential documents. Every fact it writes needs checking against the record.
Features added to the platform that already holds the RFIs, submittals, daily logs and drawings. It works on records that are already there and inside that platform’s workflow. Ask what it can read outside the platform and what the vendor does with project data.
A product built for one task, such as schedule scenario analysis, photo-based progress capture, quantity takeoff or document search. Strong when the task matches what it was built for; it adds a subscription, an integration and another place where records live.
A workflow built around the firm’s own process: it reads the systems involved, prepares a draft, runs fixed checks in software and routes the result to a named approver. It costs more up front and needs an evaluation on real cases, so it is only worth it when the tests in section 03 apply.
02 / Task map
The usual first choice for common project management tasks, and the condition that moves the task to a custom workflow. The examples linked below use synthetic project data.
Start with a general assistant, with the author checking every date, quantity and commitment. Nothing here needs a build unless the same letter is produced from the same records every week.
Start with the platform’s own features or a progress-capture tool. Consider a custom workflow when the report combines several sources, must follow an owner’s format and has to state which evidence each line rests on.
Start with the platform’s features when every record already lives in it. Consider a custom workflow when intake arrives by email, the specification sits elsewhere and the register is a separate system.
Start with the document control already in use. Consider a custom workflow when revisions arrive from several parties and someone must trace which open items each change affects.
Start with a specialist scheduling tool or the scheduling software already in use. The planner keeps the decision on any re-sequence.
Keep the numbers in the cost or ERP system, where they are calculated rather than generated. A model can draft the narrative around a variance; it should not produce the figures.
Start with the platform’s closeout tools. Consider a custom workflow when the handover set must be assembled from several systems against an owner’s list and every gap has to be shown.
03 / Build tests
A custom workflow is justified when several of these are true for one named task. One alone is rarely enough.
If none of these is true, buy or configure the tool. A custom build that repeats what a product already does is cost without a reason.
04 / Buy tests
The last point is not a reason to buy a product. It is a reason to name an owner first, because neither a product nor a build can be judged without one.
05 / Trial
The same trial works for an assistant, a platform feature, a specialist tool or a custom build, so the options can be compared on equal terms.
The AI proof of concept guide turns this trial into a one-page scope with thresholds, and lists the data-handling questions to put to a vendor first.
06 / Control
Whichever option drafts the work, the split of responsibility should be written down before use: the model proposes, software enforces the fixed checks, and a named role approves.
Cost
A general assistant is a per-seat subscription and the cheapest to try. Platform AI may be included in a plan or sold as an add-on, so ask before assuming either. Many specialist tools publish no price and quote by project or seat.
A custom workflow starts with an AI readiness assessment of the one task, not with a build.
FAQ
It can draft minutes, letters, summaries and status updates, provided the firm’s rules on confidential documents allow it and the author checks every fact. It only sees what is pasted or uploaded and keeps no project audit trail, so it should not decide RFI answers, submittal statuses or change order values.
Continue with the evidence
Assessment / one workflow
Describe the task, the systems it passes through, its owner and how it is measured today. A person reviews every assessment.