ChatGPT for construction: what works and what breaks

Generic ChatGPT is genuinely useful on a job site, just not for the things people assume. It is a strong writing assistant when you give it text. It is unreliable the moment a task depends on your drawings, your specs, or numbers that have to be right. Knowing where that line sits saves you from sending a hallucinated quantity to an owner or pasting a confidential bid into a public tool. Below is the honest split between what works, what breaks, and when a contractor should move to AI built for the work.

What does ChatGPT do well for contractors?

The reliable wins are language tasks where you supply the facts and the model shapes them.

• Drafting RFIs from a description you type in, then tightening the wording so it is specific and answerable.
• Turning rough scope notes into a clean subcontract scope or a clarification list.
• Writing and softening emails: chasing a late submittal, pushing back on a directive, documenting a verbal in writing.
• Summarizing a meeting from a transcript or your notes into action items with owners.
• Explaining an unfamiliar spec section or code reference in plain language so you know what to verify.

In every one of these, you are the source of truth. ChatGPT is rephrasing, structuring, or speeding up the typing. That is a fair use of it, and it is faster than a blank page.

Where does ChatGPT break on real project work?

It breaks wherever the task quietly requires access to your actual documents or a number that must be defensible.

Scanned and image-based drawings. A large share of construction documents are scanned PDFs, photographed sheets, or flattened plan images. Generic ChatGPT can describe what a clean sheet looks like, but it cannot reliably read a marked-up scan, follow a detail callout across sheets, or trace a dimension string. It will guess at what it cannot see.

No access to your specs, contract, or submittals. ChatGPT does not know your Division 09 finish schedule, your contract's notice provisions, or which submittal revision is current. Ask it about "the spec" and it answers from generic industry patterns, not your project. That is the difference between sounding right and being right.

Hallucinated quantities and clauses. This is the dangerous one. Ask for a takeoff, a unit count, or a contract citation and it will return a confident, specific, wrong answer when it does not actually have the data. There is no flag that says "I am guessing." For estimating and quantities, you want a system that ties every number back to a marked sheet you can open, which is the model behind purpose-built AI construction takeoff and estimating software.

Data privacy. Pasting drawings, bid numbers, or contract language into a consumer chat tool means that content leaves your control. For competitive bids and owner-confidential documents, that alone is a reason to keep this work inside a system you govern.

A worked example: drafting an RFI the safe way

Here is the same task done two ways so the boundary is concrete.

Task: A foreman finds a conflict between the structural sheet and the mechanical sheet at a beam penetration.

Step by step with generic ChatGPT:

1. What goes in: You type a plain description of the conflict, the two sheet numbers, and the affected gridline.
2. What the AI does: It drafts a clean, properly formatted RFI with a clear question, the impact, and a proposed resolution for the engineer to confirm.
3. What the human checks: You verify the sheet numbers, the gridline, and the dimensions against the actual drawings yourself, because ChatGPT never saw them. You confirm the proposed resolution is reasonable. You add the cost and schedule note.
4. Where it would break: If you had asked ChatGPT to "find conflicts in these drawings," it cannot, because it cannot read your sheets. The drafting is the safe part. The reading is not.

The lesson holds across tasks. Use generic AI to write from facts you confirmed. Do not use it to source facts from documents it cannot open.

What does purpose-built construction AI do differently?

The core difference is grounding. A construction-specific system works from your documents instead of from general training data.

• It reads your actual sheets, including scans, using document processing tuned for plans, schedules, and specs rather than expecting clean typed text.
• It runs on your project record: drawings, specs, submittals, contracts, and correspondence, so an answer reflects this job, not a generic pattern.
• It cites its source. A quantity or a clause comes back with the sheet or page it came from, so you can verify in seconds instead of trusting blind.
• It runs in your tenant. With Ruh, the AI reads your documents inside your own environment, does the work, and routes it to your team. Your bid numbers and contracts do not get pasted into a public tool.

This is the practical answer to hallucinated quantities and privacy at once. When every output traces to a document you control, "is this right?" becomes a check you can actually perform.

How should you decide which tool for which task?

A simple test: does the task require reading your project documents, or producing a number or clause that has to be defensible?

• No to both: generic ChatGPT is fine. Email polish, a meeting recap from your notes, plain-language explanation of a code section, a first-draft scope from your bullets.
• Yes to either: use purpose-built construction AI. Takeoffs and quantities, reading scanned drawings, anything quoting your spec or contract, anything competitive or confidential.

You do not have to pick one tool for everything. Many teams keep ChatGPT for writing and bring in a grounded system for document-heavy work like estimating, submittal logs, and change order review.

Does the human still own the decision?

Yes, on both paths, and that should not change. AI assists. The estimator, the PM, and the principal keep judgment and sign-off. The right setup makes that review fast and honest: the AI does the first pass and shows its work, a human checks against the cited source, and a person approves before anything goes to an owner, an architect, or a sub. A tool that hides its sources makes that review impossible. A tool that grounds and cites makes it routine.

The takeaway is not "avoid ChatGPT." It is to match the tool to the task. Let generic AI carry the writing where you already hold the facts. For the work that touches your drawings, your specs, and your money, use construction AI that runs on your documents in your tenant, cites what it found, and hands the call back to your team. That keeps the speed without betting a bid or a claim on a confident guess.