ChatGPT for Property Condition Reports: What It Can and Can't Do for a PE

ChatGPT can turn your field shorthand into clean paragraphs for a property condition report, and it is good at that. It cannot tell you which of those sentences rest on what you observed, what the manager told you or what a document says, and it will supply a quantity, an install year or a cost if the prompt leaves a gap. Used for wording, with every fact checked against your notes, it saves typing; used as the author of the report, it puts unsupported facts above your signature.

This page sets out where it helps, where it breaks against ASTM E2018-24, a prompt that limits the damage, and what to check before anything it wrote goes out.

What can ChatGPT do well in a PCR?

Language work on material you supply.

• Turning shorthand into prose. "RTU-3 Carrier 2004 R-22, coil corroded, mgr says leaks, rec replace" becomes two readable sentences in the register lenders expect.
• Consistency. Rewriting ten system sections so they share tense, condition vocabulary and structure.
• Boilerplate drafts. First versions of a purpose and scope paragraph, a limiting-conditions paragraph or a transmittal letter, which you then correct against your firm's approved text.
• Explaining significance. Plain-language sentences on why a deficiency matters, which E2018-24 asks for where the significance is not obvious.
• Checking your own draft. Asking it to list every number in a section, or every sentence that states an age, is a fast way to audit your writing.

In every case the facts come from you and the output is checked by you.

Where does it break?

It doesn't keep a record of where a fact came from

E2018-24 asks the report to distinguish what the field observer observed, what was reported by the owner or point of contact, and what is documented. A chat session holds your notes, the owner's documents and the model's general knowledge in one undifferentiated context. When it writes "the roof was replaced in 2011", nothing records whether that came from the warranty, the manager or nowhere. You have to reconstruct the source of every statement by hand, which is most of the work you hoped to save.

It fills gaps with plausible numbers

Ask for a roofing section from notes that give no area and you may get "approximately 24,000 square feet of modified bitumen". Ask for an age and you may get a typical one. These read like facts and are the most dangerous sentences in a report, because a lender's reviewer will treat them as observed. The safe instruction is "if a number is not in my notes, write that it was not provided", and even then you have to check that it obeyed.

It doesn't build the cost tables

Table 1 and the replacement reserve are arithmetic: quantity times unit cost, segregated by category, the reporting threshold applied, years placed by remaining useful life, inflation compounded. A language model can format a table, but its sums and its unit costs are not a cost basis you can defend. Build the tables in a spreadsheet from your firm's unit costs and a published useful-life table such as HUD's, and paste them in.

It cites the standard from memory

It will produce plausible section numbers for E2018-24 and other standards. Some will be right. Check every citation against your copy of the guide; a wrong clause on a lender report is an easy credibility hit.

Confidentiality depends on the plan

OpenAI's policy says content from its services for individuals, such as ChatGPT, may be used to improve its models unless you turn that off in data controls, while its business products (ChatGPT Business, Enterprise and the API) do not use your data for training by default. Owner documents, rent rolls and lender correspondence are often under confidentiality terms; check your engagement letter before pasting them into any individual account.

What does a safer prompt look like?

Give it a role, the evidence, and rules that make gaps visible. A workable structure, in five parts:

1. Role. You are drafting one system section of an ASTM E2018-24 property condition report for a licensed engineer who will review and sign it.
2. Evidence. Paste the field notes for that system only, then any document excerpts, each labelled with its source (field notes, POC interview, roof warranty dated..., and so on).
3. Source rule. Write what was observed as observed, what the POC said as reported, and what a document says as documented, naming the document. Never upgrade a reported fact to an observed one.
4. Number rule. Use only numbers that appear in the evidence. If a quantity, age or date is needed and is not there, write "not provided" and list it at the end under "Missing for the engineer".
5. Output. Description, observations, recommendations with a suggested remedy for each deficiency, and no costs in the narrative.

Run one system per session so evidence does not bleed between sections, and keep the "Missing for the engineer" lists: they are your follow-up questions to the owner.

What should you check before it goes out?

• Every number in the prose against your notes and documents.
• Every age and date for its basis, and that the wording says observed, reported or documented correctly.
• Every standard citation against the guide.
• Every cost figure against your tables, which you built separately.
• Nothing from the model's general knowledge presented as a finding at this property.
• The report says, in your firm's words, that drafting was AI-assisted and that you reviewed it, if your firm or client requires that disclosure.

How is a drafting tool different?

The difference is not the model; it is what surrounds it. A drafting tool built for PCRs keeps the evidence structured, so each fact carries its source tag from the notes to the PDF, and it computes the tables itself rather than asking the model for numbers. Baseline PCR, for example, tags every fact observed, reported, documented or not provided, builds Table 1 and the reserve as quantity times your firm's unit cost with HUD useful lives and a basis sheet for every line, checks every number in the narrative against the data, and lists what it could not support for the reviewer before approval. The engineer still reviews, edits and signs; nothing is issued without that.

ChatGPT costs about $20 a month for an individual Plus plan on OpenAI's pricing page. The cost that matters is the desk time spent checking its output. If you want to compare, read the 65-page sample report drafted from 90 photographs and 7 owner documents, or run one job free. For the wider market, see PCA software compared, and for the insurance side of AI-assisted drafting, AI exclusions and E&O for engineering reports.