Build an AI-assisted photo evidence pack for HVAC proposals

Start with the buyer’s question

A proposal photo pack should answer a specific question: which asset is in scope, what was observed at the site, or what a comparable completed installation looks like. Those are different uses. Separate current-site evidence from reference examples before asking an assistant to draft captions. A completed installation at another property does not prove that the proposed equipment will fit this site.

Create a small manifest containing image ID, authorized location label, capture date, photographer, intended use and the source file link. Add an approval field for customer-facing reuse. Leave it pending when permission is unknown. Exclude access instructions, security details, unrelated people and private customer records from the working packet.

Draft captions from a reviewed manifest

Give the assistant the manifest and approved written observations. Request a table with image ID, proposed caption, supported proposal claim and an unanswered-question column. Require it to quote the relevant observation or record identifier next to each claim. Do not let it infer equipment condition, capacity, cause of failure or code compliance from an image. Qualified staff own those assessments.

For a fictional example, photo P-14 might show the labeled rooftop unit discussed in observation O-7. The caption can identify that relationship without claiming the photo proves replacement is necessary. If P-15 comes from a different property, its caption must say it is a reference example and explain why it is included.

Review the export, not just the draft

The estimator checks the photo-to-scope relationship; the account owner approves reuse and presentation. Inspect the actual proposal export for wrong-site images, cropped labels, unreadable annotations and captions that moved away from their image. A correct manifest does not guarantee a correct PDF. Keep the approved file version with the opportunity.

Try this on one proposal with a small image set. Record missing provenance, rejected captions and time spent finding originals. If review takes longer because sources are unclear, improve the image naming and approval process before expanding the workflow. Do not use AI-generated job photos as evidence of work your company performed.

Before you use the result

• Label every image as current-site evidence or a reference example.
• Record the capture date, source location and permitted audience.
• Require qualified review of technical statements; do not derive diagnoses from photos.
• Remove unapproved customer identifiers from reused examples.
• Check the exported proposal against the approved manifest.

What the sources support

Jobber documents a reusable files/media library. NIST describes the risk of confidently incorrect generated statements. The manifest, caption review and export checks here are our proposed workflow, not a vendor-provided approval system or a tested product result.