Reviewing your safety inspections with ChatGPT

How to connect ChatGPT to your project

Connecting takes a few minutes and happens once. Your own Dashpivot permissions carry across, so what you see in Dashpivot will be reflected in ChatGPT.

Prompts to start with

One rule matters more than the rest: it queries one form template at a time, so name the exact template, the project or team, and the dates. Ask for "all forms" and you get whatever it reached first.

• "What workspaces do I have access to? List the projects and teams."
• "What form templates are available in [team name]?"
• "Show me all [exact inspection template name] forms from [team], [project], created between [date] and [date]. For each one show the date, location and findings."
• "From those, which corrective actions are still open, and how long has each been open?"
• "Group those findings by location and tell me which one repeats most. Tell me if the results are complete or if some might be missing."

What a safety inspection is meant to produce

An inspection isn't paperwork for its own sake. It exists to find a hazard before it finds someone, put a name and a date against fixing it, and prove that happened. Most formats carry the same things:

• The hazard or non-conformance, described plainly
• Where on site it is
• A risk rating
• The control or corrective action
• Who owns it, and by when
• Photos of the finding
• Sign-off that it was closed out

If the answers come back thin

Two different problems look the same from the chat window. Either the connection is not doing what you think, or the form template never captured the thing you are asking about.

• Nothing connects at all: MCP Access is probably still off, and only an Org Controller can turn it on
• Sitemate does not appear in the app list: search "Sitemate" exactly
• The tools do not show in a chat: start a new chat, or disconnect and reconnect the app
• Results look short: it queries one template at a time, so run a separate query per template rather than asking for everything
• Findings sit in one free-text box, so nothing can be grouped. Split them into a repeating section with a field per finding
• Location is typed by hand and spelled differently each time. Use a register list so the same place resolves to the same value

Why findings go cold

Safety officers and supervisors spend hours after an inspection typing up findings, corrective actions and compliance notes. Photos get lost, handwritten notes are illegible, and reports sit incomplete for days while urgent issues wait on documentation.

The second failure is quieter. Once a finding is filed it stops being visible. Nobody is reading thirty inspections looking for the hazard that keeps coming back, so the pattern only becomes obvious after an incident.

What it can and cannot do

The honest limit is the records themselves. ChatGPT reading your project cannot tell you about a hazard nobody inspected for, and it won't invent one. If an area hasn't been inspected since July, that absence is the finding.

How the record gets there

Asking a question of your inspections only works because the inspections are already structured records rather than photos in a phone. Three parts do that, and only the last one involves ChatGPT.