Agents in the Trades
Safety

Incident and near-miss reports by voice

A near miss happens at 9:40 on a rooftop array. A module slips from a rail clip while two installers are carrying it, and it lands on the deck a few feet from a worker's boot. Nobody is hurt. The...

The problem

A near miss happens at 9:40 on a rooftop array. A module slips from a rail clip while two installers are carrying it, and it lands on the deck a few feet from a worker's boot. Nobody is hurt. The crew stops, talks it over, and goes back to work.

The report is a different story. The foreman means to write it up, but there is a delivery to receive, an inspector on the way, and a crew waiting on direction. By the time he sits down with the form it is late afternoon. He remembers the module and the clip, but not which row, who was on which end, or what the wind was doing.

Late reports are thin reports. Thin reports hide the cause, and the cause is the reason to write the report in the first place. Near misses get skipped entirely because the form feels like a lot of effort for something where nobody got hurt.

How the agent workflow works

The foreman talks. Right after the crew is safe and the area is secured, he records a voice note on his phone and just says what happened, in the order it happened. Who was there, where, what task was underway, what went wrong, what was done right after.

The agent reads that transcript and does four things:

  1. Drafts the company's incident or near-miss form from the transcript, filling in only what was actually said.
  2. Pulls context from sources it can read, such as that day's daily log, the crew roster, and the weather entry, and marks each pulled item as sourced from the log, not from the foreman.
  3. Flags gaps as plain questions. Which row? Was anyone below? Was the area barricaded? Was the module being carried by hand or on a cart?
  4. Sends the draft back to the foreman's phone.

The foreman answers the questions by voice, fixes anything wrong, and reads the draft against what he saw. Then he signs off. That signature is the checkpoint. The agent does not file, send to the safety manager, or notify anyone outside the foreman until he approves. If he never approves, the draft stays a draft.

The agent also does not assign cause or fault. It can list contributing factors the foreman mentioned. Deciding root cause is for the foreman and the safety lead.

What you need to set it up

Setup is a few hours, and most of it goes to the form template. Expect to run it against two or three old incidents to see where the drafts miss. Ongoing cost is whatever the transcription and AI subscriptions run, with no per-report fee unless your tools charge one.

Where it breaks

Transcription mangles noisy recordings. Wind, generators, and inverters can garble names and part numbers. The foreman has to read the draft, not skim it.

The agent fills gaps with guesses if the instructions are loose. Tell it to leave a field blank and ask rather than infer. Check this during testing.

Pulled context can be wrong. If the daily log was written hours earlier or the roster was not updated, the agent repeats the error with confidence. That is why sourced items are labeled.

Serious injuries are not for this tool. Those need your formal reporting process and the people responsible for it, and the draft is not a substitute for notifying them immediately.

Foremen may start signing without reading, since the draft looks finished. If sign-off becomes a formality, the report is only as accurate as the agent's guesses. Spot-check signed reports against what crews remember.

Witness statements are another limit. The agent can draft from the foreman's account but cannot collect what others saw unless they record their own notes.

The bottom line

The agent shortens the gap between the event and a usable report, so the details are captured while they are still sharp. It does the typing and the question-asking. The foreman still decides what is true, and nothing is filed without his signature. Start with near misses, where the stakes of a rough draft are low, and tune the form template before extending it further.

Related use cases

← All use cases

One working agent workflow, every day.

Real AI agent use cases from solar, construction, and field operations — what they do, what they cost, and where they break. No hype, no theory.