A safety walk produces two things: hazards corrected today, and a record. Most programs do fine on the first and poorly on the second, because the record is written in a hurry by someone whose real job is making the site safe, not making the file defensible. This guide covers what a safety-walk record needs to contain, why tying findings to OSHA standards is worth the extra step almost nobody takes, and how to get both without slowing the walk down.
What a defensible finding contains
A safety observation that will still be useful in six months—to a safety director spotting trends, or to counsel after an incident—needs more than "fall hazard near stairs." Each finding should record where the condition was (building, level, gridline or room, precisely enough that a stranger could stand in the spot), what the condition actually was in concrete terms, a photograph, the corrective action required, who is responsible for it, and when it was found and when it was corrected. The photo matters more than any of the text: it establishes the condition as it existed, and after abatement it establishes that too.
Two disciplines separate strong programs from weak ones. First, findings name a responsible party—a subcontractor, not "site"—because unowned corrections don't happen. Second, closure is documented with the same rigor as discovery. An open finding with no recorded correction is worse than no finding in a dispute, since it proves the hazard was known.
The citation gap
Federal construction safety standards live in 29 CFR 1926, organized into subparts by subject—Subpart M for fall protection, Subpart L for scaffolds, Subpart K for electrical, Subpart X for stairways and ladders, and so on. In principle, every finding on a safety walk maps to a part and subpart. In practice, almost no walk reports cite them, because doing it properly means a code book and time, and the walker has neither on a Tuesday morning walk. So findings get filed under free-text descriptions, each written in whatever words occurred to that walker that day.
The loss isn't pedantry. Uncited findings can't be aggregated: "guardrail gap at slab edge," "missing top rail," and "fall exposure at leading edge" are the same Subpart M problem wearing three descriptions, and no spreadsheet filter will ever group them. Cited findings turn a pile of walk reports into a dataset.
What classification buys you
Once findings carry standard references, questions that were unanswerable become routine. Which violation categories recur across your projects, and are they the ones your training program emphasizes? Which subcontractors account for the repeated Subpart L findings, and does their prequalification score reflect it? Is this quarter's fall-protection trend improving or worsening, project by project? OSHA publishes its most-cited standards annually—fall protection has led it for years—and a classified dataset lets you compare your own sites against that list rather than against impressions.
The same data sharpens the response. Toolbox talks can target the subparts actually appearing on your walks instead of rotating through a generic calendar, and a sub with a pattern gets a specific conversation backed by specific findings. Safety programs run on credibility, and nothing is more credible than your own classified numbers.
Getting citations without a code book
The reason citation programs fail is workload, so the fix has to remove work rather than add it. This is a problem AI happens to fit well: recognizing that a photographed condition implicates a particular standard is exactly the kind of reading-and-matching that current models do reliably, with a person confirming the result.
In Bex Walks, a walk run in safety mode works the way the walker already works—narrate each finding by voice, photograph it, move on. Server-side, Bex transcribes and professionalizes the narration, then analyzes every photo for safety violations, including conditions the walker didn't mention, and classifies each finding by OSHA part and subpart with its reasoning stated. Firms with internal standards stricter than 1926 can have their own rules applied as well. The walker's expertise stays in charge—the AI's classification and reasoning are there to be reviewed, consistent with how we keep humans in the loop—but the code-book hour disappears, and every finding arrives in the report ready to assign and ready to report on.
Because the report is generated at submission, the safety walk also stops competing with the walker's afternoon. The two-hour walk plus one-hour writeup becomes the walk alone, which in our experience is the difference between a walk cadence that holds and one that erodes under schedule pressure. The capture mechanics—voice, photos, offline operation, GPS—are the same ones covered in voice-first site documentation.
A note on scope
Documentation discipline is not a safety program by itself; it presumes competent-person inspections where standards require them, real corrective authority, and a culture where crews flag hazards without waiting for the walk. What the record does is make the program legible—to your own leadership, to insurers, and to anyone who later asks what you knew and what you did about it. On that question, "here is every classified finding and its closure date" is the answer you want to be holding.
If your safety walks end in vague findings and a typing session, Bex can change both on the next walk you take. Get in touch below and we'll show you a safety-mode report from the kind of site you run.