The AI-Assisted Safety System

Construction remains one of the most dangerous industries in the country: 1,075 workers died on the job in 2023, roughly one in five of all U.S. worker deaths, with the "Fatal Four" — falls, struck-by, caught-in/between, electrocution — accounting for the clear majority. Falls alone killed 421 people.

Which is exactly why this post has to begin with what AI cannot do. This series drew the boundary a month ago; safety is where the boundary is load-bearing.

AI may help retrieve the rule. It cannot see the ground move, smell the gas, inspect the rigging, or assume the responsibility of the competent person standing on the site.

No model output substitutes for a competent person, a qualified person, site-specific judgment, formal training, required inspections, engineering, or the employer's non-delegable responsibility. A vendor claiming their AI "ensures OSHA compliance" is describing something that cannot exist, and a contractor who believes it is buying liability with a subscription fee.

The real problem: your safety program is information-rich and retrieval-poor

Here is what an honest look at most contractors' safety programs finds — not an absence of safety information, but a scatter of it. Procedures in a binder from two revisions ago. The current versions in a PDF someone emailed. Training records in a spreadsheet, certifications in a drawer, manufacturer instructions in a truck, toolbox talks wherever the last safety consultant left them, and the sharpest hazard knowledge in the head of one superintendent who may retire within the decade along with 41 percent of the workforce.

The dangerous moment is not the missing rule. It is the foreman with a real question at 6:50 a.m. who cannot reach the answer — so the crew proceeds on memory and habit. Fragmentation converts good documentation into folklore.

What AI honestly contributes

Scoped to retrieval, documentation, communication, and escalation, the contribution is substantial:

• Retrieve the current procedure — with its source, revision date, and applicable jurisdiction attached, so "current" is a fact and not a guess.
• Draft project-specific checklists from your approved program, for a human to edit and own.
• Prepare toolbox-talk materials on the hazard that actually matters this week.
• Translate approved instructions for multilingual crews — retaining the English-controlled version, because translation is delivery, not authorship.
• Record attendance and acknowledgments without a clipboard chase.
• Convert field observations — a voice note, a photo — into structured reports that preserve the evidence.
• Track certifications and expirations; flag the gap before mobilization, not after the incident.
• Route incidents and near misses to the right people, on time, documented.
• Analyze recurring observations: the same ladder finding on three jobs is a pattern, not three coincidences.
• Escalate — the most important verb on the list. A question the system cannot answer from approved sources goes to the competent person, marked unresolved. It never improvises an answer to a hazard question.

Notice what every item shares: the output lands in front of an accountable human with better information than they had, sooner. Nothing on the list decides anything on a site.

What this buys, concretely

Three quiet wins, none of which show up in a demo:

1. The program becomes answerable. The 6:50 a.m. question gets the current, cited procedure in seconds — folklore loses market share to documentation.
2. The paperwork becomes an asset. When an inspector, an insurer, or a plaintiff's attorney asks for training records, acknowledgments, and inspection documentation, the answer is organized and timestamped rather than archaeological.
3. Patterns surface before incidents do. Near-miss and observation data, actually analyzed, is the cheapest leading indicator in safety — and it is exactly the data most companies collect and never use.

The through-line of this series holds: prepare, retrieve, flag, escalate — never decide. Next week: the practical build — how a scattered safety program becomes one controlled, citable source of truth, and why "controlled" is the word that does all the work.