Your Safety Program Lives in Nine Places

Ask a contractor whether they have a safety program and the answer is yes. Ask where it is, and the honest answer is a list:

Manuals. Binders. PDFs. Email threads. SharePoint. A training platform. Manufacturer documents. Regulator websites. And the ninth place — the deepest and most fragile — the memory of your most experienced superintendent, one of the fifth of the workforce now over 55.

Last week's post argued that AI's honest role in safety is retrieval, documentation, and escalation. This one is the build guide: how nine places become one — and why the defining property of the system is not intelligence but control.

Why "controlled" is the word doing all the work

The naive version of this project is: dump everything into an AI tool, let crews ask questions. That version is worse than the binder, for a reason worth being precise about.

A general-purpose model answers from everything: your 2019 procedure and your 2024 revision, your state's rule and another state's, the manufacturer's manual and a forum thread — all delivered with identical confidence. Fragmentation at least looked unreliable. An uncontrolled assistant launders folklore into fluent, cited-sounding prose. For hazard questions, that is not a quality problem; it is a liability engine.

A controlled assistant inverts the design. It may answer only from an approved library, and every answer carries its papers:

| Every answer must include | Example | |---|---| | Source document | Company Trenching & Excavation Procedure | | Revision / effective date | Rev. 6, effective March 2026 | | Applicable jurisdiction | Federal OSHA + state plan, if applicable | | Exact section | §4.2, Protective Systems | | Human-approval flag | "Competent person must verify on site" |

And one more property, the one that makes it trustworthy: when the library does not contain the answer, the system says so and escalates to the competent person. A safety assistant is defined by what it refuses to improvise.

The build, in order

1. Inventory the nine places. Collect every safety artifact into one staging pile: program, procedures, training records, certs, toolbox talks, manufacturer instructions, applicable standards.
2. Adjudicate versions. For every document, one question: is this current? Superseded revisions get archived out of the answer path — this single step retires years of folklore.
3. Tag what remains. Source, revision date, jurisdiction, work types it applies to. Tedious, mechanical, and exactly the preparation work AI now does cheaply with human spot-checks.
4. Wire the escalation path. Unresolved questions route to a named person, and the routing is logged. The log itself becomes safety data: what crews keep asking is a map of where the program is unclear.
5. Put intake on rails. New revisions enter through one door — reviewed, tagged, replacing their predecessors — or the library rots back into a tenth place.

Notice the sequence is 80 percent librarianship and 20 percent AI. That ratio is the point. The intelligence layer is a commodity; the controlled library is the asset, and it is yours.

The eleven-year question

NCCER estimates it takes about eleven years to train a craft professional to full equivalence with the veterans now heading toward retirement. Nothing about a document library replaces that. But every hazard judgment your best superintendent has ever written down, corrected in a toolbox talk, or flagged in a near-miss report is either captured in the controlled library — with his name on the revision — or it retires with him.

The safety case for this system is the strongest one. It is also a rehearsal: the same architecture — controlled sources, versioning, provenance, escalation — is exactly what the next two posts apply to a domain with less blood and just as much money: specifications, codes, and the question of which document governs.