Most EHS teams are still early with AI, and most of them started in the last 18 months. The uses that pay off first are the unglamorous ones: writing reports, tailoring training, cutting repetitive work. Predicting serious injuries and cutting costs are still out of reach, even for the teams furthest along. Data quality is the barrier that never goes away.
This summer the American Society of Safety Professionals surveyed nearly 1,000 safety professionals about how they use AI, what has worked, and what worries them. The report, Trust, Ethics and Governance: How Safety Leaders Are Thinking About and Using AI, was run with Kalorama Research and supported by the ASSP Foundation. It is one of the better benchmarks the profession has, because it asks practitioners rather than vendors. ASSP points readers to its AI in Safety user group to continue the conversation.
Here is what stood out to us, and what we think it means if you run a safety consulting firm or an in-house EHS program.
Almost everyone is early
ASSP sorts organizations into five stages. The split looks like this:
- 12% are not using or exploring AI for EHS.
- 28% are in discovery, learning what it can do.
- 30% are piloting a few narrow uses.
- 23% are scaling it across the organization.
- 5% consider it fully integrated.
Two thirds began in the last 18 months, and the report estimates it takes one to two years to move from experimenting to scaling. So if you feel behind, you mostly aren't. The people who feel ahead started about a year earlier.
Size matters, though. Among organizations with 5,000 or more employees, 45% are in the scaling or integrated stages. Among those with fewer than 250, it's 19%. Insurance leads every industry, with half of insurance organizations in those later stages. That last one is worth noting for anyone who writes loss control reports: the people reading them are further along than the people writing them.
What pays off first
The early wins are practical. Reducing time on repetitive tasks and getting more usable insight from data show up early and hold as organizations scale. When the survey asked people to rate specific uses, the most accessible ones also scored well:
- Report development: rated effective by 79% of the people using AI for it.
- Incident analysis and pattern recognition: 70%.
- Tailored education and training: 68%.
The weakest was real-time monitoring and alerts, at 49%. The pattern is consistent: AI is strongest at helping a safety professional organize and understand information, and weakest when it is asked to act on its own in the moment.
What is still out of reach
Two outcomes everyone would like are not showing up yet. Lower serious injury and fatality (SIF) rates and lower costs remain largely unrealized, even at fully integrated organizations. The report suggests cost may be a timing issue, with upfront spending ahead of returns. SIF reduction is harder. Respondents are far more confident in AI's ability to sort information and spot patterns than in its ability to identify or control SIF exposures.
That is a useful reality check for anyone being sold "predictive safety." The evidence so far says AI makes the safety professional faster and better informed. It does not replace the judgment that prevents the serious ones.
62% of organizations call data quality a barrier, and it stays high at every stage, including the most advanced. We wrote a separate piece on why AI readiness starts with your field notes.
Training lags use
About 2 in 5 safety professionals have had no formal AI training at all. Those who have report more knowledge, more comfort reading AI-generated metrics, and more belief that AI improves safety decisions. Most of the training on offer is general prompting, though. Even at fully integrated organizations, only 32% have had AI training specific to EHS. The topics people asked for most were time savings, incident analysis, and validating AI output, which half of respondents wanted.
Ethics: experience helps, but doesn't settle it
Concern drops as people get more experience, from about half of non-adopters expressing a great deal of concern to about 3 in 10 at integrated organizations. But the concern shifts rather than vanishes. It centers on AI outputs that drive decisions about workers: interventions, risk-cause assignments, and discipline. Governance hasn't caught up either. Nearly half of organizations have no AI ethics policy. We cover that in where AI should stop in a safety program.
What we'd take from it
- Start where the evidence is strongest. Report writing and training are low risk, quick to show value, and the people using AI for them rate it highly.
- Keep a qualified person in the loop. The profession is telling you where its trust ends. Build the workflow so a person reviews before anything reaches a client, a worker, or a disciplinary file.
- Fix the data at the source. Better models won't rescue vague observations and missing closeouts.
- Smaller firms shouldn't wait for perfect. The gap between large and small organizations is mostly experience, and experience is what builds confidence.
Where Safely Track fits
Safely Track sits in the part of this survey with the strongest results. It turns voice notes, photos and text from the field into the report your client requires, and you review and confirm it before it goes out. The course builder turns the safety decks you already teach into narrated courses that stay a draft until you publish. If you want a quick read on where your own firm sits, the AI readiness quiz takes a few minutes.