Most organizations that employ safety professionals are now experimenting with artificial intelligence in their safety programs, but the money, the data and the internal ownership have not caught up. That is the picture from a survey report the American Society of Safety Professionals released Oct. 14, and the gap between enthusiasm and infrastructure is the real story.
Of the nearly 1,000 professionals surveyed between July 6 and Aug. 2, 86% placed their organizations in an exploratory or transformative phase of AI adoption. ASSP defines the first as discovering and piloting AI, the second as scaling and integrating it. Two-thirds of that group started their AI work in the last two years. For workers, the practical question is what happens when a system that has not been paid for, fed clean data or given a clear owner starts shaping safety decisions.
Workplace safety has always run on records: incident logs, inspection notes, training files. AI enters that world as a tool for reading those records faster than any human can. According to Occupational Health & Safety, respondents reported using AI to develop reports, tailor education and training, analyze incidents and spot patterns, optimize and manage resources, run predictive analytics, automate compliance tasks and monitor conditions with real-time alerts.
What is actually being adopted, and how fast?
The adoption is broad but shallow. Nearly every respondent's organization sits in an exploratory or transformative phase, and most of that group arrived recently — two-thirds began within the past two years. That timing matters. These are programs young enough that their track records are thin, and the survey itself reports barriers at every stage rather than successes at scale.
The uses cluster around paperwork and pattern-finding. Drafting reports, automating compliance tasks and tailoring training are administrative applications. Analyzing incidents, spotting patterns and running predictive analytics are analytical ones. The second category is where the stakes rise, because a prediction about where the next injury happens is only as good as the incident data behind it.
Why the barriers look like a maturity curve, not a rejection
The survey's barrier list reads less like resistance to AI and more like the ordinary costs of building a capability. Respondents cited:
- Limited or early-stage investment — 74%
- Data quality — 62%
- No clear owner, champion or organizational priority — 59%
- Mistrust or discomfort with the tools — 55%
- Concerns about data privacy, security or bias — 55%
Read together, the top three are resourcing problems: not enough budget, not enough clean data, not enough executive sponsorship. Only the bottom two — mistrust and privacy or bias concerns — are questions about the technology itself, and even those are tied at 55%.
The report found that workplace culture, views on AI's importance to environmental, health and safety work, fewer ethical concerns and data quality strongly predict adoption. In other words, organizations adopt AI when leadership treats it as a priority and the underlying data can support it. That is a finding about management, not software.
Why data quality is the barrier that never goes away
Data quality was the one problem the report found at every stage of adoption. That detail deserves more attention than it will likely get, because safety data is famously imperfect. How an organization counts and classifies incidents shapes what any analysis of those incidents can find — a problem that predates AI and now gets magnified by it.
For anyone following how injury data is collected and what it can and cannot show, our explainer on how OSHA counts injuries at work covers the recordkeeping rules that feed these systems. An AI model trained on inconsistent records does not clean them up. It automates their inconsistencies.
What this means for workers and their representatives
ASSP President Monique Parker, CSP, said human oversight remains essential as safety professionals apply AI, and that the technology does not alter their basic responsibility for people, property and environmental protection. That framing puts the accountability question squarely on the organization, not the tool.
For unions and worker representatives, two items on the barrier list are worth watching. Concerns about data privacy, security or bias — cited by 55% — overlap with the questions raised by workplace monitoring, which our explainer on workplace surveillance and the law examines. And the predictive uses — spotting patterns, forecasting incidents — connect to the broader rise of algorithmic decision-making at work, covered in our analysis of algorithmic management. A system that predicts where injuries will happen can also, in the wrong hands, become a system that scores workers.
ASSP's own response to the trust problem is training. The survey found training was linked to greater knowledge of and confidence in AI for EHS work, and the organization offers an Applied AI for EHS Certificate with the University of Alabama at Birmingham along with an AI in EHS: A Practical Framework for Responsible Adoption course.
What happens next
The findings will inform ASSP's free Standards-Based User Group on AI in Safety, which is identifying further barriers and testing solutions. The ASSP Foundation funded the survey; ASSP is based in Park Ridge, Illinois.
ASSP said organizations that have not yet started using AI risk falling behind their industry peers. That is the association's stated view, and it is worth reading carefully: it is a claim from a professional body that also sells AI training. The survey data itself supports a more measured reading. Most organizations are piloting, investment lags at 74% citing limited funding, and the hard problems — data quality, ownership, trust — are organizational ones that a pilot project cannot purchase its way past. The next signal to watch is whether the transformative-phase group can scale without the data-quality problem following it there, as it has followed every stage so far.
