Most EHS teams say their data holds back their AI, and the problem doesn't fade as they get more advanced. The fix isn't a data project at headquarters. It's the observation written on site: a consistent category, an exact location, a photo, an owner, a due date, and proof it was closed. Get those right and every tool you add later has something to work with.
In ASSP's 2026 survey of nearly 1,000 safety professionals, Trust, Ethics and Governance, data quality came up again and again. It was one of three factors that predicted how far an organization had come with AI, alongside culture and ethical concerns. It was also the barrier that refused to go away. (Our overview of the survey covers the rest.)
The barrier at every stage
62% of organizations named data quality as a barrier to AI in EHS. Broken out by stage, it barely moves: 54% of non-adopters, 58% in discovery, 67% piloting, 66% scaling, and 62% even at fully integrated organizations. Teams that are piloting run into it hardest, which makes sense. That's when you first point a tool at your records and see what comes back.
Ratings of data quality do improve with experience. Only 33% of organizations in discovery call their data good or excellent, against 56% at scaling and 70% at integrated. And organizations that did predictive analytics before they adopted AI report better data. ASSP's reading, and ours: AI readiness starts well before you buy an AI tool.
What EHS teams actually collect
- Incident data is the norm. About 85% collect it and are comfortable using it.
- Equipment and environmental data is collected by about half, though nearly three quarters would be comfortable doing so.
- Ergonomic data is collected less, and people are less comfortable with it.
- Biometric data is collected by only 8%, and most aren't comfortable collecting it.
Notice what's missing from that list: the day-to-day observations and corrective actions that make up most of a safety program's paper trail. Incidents are lagging and rare. The leading signal lives in what inspectors and consultants write down every week, and that is where data quality usually falls apart.
Why field data goes bad
Not because people are careless. It goes bad for structural reasons:
- Every inspector writes it differently. One says "fall hazard," another says "unprotected edge," a third says "guardrail missing." A person reads those as the same thing. A trend report doesn't.
- It's written up later. Notes typed in the truck at the end of the day lose the location, the crew, and half the detail.
- The photo lives somewhere else. On a phone, in a text thread, in a folder nobody linked to the record.
- Nobody records the close. The hazard was fixed, but the record still says open, or says "closed" with no evidence. We call this the closeout gap.
- Each client gets a different template. Consultants reformat the same observations for every client, and the underlying data never lines up across the portfolio.
What usable safety data looks like
You don't need a data warehouse. Each observation needs the same handful of fields, filled the same way, every time:
- A category from a fixed list, mapped to the OSHA standard where one applies, so the same hazard is counted as the same thing.
- An exact location: site, area, level, side.
- A photo attached to the record, not stored separately.
- An owner and a due date, so it is a task and not a remark. (More in how to write an observation someone will fix.)
- The client, project and contractor it belongs to, so you can filter across a portfolio.
- Proof of the close: who fixed it, when, and an after photo. (See what counts as proof of fix.)
Capture it on site, not later
The single biggest improvement is when the data is captured. Something said into a phone while standing at the hazard holds more than something typed from memory hours later. Voice and photo capture are also where AI helps most safely today: turning a spoken note into structured fields is drafting, and a person still checks it. That lines up with the survey, where report development was among the uses people rated most effective.
Pull ten closed observations from last quarter at random. How many have a photo, a named owner, a due date, and an after photo? That number is your data quality, and it's a better predictor of AI readiness than any tool you could buy.
Where Safely Track fits
Safely Track captures observations as voice notes, photos and text on site, and turns them into structured fields: location, hazard type, crew and corrective actions. You review them before they go anywhere. Each record is formatted into the template your client requires, but the fields underneath stay consistent, so you can filter by client, project, contractor, issue type or OSHA category across the whole portfolio. Corrective actions go to the person responsible and come back with an after photo, a name and a time.