AI in EHS: 8 Decisions Executives Must Own

AI is entering EHS at the same time OSHA recordkeeping, ISO 45001 accountability and board scrutiny still place responsibility on human leaders. This article defines 8 executive decisions that decide whether AI in EHS becomes an early-warning system or another dashboard that looks impressive while field risk stays untouched.

Why AI in EHS is a leadership test before it is a technology project

AI in EHS is any use of machine learning, language models, computer vision or predictive analytics to support occupational safety decisions, incident review, observations, audits, training, inspections or risk prioritization. The European Commission states that the EU AI Act entered into force on August 1, 2024, and ISO/IEC 42001:2023 now gives organizations a management-system reference for artificial intelligence, which means boards can no longer treat AI as a side experiment owned only by IT.

The Headline Podcast exists for real conversations with constantly learning people, and this topic belongs in that spirit because AI exposes the gap between declared safety ambition and real leadership discipline. Andreza Araujo and Dr. Megan Tranter often bring executive safety back to the same question: who owns the decision when the system produces information that people would rather ignore?

The main trap is not that AI will replace EHS judgment. The sharper risk is that executives will buy a tool that processes weak data, produces elegant rankings and then gives leaders a new excuse to delay hard decisions about staffing, production pressure, contractor control and escalation.

1. Decide what problem AI is allowed to solve

AI in EHS should begin with one defined safety decision, not with a broad promise to transform the department. A useful first scope might be classifying high-potential near misses, finding repeated audit findings, flagging overdue corrective actions or detecting patterns in observation quality, while a vague scope such as improving safety culture usually becomes impossible to test.

The executive angle matters because many AI pilots fail before the first model runs. They fail when leaders skip the question of what decision will change if the output is different from the current dashboard. This is why safety decision rights should be clarified before a vendor demo, since an algorithm cannot compensate for a leadership team that has not assigned authority.

For a senior EHS leader, the practical test is simple enough to write on one page. State the decision, the current evidence, the expected new evidence, the owner, the escalation path and the threshold for action. If those 6 fields cannot be completed, the organization is not ready for AI in that area.

2. Decide who owns model output when it conflicts with production pressure

An AI system can identify a high-risk pattern at 9 a.m., while the plant still has a production commitment at 2 p.m. The value of the signal depends on whether the operations leader, EHS leader and site director have already agreed what happens when the model challenges the production plan.

This is where many executive teams reveal their real safety culture. Across 25+ years leading EHS in multinationals, Andreza Araujo has seen that risk information only becomes protection when it changes authority, schedule or resources. A dashboard without decision authority only records tension.

Executives should require an escalation rule for every high-risk AI output. That rule should say who can stop work, who validates the data, who communicates the decision and how long the organization can keep running while the risk is unresolved. Without that rule, AI becomes another warning that everyone can admire and nobody has to obey.

3. Decide which data is clean enough for life-critical use

AI models learn from records, and safety records are often distorted by underreporting, inconsistent severity classification, delayed closeout and incentive programs that make bad news politically expensive. In many organizations, 3 familiar data sources, incident reports, audits and observations, carry enough bias to mislead a model if leaders do not test them first.

What most software proposals understate is that EHS data is not neutral. If one plant reports 200 near misses and another reports 12, the safer plant may be the one with the larger number, because it has a reporting climate that exposes weak signals instead of burying them. AI that reads volume without cultural context can punish transparency.

The first executive control is a data-readiness review that compares reporting rates, classification quality, corrective-action aging and field verification. Link the AI project to the existing safety KPI weighting problem, because the same weak metrics that mislead executives can train the model to rank the wrong risks.

4. Decide what a human must verify before action

Human review is not a decorative approval step after AI has already framed the conclusion. In safety, human verification means checking whether the source record, field condition, task design and operational context support the recommendation before the organization changes work.

As Andreza Araujo argues in *Safety Culture: From Theory to Practice*, culture is visible in what leaders reinforce, tolerate and inspect. AI can surface a pattern, but the leader still has to go where the risk lives, ask whether the output matches the work and decide whether the control is present in the field.

The practical rule is to divide AI outputs into 3 classes. Low-risk administrative outputs can be sampled, medium-risk recommendations need review by EHS and operations, and high-risk outputs that affect work authorization, discipline or shutdown require named executive accountability. This keeps AI from becoming an invisible decision-maker in life-critical work.

5. Decide how privacy and psychological safety will be protected

AI in EHS often touches worker observations, camera feeds, wearable data, investigation notes, health indicators or behavioral records. Those data sets can support prevention, although they can also create surveillance fear if leaders introduce them without boundaries.

On Headline Podcast, conversations about visible felt leadership keep returning to trust because people will not report weak signals if they believe the organization is building a punishment machine. Dr. Megan Tranter and Andreza Araujo frame leadership as a discipline of clarity, and AI deployment needs that clarity before the first pilot begins.

The executive decision should define what data is collected, what data is excluded, who can see individual-level information, how long records are kept and which uses are forbidden. This must connect to psychological safety audits, since a tool that silences voice can make the organization less safe even when the analytics look advanced.

6. Decide how AI will handle weak signals, not only injuries

AI becomes more useful when it reads weak signals before an injury occurs, because traditional lagging indicators tell leaders what already happened. Near misses, repeated deviations, maintenance deferrals, overtime spikes, audit repetition and quality complaints may reveal the drift that a TRIR chart hides.

The board-level implication is direct. If AI is used only to write faster incident summaries, the organization has automated the past. If it is used to connect weak signals across functions, it can support safety as material risk, where operational signals reach people who control capital, staffing and production promises.

Executives should require at least 5 weak-signal families in the pilot design: exposure, control quality, reporting quality, corrective-action age and operational pressure. The model should be judged by whether it changes an upstream decision, not by whether it produces a prettier chart.

7. Decide how vendors will prove field validity

A vendor can show accuracy on a demo set and still fail in a noisy plant, mine, warehouse or construction site. Field validity means the system works with the organization's real language, incomplete records, lighting conditions, reporting habits, job names, contractor mix and escalation behavior.

This is the difference between a technology purchase and a safety control. In more than 250 cultural transformation projects supported by Andreza Araujo, one repeated lesson is that safety systems fail when the presentation is stronger than the field test. AI should not escape that standard simply because the interface looks more modern.

The procurement test should include a 30 to 60 day pilot, false-positive review, false-negative review, frontline input, data-quality exceptions and a documented kill criterion. If the tool cannot show where it is wrong, leaders should not trust it where people can be hurt.

8. Decide what will be stopped if AI reveals intolerable risk

The final executive decision is the one many organizations avoid. Before AI goes live, leaders must define what work, plan, contract, target or practice can be stopped when the system reveals intolerable risk.

Co-host Andreza Araujo's work in *Antifragile Leadership* describes how leaders mature when pressure becomes a trigger for better decisions instead of denial. AI can support that maturity only if the organization is willing to act on uncomfortable evidence, especially when that evidence threatens a production promise or exposes an expensive design flaw.

The practical output is an action protocol with thresholds. A red pattern in high-energy maintenance might trigger a same-day control review, while repeated contractor deviations might freeze mobilization until supervision and permit controls are corrected. AI earns its place in EHS when it changes the work before the serious event, not after the investigation.

Comparison: AI as safety control versus AI as dashboard theater

Each quarter spent piloting AI without ownership rules gives the organization more data without more control, while production pressure continues to shape field decisions faster than the safety dashboard can react.

Conclusion

AI in EHS is valuable only when executives decide ownership, evidence, privacy, validation and escalation before the tool starts ranking risk.

For the Headline Podcast community, the better question is not whether safety leaders should use AI, but whether they are prepared to act when AI reveals what the organization has normalized. Listen to the conversations at Headline Podcast and bring this 8-decision test to your next executive safety review.