How to use AI in a safety management system without losing accountability

Safety teams are being asked to use AI. That is no longer a future conversation. The harder question is how to use it inside a safety management system without blurring who is accountable for the decision.

In high-risk operations, accountability is not a slogan. It sits with named people: a Site Senior Executive, a Coal Mine Operator, a supervisor, a document owner, a competent person who signs that a control is fit for purpose. AI can help those people work faster. It cannot take their place.

This article sets out a practical way to use AI in a safety management system so the tool assists the review, and people still own the outcome.

The risk is not AI. It is unowned AI.

Most safety professionals are not rejecting AI. They are worried about overreliance. If a model drafts a procedure, summarises an incident, or flags a gap, someone still has to decide whether that output is right for this site, this hazard, and this workforce.

Accountability is lost when:

• people paste site documents into a public chatbot and treat the answer as advice
• AI writes generic content that is not grounded in the organisation’s own SHMS, risks or incidents
• there is no record of who accepted, edited or rejected the output
• AI is allowed to publish, approve or cull controlled documents
• the system of record is the chat history, not the safety management system

In those cases the organisation has activity, not assurance.

What “human in the loop” has to mean

Human in the loop is not a person glancing at a paragraph and clicking accept. In a safety management system it means a defined review step, with a named role, before anyone relies on the output operationally.

A workable standard looks like this:

1. The record already lives in the safety management system — a procedure, risk, incident, audit finding or training document.
2. AI reviews that record against related information the organisation already holds.
3. The output is a draft: gaps, inconsistencies, missing references, questions to ask. It is not the approved record.
4. A competent person accepts, edits or rejects it.
5. The decision is stored with the record, alongside the usual version history and audit trail.

If step 4 is missing, you do not have a safety workflow. You have a writing assistant.

A practical way to use AI in a safety management system

Start with review, not generation. Generation creates more documents. Review helps people see what the current documents, risks and records are actually saying.

1. Keep AI inside the system of work.

The useful context is already in the SHMS: controlled documents, risks, critical controls, incidents, training, contractor records, audits and actions. AI-assisted review that reviews those records in place is more defensible than AI that invents content from the open internet.

2. Give AI a narrow job.

Good jobs include:

• summarise an incident record so the investigator starts closer to the point
• compare a procedure with related risks, obligations or other controlled documents
• ask whether a listed control looks like an actual control or a supporting activity
• highlight missing owners, stale review dates, broken references or inconsistent wording

Poor jobs include:

• rewrite the SHMS overnight
• decide who is allowed on site
• close an action
• declare a critical control effective

3. Send only what the review needs.

Do not dump an entire community into a model. Limit the input to the record being reviewed and the related fields required for that task. Keep customer data in the application first. If the AI service is unavailable, the underlying workflow must still run.

4. Make acceptance visible.

If a person cannot see who reviewed the output, when they reviewed it, and what they changed, auditors will treat the AI text as unowned. Store the draft against the record. Keep publish, approve and close-out with people.

5. Leave legal and statutory judgement with people.

AI can help a team prepare. It does not replace legal advice, professional judgement, leadership accountability or an organisation’s own obligations. Say that plainly, including to your own users.

What should stay with a competent person

Keep these decisions human, even when AI has done the first pass:

• whether a document is current, adequate and ready to publish
• whether a control is critical, and whether it is an object, an act or a system
• whether training content matches the procedure people are expected to follow
• whether an investigation has identified the right contributing factors
• whether a gap is a paperwork issue or a real-world risk
• whether a document should be withdrawn, merged or left in place

AI is useful when it reduces collation. It is harmful when it becomes the author of the system of work.

Connected records make the review safer

AI is only as good as the information it can see. If procedures live in one drive, risks in a spreadsheet, incidents in email and training in a separate LMS, the model will review a fragment and sound confident about a whole system.

A connected safety management system changes that. A procedure can be reviewed against the risks it is meant to control, the training that is meant to teach it, the incidents that tested it, and the actions that were meant to close the last gap. That is the difference between a chatbot and AI-assisted review.

RSURED is built as that connected record — documents, risks, incidents, audits, training, contractors and actions in one platform. AI-assisted review, where enabled, is optional and customer-controlled. It is designed to assist users, not replace operational judgement, and it is not on by default.

A simple test before you switch it on

Before AI touches a live SHMS, ask:

• Who is allowed to run a review?
• What records can it see?
• What is it not allowed to do?
• Where is the output stored?
• Who must accept it before anyone relies on it?
• What happens if we turn it off tomorrow?

If you cannot answer those questions, do not use AI in the safety management system yet. Get the data, the workflow and the ownership clear first. Then use AI to help competent people review the work they already own.

Related reading

For more on how RSURED connects compliance, safety and workforce records across the platform:

• HSEQ Documents & SHMS
• Risk & Hazard Management
• Incidents & Injuries
• Audits & Assessments
• Security & Compliance

Want to see how AI-assisted review can sit inside a connected safety management system — with people still accountable for the outcome? Book a demo with the RSURED team.