SafetyPro
  1. Home
  2. Blog
  3. AI in incident management: from first report to closed actions
Blog

AI in incident management: from first report to closed actions

Most incident systems are digital forms. SafetyPro adds AI at the points where investigators lose the most time: classifying the report, starting the root cause analysis and turning a long file into a clear story. People still decide, sign and approve.

SafetyPro Team · · 7 min read

Where incident investigations lose time

The method for a good investigation is well known. We cover it in our step-by-step guide to incident investigation and root cause analysis. The hard part is doing it on every incident, with busy supervisors, on a deadline. Reports arrive with the wrong type or severity. Root cause sections are left empty or filled with "human error". Investigation files grow long, and managers struggle to see what actually happened. Corrective actions are agreed in a meeting and never reach a tracker.

This article explains how SafetyPro handles that work in the platform, which parts use AI, and which parts always stay with a person.

The incident lifecycle in SafetyPro

Every incident moves through five statuses, shown as a workflow bar on the incident page:

  1. Draft: the report is being written. Drafts are saved so a reporter can resume later.
  2. Submitted: the reporter has signed and submitted the report.
  3. Under Review: an approver has accepted the report for review.
  4. Approved: the approver has approved the investigation with a drawn signature.
  5. Closed: the approver closes the incident, again with a signature, when the actions are verified.

The report itself is built in a six-step wizard: classification and location, description and response, injured person and witnesses, injury locations, analysis methods, and corrective actions with the reporter signature. The AI features sit inside these steps, so nobody has to leave the form to use them.

AI classification of a new report

In the description step, the reporter can press AI Classify once the description has at least 10 characters. The AI reads the description and the activity at the time, and returns a suggestion card with:

  • One to three incident types from the platform list: Injury, Near Miss, Property Damage, Environmental, Vehicle, Fire or Other.
  • A severity for each type, chosen only from that type's own scale, for example First Aid to Fatality for injuries, or Minor Damage to Total Loss for property damage. The AI is instructed to be conservative and not over-classify.
  • The most likely root cause category, such as human factors, equipment, work environment, management, materials or design, with a short explanation.
  • Three to five recommended immediate actions and two or three areas to focus the investigation on.
  • A confidence percentage and a short line of reasoning.

Nothing changes in the report until the user presses Apply to Step 1. The user can also dismiss the card. When applied, the platform keeps its own rule for the primary type: an injury beside a near miss stays an injury, so the real severity is never replaced by a potential one. The same step also holds the regulatory fields, including OSHA, ADOSH and RIDDOR classifications and the authority notification status (Not Required, Pending, Notified or Overdue).

AI-assisted root cause analysis

Step 5 offers eleven root cause tools in two tiers. Basic: 5 Whys, Fishbone (Ishikawa), Brainstorming, Checklists and Pareto. Intermediate: Fault Tree, FMEA, Barrier Analysis, Change Analysis, Event and Causal Factor charting, and Root Cause Tree. A notes field records findings from external tools such as TapRooT or Apollo, if your company uses them.

Five of these tools have an AI Suggest button. Each one sends the incident type, severity, location, activity, immediate actions and description, and fills the tool's own fields:

  • 5 Whys: five linked questions and answers, each building on the previous one, ending in a one-sentence root cause. The AI is told to avoid generic answers such as "human error" or "lack of training".
  • Fishbone: contributing factors on six branches: people, process, equipment, environment, materials and management.
  • Brainstorming: a wide set of hypotheses and questions to investigate, including contractor interfaces, recent changes and fatigue.
  • Checklists: short observations on eight areas: PPE, procedures, training, equipment, environment, supervision, communication and maintenance.
  • Barrier Analysis: barriers in place, barriers that failed, and barriers needed, with the needed barriers ordered by the hierarchy of controls.

The suggestions are written into normal text fields. The platform tells the user to review and edit them before submitting, and the button stays disabled until there is a real description to work from. The AI gives investigators a structured first draft. It does not replace interviews, site evidence or the team's judgement.

AI incident reconstruction

On the incident page, users who can edit or investigate incidents see an AI Reconstruction button. It reads the whole record: description, weather and shift, injured person and body parts, witness statements, investigation team, every root cause tool that was filled in, photo names and comments, the incident timeline and the corrective actions. It then builds a full-screen reconstruction with six parts:

  • A two to three paragraph narrative of what happened, written like a safety report.
  • A timeline of four to eight events, each marked as trigger, escalation, response or resolution. Where an event matches an incident photo, the photo is shown beside it.
  • A Swiss cheese model with four or five defence layers and the holes in each, plus the path the hazard took through them.
  • Three to six contributing factors grouped as human, organisational, environmental or equipment.
  • Two to four lessons learned, each with a recommendation and a priority.
  • An anonymised training scenario with three discussion questions, ready for a toolbox talk.

The reconstruction can be printed. It is a view, not part of the signed record: it is not saved into the incident. Because the AI is allowed to infer reasonable details when the record is thin, treat it as a draft to check against the evidence, never as a finding.

Automatic AI triage and actions in the tracker

When a new incident is submitted, SafetyPro also starts the Incident Triage agent, if the submitting user has the permissions the server requires. The agent compares the incident with the company's 20 most recent incidents. For that comparison it receives only type, subtype, severity, location and date, never other incidents' narratives. It returns an assessed severity, whether the report looks under-reported, an investigation urgency, a repeat-pattern check, root cause hypotheses, a suggested investigation team and an escalation recommendation.

The result is stored as an AI agent finding and shown as a badge on the record. The agent also creates up to five open corrective actions in the Action Tracker, labelled as generated by the AI triage agent, with a default due date of seven days when it does not give one. A notification goes out, marked critical when the urgency is immediate or within 24 hours, and a separate critical escalation notice is sent when the agent recommends escalation. These actions are created without a confirmation step, so the HSE team should review, reassign or close them like any other action.

The investigator's own corrective and preventive actions come from step 6 of the wizard, each with a type, owner, priority and due date. They are saved as Action Tracker rows linked to the incident. If one has not reached the tracker yet, the incident page flags it as "Not in Action Tracker" instead of hiding it.

Approvals, signatures and notifications

Review and approval are controlled by a dedicated approve permission. The assigned approver accepts a submitted report for review, then either approves it or returns it for revision with a written reason. Approving and closing both require a drawn signature, with optional comments. Every step is written to an approval path on the incident with the user, time and comment, plus the signature for approval and closure.

Notifications follow the status. The assigned approver is notified when a report moves to Under Review. The reporter is notified when a report is returned for revision, approved or closed. Witness statements are protected on the server: users without investigation or approval rights see the witness list without statements, signatures, statement files or phone numbers.

What is AI and what is human-confirmed

  • AI classification: a suggestion. Applied only when the user presses Apply, and editable afterwards.
  • AI Suggest in root cause tools: a draft written into the fields. The user reviews and submits it.
  • AI reconstruction: a printable view, not saved into the record.
  • AI triage: runs automatically on new incidents, stores a finding and creates open actions that people must review.
  • Status changes, approvals, returns for revision and closure: always a person with the approve permission, with a signature for approval and closure.

A practical rollout plan

  1. Agree who holds the approve permission on each project, and assign approvers on reports.
  2. Train supervisors to write a clear description first. Every AI feature depends on it, and the AI buttons need at least 10 characters.
  3. Ask investigators to use AI Suggest as a starting point, then edit the text with what they found on site.
  4. Review AI triage actions every day during the first weeks. Close duplicates and assign real owners and dates.
  5. Use the reconstruction and its anonymised training scenario in toolbox talks to share lessons across projects.

HSE insights and practical guides

See it on your own projects

Tell us about your sites and we will show you how SafetyPro fits your HSE operations.