The problem: the signals are there, but scattered
Before most serious incidents, the warning signs were recorded somewhere. Three trip hazards at the same stair in one week. Observations about cable routing that were closed with the word "done". Two hot work permits active in the same area. Each record alone looks minor. Together they tell a story, but only if someone connects them.
Dashboards show totals. They rarely show that a specific category in a specific location is rising faster than its own recent history, or that the same problem keeps coming back after closure.
What a useful prediction engine does
- It works on your own records, project by project.
- It separates hard facts from AI interpretation.
- It shows the records behind every signal, so a person can check them.
- It states its limits openly.
- It turns signals into owned, tracked work.
Layer 1: rule-based signals you can check by hand
The Analysis / AI Intelligence tab of the Action Tracker reads the project's hazards, actions, observations and NCRs, filtered by what the user is allowed to see. Risk is grouped on the 25-point scale: 17 and above is Critical, 10 and above High, 5 and above Medium. The engine then calculates findings with fixed rules:
- Hotspot: the location with the most High or Critical records.
- Pareto focus: how many categories make up 80 percent of all records.
- Three-week increase: records created rose in three consecutive weeks, with the category that drove the latest week.
- Overdue profile: overdue items by category, owner, contractor and location, with the oldest overdue item and the average days late.
- Weak closures: closed records whose closing note is missing, shorter than 10 words, or only a word like "done" or "fixed".
- Contractor detail: open, closed and overdue items and average closure days per contractor.
Each finding links to the records behind it, and any number can be drilled down to the list of records.
Emerging Risk: a defined early warning
The Emerging Risk signal looks at every category and location pair in the project. It raises a flag when all of these are true:
last 7 days >= 3 records AND at least 1 High/Critical AND >= 2 x average week of the previous 28 days
Up to 8 such pairs are shown, each with its records. Alongside it, a historical view names categories that stayed in the top 3 in each of the last four months, when every month has at least 5 records. The screen states that Emerging Risk is a review signal, not a forecast. That honesty matters: it tells the reader exactly why something was flagged.
Layer 2: AI that must show its evidence
On top of the rules, AI reads the narratives, the free text that people write in each record. Here the engine is strict about evidence:
- Detect recurring issues: the AI labels each record with a specific issue, such as cable routing and trip hazard, and must quote a short passage from that record. If the quote is not found in the record text, the label is dropped. Groups need at least 2 records, and the engine flags possible repeats at the same location after an earlier record was closed.
- Ask AI: you can ask a question of up to 500 characters about the scope. Every claim in the answer must cite exact source records. Claims with missing or unknown references are withheld, and the answer says so.
- Deep review: insights are labelled as a connection, priority, possible factor or evidence gap. A connection between two modules is only shown if it cites records from both. Possible causes are labelled as possible, and recommendations require human approval.
Teams can add evidence-backed corrections and lessons to a project memory. That guidance is used in later reviews only after someone with edit rights on every referenced module approves it, and current records remain the source of truth.
Layer 3: AI forecasts and cross-module patterns
For a broader view, the dashboard has two on-demand AI cards. AI Pattern Detection sends a project summary (incidents by type, location and weekday, hazards by category and location, inspection scores, overdue actions, closure rate and training) and returns patterns, correlations across modules, hotspots and actionable insights. AI Predictive Analytics adds six months of trends and returns a 30-day risk forecast with a confidence level, a four-week outlook, a risk heatmap by category, scenarios and positive and negative leading indicators.
On the AI Agents page, the Risk Monitor compares the last 30 days with the previous 30 days across incidents, hazards, observations, inspections, open actions and permits, and looks for drift, hotspots and emerging risks. The Trend Predictor looks back 90 days. These outputs are AI estimates, and the instructions require them to come from your actual data and to admit when data is too thin.
Proactive warnings where people work
The AI Copilot bar adds warnings inside each module, built on simple checks first. Incident locations with 2 or more incidents are passed as repeating locations. Observation locations with 3 or more reports become hotspots. Two or more active permits in the same zone are flagged as possible simultaneous operations. Inspection scores are averaged by location, and workers on active permits with expiring certificates are listed. The AI then writes up to 4 short warnings or actions for that page.
From prediction to action
- When the Risk Monitor rates an emerging risk as High probability, it creates an open hazard record of type Predictive, marked as AI-generated, with the early warning signs and suggested mitigation.
- The Incident Triage agent creates up to 5 corrective actions for a new incident in the Action Tracker.
- High-severity drift, critical trends and high-probability risks send notifications that link back to the source.
- From the analysis tab, teams can compare against another period or project and generate scoped reports, such as High-Risk Analysis or Overdue Action Analysis.
Every created hazard or action is an ordinary record. It gets an owner, a due date and the normal closure and verification steps.
Limits you should know
- Contractor comparisons are not adjusted for workforce size or man-hours. The screen says this directly.
- Rule-based signals depend on consistent categories and locations. Records without them cannot join a hotspot.
- Forecasts and patterns are AI estimates, not certainties, and get better with more complete data.
- The Risk Monitor, Trend Predictor, pattern detection and forecast run when a user starts them. Only permit review and incident triage start on their own.
- Long narratives are shortened for AI review, and the answer notes when coverage was partial.
A practical rollout
- Agree a fixed list of locations and categories for each project, and use it in every record.
- Ask for closing notes of at least a full sentence, so weak closures fall.
- Open the Analysis tab in each weekly HSE meeting and review Emerging Risk first.
- Run Detect recurring issues monthly and assign an owner to each repeat group.
- Run the Risk Monitor before monthly reviews, and check every AI-created hazard.
- Record corrections when the AI is wrong, and approve the useful ones.