SafetyPro
  1. Home
  2. Blog
  3. Safety Brain, AI agents and the Copilot: how AI works inside SafetyPro
Blog

Safety Brain, AI agents and the Copilot: how AI works inside SafetyPro

Most HSE teams already have the answers they need. They are buried in incident reports, permits, inspections and procedures that nobody has time to read again. The AI layer in SafetyPro reads that data for you, answers in plain language, shows where each answer came from, and leaves the decisions to qualified people.

SafetyPro Team · · 6 min read

The problem: the knowledge is there, but nobody can reach it

A safety manager preparing for a weekly meeting needs quick answers. Which high-risk hazards are still open? What did we learn from the last dropped-object incident? Which procedure covers confined space rescue on this project? The information exists, but it is spread across modules, attachments and spreadsheets.

Generic chat tools do not help much. They do not know your project, they cannot see your records, and they do not tell you where an answer came from. In HSE, an answer without a source is a risk in itself.

What good looks like

  • The assistant answers from your own records and documents, not only from general knowledge.
  • Every answer shows the records it used, so a person can open them and check.
  • Company data stays inside the company, and each user sees only what their permissions allow.
  • Routine reviews run automatically, but people approve what matters.
  • The system is honest about its limits.

Safety Brain: a chat assistant that knows your project

Safety Brain is the chat assistant of the platform. When you ask a question, it receives a live summary of the project you are working in: incident and hazard totals, open and overdue actions, the latest 10 incidents, open hazards with a risk score of 10 or more, overdue actions with their owners, active permits and the 5 most recent inspections. It refers to records by their IDs, so you can find them.

Typical questions include:

  • Analyze our open incidents and suggest priorities.
  • Review overdue actions and recommend next steps.
  • Help me investigate a near-miss incident.
  • Generate a toolbox talk on working at heights.
  • Summarize this month's HSE performance.

You can also attach a site photo (up to 10 MB). Safety Brain then returns a structured hazard assessment: identified hazards, a risk level for each, applicable standards, recommended actions and anything that must stop immediately. The chat screen carries the reminder that AI can make mistakes and safety-critical information must be verified.

Answers grounded in your records, with sources

Behind Safety Brain is a search index of your company data. When records are created or updated in 11 areas (incidents, hazards, inspections, observations, permits, audits, actions, toolbox talks, HSE meetings, QHSE documents and knowledge base entries), they are queued for indexing. A background job processes the queue every minute and retries failed items up to 5 times.

Documents get special treatment. Files uploaded to the QHSE document system are indexed with their content, not only their title. PDF text is read directly, scanned PDFs and images are read by AI vision, and Word and Excel files are unpacked to text. If a file could only be read in part, the index records that. Teams can also add their own text, such as a site rule or a lesson learned, in the Knowledge Base tab.

For each question, the platform searches the index by meaning and passes up to 8 of the most relevant records to the AI, with an instruction to cite them as [Source N] and to say honestly when the data is not enough. If the meaning-based search finds nothing, a keyword search is tried. Under the answer, a sources panel lists each record used, with its module, code, type, status, date and match percentage.

Tenant and project isolation

Isolation is enforced on the server, not only in the screen:

  • Every AI request requires a signed-in user. Suspended company accounts are blocked.
  • The search runs only inside the company the user is working in, and only after their active membership in that company is confirmed.
  • Results are limited to the current project, plus company-wide records that belong to no project.
  • If a user does not have permission to view incident details, incident records are removed from the search entirely.
  • Access to Safety Brain and to the AI Agents page is controlled by their own permissions, which an administrator can switch on or off per role.

AI agents: routine reviews that run on their own

The AI Agents page groups seven agents. Two of them start automatically:

  • Permit Processing Agent: when a permit is submitted, it reviews the application for missing fields, risk flags, extra controls, extra approvals and PPE. A High or Critical result creates a notification that links to the permit.
  • Incident Triage Agent: when a new incident is reported, it suggests a classification, assesses severity, checks for repeat patterns, sets an investigation urgency and recommends the investigation team. It creates up to 5 corrective actions in the Action Tracker, marked as AI-generated, and sends an escalation notification when it judges one is needed.

The other five run when a user clicks Run Now: the Compliance Monitor (certificates, legal requirements, statutory inspections, training, permits and overdue actions), the Risk Monitor, the Audit Auto-Scheduler, the Trend Predictor and the Report Narrative Generator. Agent findings are stored, listed in an activity feed, and shown as small badges on the related records, so people see them where they work.

The Copilot: a second pair of eyes on every page

The AI Copilot is a bar at the top of most modules. When you open a page, it reads a summary of that module and returns up to 4 suggestions, each marked as a warning, information or an action, with a priority. Each module has its own focus. On permits, it looks for active permits sharing the same zone and expired permits not closed. On inspections, it looks at average scores by location. On training, it checks workers on active permits whose certificates are expiring. You can also type a question about the module.

In text fields across the platform, an inline writing assistant can improve, correct, rewrite, shorten or expand what you wrote, with an Undo button if you prefer the original.

Safeguards and honest limits

  • AI output is advice. The system prompt tells Safety Brain to remind users that suggestions must be reviewed by qualified HSE personnel.
  • Agent-created actions are ordinary actions with an owner and a due date. If the AI gives no deadline, the default is 7 days, and the team can edit or close them like any other action.
  • The triage agent sees only basic facts of earlier incidents (type, subtype, severity, location and date), not their narratives.
  • The triage agent runs only for users who are allowed to view incident details.
  • A Safety Brain conversation lives in the open page only. Starting a new chat or reloading the page clears it.
  • Answers depend on data quality. Sparse or vague records give weaker answers, and the AI is told to say so.

Getting started

  1. Decide which roles get Safety Brain and AI Agents access, and set the permissions.
  2. Upload your key procedures and method statements to the QHSE system so they are indexed.
  3. Add short Knowledge Base entries for site rules that live only in people's heads.
  4. Ask Safety Brain the questions you answer every week, and open the cited sources to check them.
  5. Review AI-created actions in the Action Tracker during the weekly HSE meeting.
  6. Use thumbs up or down on answers, so the team can see what works.

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.