Digital Transformation: How to Automate Investigation Report Audits with an AI Assistant

Case
17 November 2026 🇷🇺 Original language: русский

The Flip Side of Transparency: Why Manual Review of Investigations Hits a Dead End

As safety culture matures, industrial companies inevitably move beyond tracking only severe workplace accidents and begin recording minor injuries, hazardous situations, and contractor incidents. While transparency provides an objective risk picture, it exponentially increases the workload on the health and safety team: the influx of documentation exceeds expert capacity, causing analytical quality to drop. Using the wood-based panel division of the timber group Sveza as a case study, Health and safety manager Arina Kuzyukova explains how to automate the initial audit of incident investigations without sacrificing analytical depth.

Birch plywood manufacturing combines historic early 20th-century mill layouts with modern machinery, hydrothermal log treatment vats, hot presses, cutting tools, and heavy forklift traffic. In 2015, the company's overall statistics covered only serious incidents—around 30 to 40 cases per year, which could be thoroughly moderated by hand. Expanding the reporting scope to include minor injuries and near misses in 2018, followed by all contractor personnel incidents in 2020, turned dozens of reports into hundreds and exposed three systemic barriers:

  • Superficial root cause analysis on the shop floor. Frontline supervisors lack the skills of professional analysts; without methodological guidance, they reduce conclusions to phrases like "human error," "worker inattention," or "that is how it has always been done," leaving the true root cause unidentified.
  • Heavy workload for safety specialists. Returning rough reports for revision, holding clarification calls, and conducting multiple rounds of proofreading consumed at least 20% of OH&S experts' working hours, crowding out hands-on shop-floor activities such as risk assessments, behavioral safety audits, and direct engagement with workers.
  • Box-ticking corrective actions. Vague causes generate ineffective action plans filled with wording like "strengthen oversight," "deepen focus," or "conduct a briefing" paired with nominal deadlines such as December 31, failing to eliminate the hazard source or prevent recurrence.

AI Assistant Architecture: Corporate Chatbot and Multi-Stage Verification

To relieve methodology specialists of routine reviews and assist area supervisors at the drafting stage, the company translated its corporate investigation standard, the 5 Whys methodology, and SMART criteria into a system of prompts and rules for a large language model. Deployed within a secure internal perimeter to prevent leaks of personal or sensitive data, the solution is accessed via a familiar corporate chatbot on a smartphone or desktop computer. Internal regulations and operating procedures are integrated directly into the assistant's knowledge base, so supervisors do not need to look for a separate IT tool or manually cross-check standards.

Working through a report in the chatbot is structured as a sequential analytical filter:

  • 5 Whys analysis. Once the event description is entered, the assistant checks the logical coherence of the causal chain proposed by the supervisor; if a manager struggles to start the analysis from scratch, it helps formulate hypotheses and build an initial line of reasoning.
  • Verification through an eight-criteria logic filter. Ten years of company practice showed that the classic 5 Whys method often falls short without additional guardrails, so the bot tests conclusions for clarity of wording, sound cause-and-effect relationships, sufficient evidence, and consideration of alternative factors.
  • Audit and development of corrective actions. The model cross-checks the proposed action plan against the identified root cause, automatically blocks generic prescriptions such as repeat briefings or stricter supervision, and recommends specific engineering and administrative controls before exporting the final report.

Implementation Impact, AI Limitations, and Further Digitalization

Automating the initial audit cut the average moderation time per report by 82%—from several days to just a few minutes. Rapid feedback accelerated final investigation sign-off and shifted teams' focus from paperwork back-and-forth to implementing engineering and systemic barriers against repeat incidents. At the same time, the presentation clearly defines the boundaries of the technology: the neural network serves as a logic coach and editor rather than a final approval authority, and it cannot independently detect the concealment of facts by department managers with a conflict of interest.

At Sveza, data reliability is addressed through organizational mechanisms and safety culture development: investigation committees include dedicated OH&S methodologists, corporate-level specialists, and independent experts from other group sites, while recurring events are investigated individually every time and systemic failures are consolidated into targeted site- or division-level projects. On the AI development front, the team is expanding the bot's capabilities to generate witness interview guides, automatically produce safety posters and toolbox talks in clear, worker-friendly language, build risk assessment assistants, and pilot conversational simulators for practicing leadership safety walks, safety briefings, and interview-based safety culture assessments.

Key Takeaways from This Presentation

  • How to translate a corporate investigation standard and the 5 Whys methodology into system prompts—including when using public language models without compromising confidential data?
  • Which logical criteria should be used to verify cause-and-effect links and filter out superficial claims of "worker inattention"?
  • How to configure automated filtering of vague corrective actions and steer frontline supervisors toward engineering controls?
  • What organizational safeguards can offset bias among local managers where artificial intelligence algorithms reach their limits?
  • How to scale AI assistants to preparing safety toolbox talks, conducting risk assessments, and running conversational training simulators for managers?
For Pro and VIP members
Structured summary with budget, timelines, team, and tools.
Choose plan

600+ cases and practices

Explore the full library of industrial safety best practices

Go to library
We use cookies to improve your experience · Cookie Notice

Join the leaders

14,000+ professionals · 128+ countries

1
Contacts
2
Profile

Registration

Tell us about yourself

Required field
Required field
Enter a valid email
Invalid number

Registration

Professional details

Required field
Required field
Required field

Please consent to newsletters. This will greatly enhance your platform experience.

Registration complete

We sent login credentials to your email. Use the password from the email to sign in.

Didn't receive the email?
Check your Spam folder
Already have an account? Sign In · Forgot password?

Welcome!

You have successfully signed in.

Don't have an account? Register · Forgot password?

Password Recovery

Enter your email to recover access

Enter a valid email

Link sent

A password reset link has been sent to the specified email. The link is valid for 1 hour.

Didn't receive the email?
Check your Spam folder
Remember your password? Sign In · Register