How Cando Rail Cut Risk-Assessment Time 83% With AI
83%
Not disclosed
Not disclosed
Cando wanted to scale its risk assessment program sixfold but faced heavy administrative burden. The existing process required 4-16 hours of meetings per assessment for up to 10 subject-matter experts, plus 2+ weeks of back-and-forth coordination that pulled experts away from safety-critical work.
Fractional AI built Peter the Safety Agent, an AI assistant with two workflows: automated draft generation for pre-meeting preparation, and a real-time in-meeting voice assistant that documents and updates assessments live. It draws on historical risk assessments, safety procedures, incident reports, and training data.
The team used GPT-4.1 for drafting and the OpenAI Realtime API for voice, orchestrated through LangChain agent flows with RAG, integrated with MS Teams and Recall.ai for meeting transcription. A key insight was that human-written assessment titles outperformed similarity-based RAG retrieval for selecting relevant context.
Best fit for safety- or compliance-heavy operators who run repetitive, expert-intensive assessment meetings and need to scale them without adding headcount.
20+ hours saved per assessment
83% faster to scale 6x
Under $0.05 per draft
Chris Taylor
Through an AI safety agent called Peter that drafts risk assessments before meetings and documents them live during them. Grounded in Cando Rail's historical assessments, safety procedures, and incident reports, it removed the manual bottleneck and cut the time to scale the program sixfold by 83%.
The agent used GPT-4.1 for drafting and the OpenAI Realtime API for voice, orchestrated through LangChain agent flows with retrieval-augmented generation (RAG). It integrated with Microsoft Teams and used Recall.ai for meeting transcription.
Three main outcomes: over 20 hours of subject-matter-expert time saved per assessment, an 83% reduction in the time needed to scale the program sixfold, and automated draft assessments generated for under $0.05 each.
A specific implementation timeline was not disclosed. Impact was measured per assessment once the agent was live, with each one saving over 20 hours of expert time.
It is best suited to safety- or compliance-heavy operators that run repetitive, expert-intensive assessment meetings and need to scale them without adding headcount.
Have a similar challenge?
