How Our AI Works

Transparent AI for safety-critical environments.

We use AI to handle documentation — not decisions. Here's exactly how it works, what it can and can't do, and why human oversight is built into every output.

⚠

Important: Every report generated by Shift Copilot is AI-assisted and requires human verification before sign-off. Supervisors are responsible for confirming the accuracy of all generated content. Shift Copilot is a documentation tool — not a safety authority.

Human in the loop, always

Every report Shift Copilot generates is presented to the supervisor for review before it becomes any kind of record. We explicitly flag missing information, ambiguous entries, and areas requiring clarification. The AI structures — the human signs off.

WA-calibrated outputs

Our AI is prompted with specific knowledge of the WHS Mines Regulations 2022, WorkSafe WA requirements, and DMIRS reporting obligations. It doesn't use generic safety templates or reference US OSHA standards. Every output is structured for the WA regulatory environment.

No data used for model training

Shift notes submitted through Shift Copilot are processed via the Anthropic API. Anthropic does not use API-submitted data to train their foundation models by default. Your operational data does not improve anyone else's AI.

Transparent about limitations

AI can misinterpret ambiguous speech, miss context not explicitly stated, or generate plausible-sounding but incorrect operational details. We flag gaps automatically and we're upfront: if the supervisor didn't say it, the AI shouldn't invent it.

Documentation tool, not decision maker

Shift Copilot structures what supervisors say. It does not recommend operational decisions, assess risk severity beyond flagging what was mentioned, or substitute for formal safety management processes. Those decisions belong to qualified personnel.

How a report is generated

Supervisor speaks or types shift notes

Raw, unstructured input in their own words. Voice is converted to text on-device — no audio leaves the browser.

Text sent to our serverless function

The transcribed text and site context are sent securely to our backend. API keys are never exposed to the browser.

AI processes against WA mining context

Anthropic's Claude model receives the notes along with a detailed system prompt calibrated to WHS Mines Regs 2022, WA operational standards, and structured output requirements.

Structured report returned and displayed

Seven sections are populated. Missing information is automatically flagged. The supervisor reviews the output before copying, downloading, or using it.

Supervisor verifies and signs off

The human step. Always. The supervisor is responsible for the accuracy of the final record — Shift Copilot is the drafting tool, not the authority.

Known limitations — what our AI can't do

It can't invent information. If the supervisor didn't mention an equipment status, a hazard, or a permit number — the AI will flag it as missing, not fabricate it.

It can't assess risk severity. The AI can note that a hazard was mentioned and flag it for attention. It cannot formally classify risk severity — that requires a qualified person on site.

It can't replace WHS expertise. Reports are designed to support handover obligations, not substitute for formal safety management system processes or HSE professional judgment.

It can occasionally make errors. Large language models can misinterpret ambiguous input. This is why every report requires human review before use. Do not use AI-generated content without checking it.