AI assists. It does not govern.

The governed system does not depend on an LLM being available

Catalyst workflows, permissions, calculations, evidence lineage, tenant isolation, approvals, and reporting remain authoritative whether AI is available or not.

AI sits above that foundation as a bounded assistance layer. It can help extract, explain, navigate, and draft, but the governed workflow remains the source of operational truth.

• Governed workflow: the source workflow owns state, authority, and evidence
• Scoped evidence: AI receives only the bounded context and task it needs
• AI assistance: assistance is limited to the approved task
• Human review: a person evaluates the structured result
• Accepted structured result: governed records and workflow rules remain authoritative

Reduce tedious work while leaving authority intact

These are places where AI can accelerate extraction, navigation, summarization, and pattern discovery without becoming the source of truth.

• Extracting SDS sections into a reviewable draft
• Transcribing field interviews and toolbox talks
• Summarizing known incident facts
• Identifying repeated themes
• Translating approved content
• Drafting plain-language explanations
• Suggesting ergonomic screening bands from reviewed evidence
• Surfacing possible evidence gaps
• Helping users navigate large controlled-document sets

Assistance accelerates work; it does not create operational truth.

Now with AI is not a control strategy

Catalyst does not add AI merely so a product card can say it. Model availability does not create professional authority.

• Determine regulatory compliance
• Certify PPE suitability
• Approve engineering controls
• Authorize workers
• Close corrective actions
• Classify incidents as final authority
• Certify readiness
• Alter tenant scope
• Replace a competent reviewer
• Become required for core workflows

A PPE assessment does not become more useful because a language model has been invited to admire it

The task already has known hazard categories, a governed control hierarchy, defined PPE requirements, site-specific conditions, human observation, and regulatory and procedural authority.

AI may help extract information or suggest questions. It should not become the mystical authority announcing that gloves are required because it noticed hands exist.

PPE selection should be grounded in the hazard, control hierarchy, task conditions, governing procedures, and responsible human review.

A prompt and confident paragraph are not operational truth.

AI governed as infrastructure

The sophistication is not a glowing robot icon. It is a server-side execution envelope with explicit scope, policy, and review.

• Explicit AI identities
• Tenant-scoped execution
• Action-specific grants
• Model-family policy
• Tenant allowance controls
• Provider credential isolation
• Task registry
• Human review requirements
• Run logging
• Guardrail events
• No frontend provider access
• No AI access through basic-auth or generic service credentials

AI is governed as infrastructure, not treated as a source of authority.

Commercial visibility without employee dossiers

Usage governance exists to manage allowance, guardrails, and repeated task patterns, not to score worker productivity or build advertising profiles.

• Included monthly tenant allowance
• Stepped warnings
• Administrator-authorized overflow credits
• No surprise metered billing
• BYOK monitoring off by default
• Telemetry focused on task families, guardrail effectiveness, and repeated work patterns
• No employee-productivity surveillance
• No advertising or cross-tenant profiling

AI should reveal where Catalyst needs a better native workflow

Repeated AI-assisted work is a signal to build a clearer deterministic capability. AI is useful as a scout for product development, not as permanent rented intelligence for work that can be made clear and repeatable.

• Overdue-obligation summaries: a governed reporting observation
• Corrective-action drafting: structured verification prompts
• SDS extraction: a governed SDS ingestion pipeline
• Training-gap questions: a readiness view

Core Catalyst workflows continue when AI is unavailable

AI degradation may remove assistance. It must not remove workflow authority, customer access to governed records, or the ability to complete essential safety work.

A product architecture should not collapse because the model provider has a bad afternoon.