AI you can put in front of procurement
EnviroAutomate uses AI to help environmental consultants work faster on the routine parts of report production. We built it for an audience that has to defend every output to a regulator, a client, and an auditor. That shaped every decision below. We use Amazon Bedrock with Anthropic Claude. This is deliberate enterprise architecture, not a chatbot wrapper.
Where your data lives
All AI processing is performed within AWS Asia Pacific (Sydney), ap-southeast-2. This is the same Australian region where your application and database run. Your data does not leave Australia for AI processing.
Your data is never used to train AI models
Amazon Bedrock does not use customer inputs or outputs to train or improve foundation models. This is a core guarantee of the Bedrock service, on by default, with no opt-in required.
Your questions, site data, and report content never enter a training dataset. Your interactions never influence an AI model's behaviour for any other user. The reference is the Amazon Bedrock Data Privacy FAQ.
Tenant isolation
EnviroAutomate is multi-tenant. Every AI interaction is strictly scoped to your organisation's data. Tenant isolation is enforced at the application layer before any data reaches the model. Prompts sent to the AI are constructed exclusively from your tenant's data. There is no mechanism by which one client's data can appear in another client's AI session.
Guardrails on every request
All AI features are protected by Amazon Bedrock Guardrails. Guardrails are applied to every request, regardless of which feature is in use.
- High-sensitivity content filtering applies to hate, harassment, sexual content, violence, criminal or harmful behaviour, and prompt injection.
- Topic restriction limits responses to environmental consulting subject matter, so general-purpose queries are blocked.
- Prompt injection defence catches attempts to manipulate the model through user input or content in your data.
- When the guardrail blocks a request, the user sees a safe generic message and the original query is not exposed.
Encryption end to end
In transit, traffic uses HTTPS with TLS 1.2 or higher from browser to application and from application to the AI service. HSTS is enforced. At rest, AES-256 covers the database, file storage, and AI chat history.
Access control
- MFA is enforced for every user, with no bypass.
- There is no public or unauthenticated access to any AI feature.
- Role-based access restricts AI outputs to a user’s permitted scope.
- AI requests are issued from private AWS subnets, server-side only; no browser ever talks directly to the model.
Full audit trail
Every AI interaction is recorded in the EnviroAutomate database:
- Each user and assistant message is timestamped to the user identity.
- Every tool call the AI made is logged with its inputs and outcome.
- Every action the AI proposed is recorded, along with who approved or rejected it.
- Every AI output is traceable to the model invocation that produced it.
Shared responsibility
EnviroAutomate controls
- Tenant isolation and scoping of AI context to your data
- Guardrail configuration and topic restriction
- Authentication, MFA, and role-based access
- Audit logging and retention of AI sessions
- Data residency (all infrastructure in ap-southeast-2)
AWS controls
- The guarantee that your data is not used for foundation model training
- Guardrail enforcement at the infrastructure level
- The underlying model and infrastructure security
- Physical data centre security (ISO 27001, SOC 2, PCI DSS)
- Regional data containment within ap-southeast-2 (Sydney) and ap-southeast-4 (Melbourne)
The human is the decision-maker
EnviroAutomate drafts, and a qualified scientist reviews, edits, and approves. The AI Assistant cannot write to your data without an explicit human Approve click on a proposed action, and the assistant's safety design is documented in detail on the AI Safety page.
Request the full AI Risk Assessment
The summary above is the procurement-ready short version. The full AI Risk Assessment (available on request) includes the residual risk register, the detailed shared-responsibility decomposition, and the operational mitigations behind each control. For other security and architecture detail, see our Technology page, or contact us.