AI transparency notice

In short: Sustain uses AI to read documents and to help you ask questions of your own data. No AI produces, adjusts, or ranks a euro figure. Deterministic simulation code prices each recommendation you see, and you can verify that yourself: turn the language model off and each number in the system stays identical. An automated test enforces it on each release.

The two places Sustain uses AI

Ask Sustain

Sustain retrieves passages from your own uploaded documents and, where relevant, a global regulatory corpus. It then sends your question and those passages to Mistral, our AI sub-processor, which composes the answer from them and cites what it used. The answer rests on real retrieved text rather than open-ended generation. If the model goes down, Sustain answers from the retrieved passages on its own.

Document extraction

Turning a messy PDF, spreadsheet or CSV into structured, provenanced facts (a supplier, a site, an emissions figure) can use a model to read a document the deterministic parser cannot interpret. It runs only when the rule-based pass finds nothing. Sustain asks it for rows, not for graph structure, and those rows go through the same validation as a spreadsheet. Sustain checks each value against unit and range rules before it reaches the graph, quarantines what the validator cannot accept for a person to review, and marks a model-read value with lower confidence than a parsed one.

The places it does not

Deciding whether a risk exposure exists
Calculating or adjusting a euro figure, an expected impact, or a p95 downside
Ordering or ranking the recommended actions
Approving, rejecting, or otherwise deciding anything on your behalf

Telling what you are looking at

Ask Sustain carries a persistent notice in the product itself stating that you are interacting with an AI system. Sustain marks any answer built from model-generated rather than retrieved text, in the interface and in the underlying API response (ai_generated).
Reports and exports carry a machine-readable line stating whether any of their content is AI-generated.
Recommendations are never AI-generated. Their rationale, financial figures, and ranking come from versioned, auditable reasoning and simulation code (see the Product page's explainability section for a worked example).

No human data used to train external models

Sustain does not use your uploaded documents or its answers to train, fine-tune, or otherwise improve a model, ours or a third party's. In the two places above, the external model processes your data for that single request and nothing else.

Human oversight

Sustain analysts govern the rules, thresholds, and assumptions behind each calculation. They review each change, test it against a fixed benchmark set, and audit it before publishing. Sustain produces the analysis; your management makes the decision. That boundary is deliberate and does not move.

AI reads your documents. It produces none of your numbers.