AI Policy

Where AI is used

Language models are used for interpretation and communication tasks, where natural language is genuinely the right tool.

• Understanding what you are asking and translating it into a structured query.
• Explaining a computed result in plain language.
• Summarising narrative content such as risk disclosures, with citations.
• Classifying transactions where deterministic rules do not resolve the case.

Where AI is deliberately not used

Every financial figure NanoFora reports is computed by deterministic queries and formulas over your data. Models do not produce, estimate or adjust numeric values.

Scenario outcomes, ratios, variances and forecasts are calculated arithmetically from stated assumptions. The model may explain them; it never generates them.

Handling uncertainty

When the available data cannot support an answer, NanoFora says so and states what is missing. Producing a plausible answer in place of an honest limitation is treated as a defect, not a feature.

Document extractions below the confidence threshold are flagged for human review rather than silently accepted.

Traceability

Every figure links back through the transactions used to compute it to the source document and page. Every summarised statement cites the passage it came from. If a claim cannot be sourced, it is not shown.

Data and training

Your company data is processed only to answer your own questions within your own workspace. It is not used to train shared or third-party models, and model providers are contractually prohibited from retaining it for training.

Human responsibility

NanoFora is a tool for analysis, not a decision-maker. Professional judgement, statutory reporting and regulated advice remain with qualified humans. We state this wherever AI output could be mistaken for professional advice.

Known limitations

We publish limitations rather than waiting for you to discover them:

• Forecast quality depends on the length and consistency of your history.
• Extraction accuracy varies with document quality, particularly for low-resolution scans.
• Categorisation of genuinely ambiguous transactions requires human judgement.
• Analysis reflects the data provided; unrecorded activity cannot be accounted for.