How we use AI

Kafu relies on an advanced family of generative AI models. The specific tier invoked depends on the task:

• Top-tier model: complex generative tasks (designing a survey from a text brief, analyzing response patterns, drafting reports).
• Balanced-tier model: conversational tasks and quick in-product summaries.
• Lightweight-tier model: lightweight repeated tasks (classification, pattern detection, logic validation).

We may add or change models as new versions are released, with this page updated and tenants notified before any material change.

Only the minimum necessary to perform the task you requested is sent to the model. Specifically:

• When generating a survey: the text brief you wrote, the survey type, and any constraints you set (target audience, question count).
• When analyzing results: aggregated, de-identified responses by default — unless you explicitly choose to include identifying fields (e.g. respondent names in a qualitative analysis).
• When chatting in-product: only the text of your question, plus the current screen context (e.g. the name of the open survey).

We never send, under any circumstance: passwords, payment data, your full account history, or other tenants' data.

Under our commercial agreement with our AI model provider, data we send via the API is not used to train or improve any model, and is not retained beyond completion of the request (except for a short safety-review window per the provider's policy). This is a contractual commitment, not a default setting that could change.

Likewise, Kafu does not train any in-house model on customer data. Product improvements rely on aggregated, de-identified usage metrics and your direct feedback only.

• Tenant-level: a tenant admin can disable all AI features from governance settings. When disabled, AI buttons disappear and no AI processing occurs.
• Survey-level: you can turn off AI features on a specific survey (e.g. conversational probing of respondents), even if enabled at the tenant level.
• Action-level: every AI generation starts with an explicit click from you. No automated generation without your authorization.
• Token-level: you can set a monthly AI spend cap to prevent overruns.

AI is an assistant, not a replacement. Every model output (a generated survey, an analysis report, a phrasing suggestion) appears as an editable draft before publishing. No AI-generated content is published automatically.

If you are materially affected by an AI-assisted decision (e.g. your response was excluded from an analysis) and believe it is inaccurate, you may request human review via privacy@kafuapp.com.

• AI outputs may contain inaccurate or biased information. Verify every output before relying on it, especially statistical analyses and interpretations.
• Language models can produce content that appears convincing but is incorrect (hallucination). The higher the stakes, the more important independent human verification becomes.
• Kafu does not guarantee 100% accuracy of any generated output, and is not liable for business or legal decisions taken on unreviewed output.
• Arabic-language bias: models continue to improve on Arabic, but quality may vary by dialect or context. Please report any bias you observe.

Every AI invocation is recorded in your tenant's audit log: who triggered it, when, which model, and the token count. The full log is available to tenant admins and is exportable.

For any question about AI use in Kafu, or to request human review of a specific output: privacy@kafuapp.com.