AI Policy

Kira ranks roles against your outcomes rather than your keywords, drafts follow-up messages you review before sending, summarises long listings, and prepares interview notes from a role and your showcase.

On the employer side Kira summarises applications, suggests a shortlist order, and drafts outreach. Every one of those outputs is presented as a recommendation with the reasoning shown, never as a decision.

Kwirl does not permit fully automated adverse decisions. An employer cannot configure Kwirl to auto-reject candidates on a Kira score, and no such feature exists in the product.

A named person must review and record a reason before any candidate is rejected. We log the reviewer and the timestamp, and a candidate can ask whether a human reviewed their application.

This is stricter than most jurisdictions currently require, and it aligns with the direction of Australian automated-decision reform, the New Zealand Algorithm Charter, and the transparency duties in United States state and city law.

We test matching and ranking outputs quarterly for disparate impact across gender, age band, name origin, and disclosed disability, using held-out evaluation sets rather than live user data.

Where a test shows material drift we hold the model version and correct it before release. Summary results are published in our transparency report.

Employers using Kwirl for candidate screening in New York City are responsible for their own annual bias audit and candidate notice under Local Law 144. We provide the model documentation and audit artefacts they need, but the audit obligation is theirs.

We do not train foundation models on your Kira conversations, your showcase, or your applications unless you switch model improvement on in settings. It is off by default and switching it off later stops future use immediately.

We use aggregated, de-identified interaction signals — such as which of two ranking orders users preferred — to tune retrieval and ranking. That data carries no identifiers and is not reversible.

We never train on sensitive information, on messages between a candidate and an employer, or on material an employer has marked confidential.

Kira session and library data you keep is retained while your account is open; sessions and items you delete are removed about 30 days later. Opted-in activity signals used for personalisation are kept for up to 90 days. See the Data Retention and Deletion Policy for the full schedule.

Language models produce fluent text that can be factually wrong. Kira can misread a listing, invent a detail about a company, or overstate a match. Read what it drafts before you send it.

Kira does not give legal, migration, tax or financial advice. Where it summarises visa or right-to-work information it is repeating what the employer published, not advising you.

Anything Kira generated is labelled in the interface. Drafts are marked as drafts and are never sent without your action.

You can read every fact Kira has remembered about you, delete any of them, and turn memory off entirely. Turning it off degrades match quality rather than breaking the product.

You can ask what data informed a particular match. The reasoning panel lists the signals used.

We use third-party model providers for inference under contracts that prohibit training on our traffic and require deletion of prompt and completion data within 30 days.

Inference may occur outside Australia. Prompts carry the minimum context needed and strip direct identifiers where the task allows.