Responsible AI Policy

1. Our position on AI

Briefable is an AI-powered productivity platform. AI is embedded throughout the product — in morning briefings, email classification, task extraction, draft responses, knowledge search, pipeline monitoring and meeting preparation.

We believe AI should reduce cognitive load, surface what matters and prepare the work — but the person using it should always make the decision. AI is the preparation layer. Humans remain the judgement layer.

This policy sets out how we design, build and operate Briefable’s AI features responsibly.

2. Core principles

3. Human oversight

Human oversight is embedded in the product architecture, not added as an afterthought:

• Draft replies are prepared by AI but staged for your review. They are never sent automatically.
• Task extraction identifies potential tasks from emails and conversations. You confirm which become tracked work.
• Content publishing follows a draft → review → approve workflow. AI generates; you decide what goes live.
• Priority recommendations are suggestions, not commands. You can override, reorder or dismiss them.
• Meeting preparation assembles relevant context. You decide what to use and how to act on it.
• Pipeline signals (stale deals, follow-up reminders) are surfaced as alerts. The response is yours to choose.

Where the Service offers automated steps (such as email classification or task extraction), these operate on your data within your workspace — they do not take external actions or communicate with people on your behalf.

4. Transparency

We are committed to being clear about how AI operates within Briefable:

• AI-generated content is identified. Where the Service produces AI-generated text (briefings, draft replies, summaries, recommendations), it is presented as AI-prepared content requiring your review.
• Source attribution. Briefings, meeting preparation and knowledge answers reference the connected sources used to generate the output, so you can verify the information.
• Model selection is visible. In the Knowledge module, you choose which AI model to use and can see the credit impact of that choice.
• System routing is documented. For non-interactive features (email classification, task extraction, priority signals), Briefable selects the most appropriate model automatically. The routing logic is documented and not hidden.
• Limitations are stated plainly. We do not overstate what AI can do. AI outputs may be wrong, incomplete, outdated or inappropriate. This is communicated in the product interface, onboarding materials and Terms of Use.

5. Data privacy and security

AI features process your data to deliver the Service. We apply the following safeguards:

• No training on your data. Your content, communications, files, contacts and pipeline data are not used to train, fine-tune or improve foundation AI models — ours or any third party’s.
• Scoped access. Briefable accesses only the data sources you explicitly authorise during onboarding. You can revoke access at any time.
• Minimisation. We send the minimum context necessary to AI model providers to generate useful responses. We do not send your entire dataset for every query.
• Encryption. Data is encrypted in transit (TLS) and at rest. Authentication uses industry-standard OAuth 2.0 via Google and Microsoft.
• Infrastructure. The Service runs on Supabase (database, authentication, edge functions) and Vercel (application hosting). AI processing uses Anthropic‘s Claude API.
• Retention. We retain your data for as long as your account is active and for a reasonable period after termination to allow data export. Details are in our Privacy Policy.

6. Accuracy and limitations

AI-generated outputs are useful but imperfect. We design the product around this reality rather than hiding it:

• AI can be wrong. Language models can produce inaccurate statements, misinterpret context, hallucinate content, miss important information or generate inappropriate suggestions. This is inherent to the technology.
• Outputs are not verified facts. AI-generated summaries, classifications, recommendations and draft responses should be treated as starting points that require human review, not as authoritative conclusions.
• Quality varies. Output quality depends on the model used, the context available, the complexity of the task and the quality of the connected data. Results are not guaranteed to be consistent across sessions.
• Not a professional service. Briefable is a productivity tool. It does not provide legal, financial, medical, tax, compliance or professional advice. Outputs must not be relied upon as such.

We continuously review AI performance and user feedback to identify patterns of inaccuracy, update system prompts and improve output quality within the constraints of the underlying models.

7. Fairness and bias

Foundation AI models carry biases from their training data. We take the following steps to mitigate their impact within Briefable:

• Task-specific prompting. System prompts are designed for specific operational tasks (classification, summarisation, preparation) rather than open-ended generation, which reduces the surface area for biased outputs.
• No profiling or scoring of people. Briefable does not use AI to assess, rank, score or profile individuals based on personal characteristics. Contact and relationship features track business context (last contact date, deal status, follow-up signals) — not personal attributes.
• Human review as a bias check. Because all consequential outputs require human approval before action, users have the opportunity to identify and correct biased or inappropriate AI suggestions.
• NZ English and cultural context. We design for the New Zealand market, including NZ English voice input and prompts that reflect New Zealand business practices and terminology where relevant.

We do not claim that AI outputs are free from bias. We design the product so that bias can be identified, reviewed and overridden by the person using it.

8. AI model providers

Briefable uses third-party AI models provided by Anthropic (Claude). Specifically:

• Claude Haiku 4.5 — used for high-volume, lower-complexity tasks such as email classification, task extraction and priority signalling.
• Claude Sonnet 4.6 — used for balanced tasks including morning briefings, draft responses, meeting preparation and knowledge search (default user-facing model).
• Claude Opus 4.6 and Opus 4.8 — available in the Knowledge module for users on Pro and Business plans who need deeper reasoning or more complex analysis.

Anthropic’s commercial API terms commit to not training on customer data sent via the API. We select Anthropic because of their published commitment to AI safety and their data handling practices. However, we cannot guarantee third-party compliance and recommend users review Anthropic’s policies independently.

We may add, change or replace AI model providers in the future. Material changes to providers or models will be communicated to users.

9. What Briefable will not do

We maintain explicit boundaries on what AI does within the product:

• Will not send emails or messages without your explicit review and approval.
• Will not publish content without passing through the draft → review → approve workflow.
• Will not make purchases, sign contracts or commit resources on your behalf.
• Will not delete your data without your explicit instruction.
• Will not access tools or systems you have not explicitly authorised.
• Will not profile, score or assess individuals based on personal characteristics, protected attributes or demographic information.
• Will not make autonomous decisions about hiring, lending, insurance, legal matters or any domain with significant human consequences.
• Will not store or process data for purposes beyond delivering the Service as described.
• Will not use your data to train, fine-tune or improve AI models.

10. New Zealand context

Briefable is built and operated in New Zealand. Our approach to responsible AI aligns with:

• New Zealand’s AI strategy and guidance for business — published by the Department of Internal Affairs and the Ministry of Business, Innovation and Employment.
• The Privacy Act 2020 — New Zealand’s primary privacy legislation, governing how personal information is collected, used, stored and disclosed.
• The Consumer Guarantees Act 1993 — where it applies to the Service, we acknowledge the consumer protections it provides.
• The Fair Trading Act 1986 — we do not make misleading claims about AI capabilities. We state limitations plainly.

We also monitor developments in international AI governance, including the EU AI Act and OECD AI Principles, to ensure our practices remain aligned with emerging standards that are likely to influence New Zealand’s regulatory direction.

11. Governance and review

This policy is reviewed at least annually and updated when material changes occur to:

• The AI models or providers used by the Service.
• The scope of AI functionality within the product.
• Applicable New Zealand or international regulations.
• User feedback indicating safety, accuracy or fairness concerns.

The policy is owned by the director of Specialeyes Ltd (trading as Changeable) and applies to all AI features within the Briefable product, including standard subscription and custom build implementations.

User reports of AI safety concerns, inaccurate outputs, biased behaviour or data handling issues are treated as priority feedback and reviewed within five business days.

12. Contact

To report an AI safety concern, ask a question about this policy, or request information about how your data is processed by AI features, contact us at:

Briefable
Email: hello@briefable.co.nz
Website: briefable.co.nz