AI Disclosures

This page is a plain-language summary of how Watari uses AI. It supplements §3 of our Privacy Policy and §4 of our Terms of Service, which are the legally controlling documents.

1. What Watari does with AI

Watari is an AI-powered ticket-to-PR loop. Customer support tickets you connect (from Zendesk, Intercom) are processed through a four-stage AI pipeline:

1. Extract— Anthropic Claude reads the ticket conversation and produces a structured bug report (severity, repro steps, expected vs. actual, customer impact).
2. Map— the extracted bug is matched against your indexed source code using vector embeddings (OpenAI text-embedding-3-small) and reranked by Claude.
3. Fix— Claude drafts a pull request against your repository.
4. Close the loop— when you approve, a customer-facing root-cause analysis is posted back to the originating support ticket.

2. Models we use

• Anthropic Claude (Haiku and Sonnet families): used for ticket-to-bug extraction, code-mapping reranking, pull-request drafting, and root-cause analysis drafting.
• OpenAI text-embedding-3-small: used only for generating 1,536-dimensional vector embeddings of source-code chunks for semantic search. We do not send chat or completion prompts to OpenAI.

We may upgrade or change models from time to time; material changes are summarised in our changelog.

3. What is sent to AI providers

• To Anthropic: ticket bodies, conversation threads, source-code excerpts, code-location candidates, and structured intermediate fields used by the pipeline.
• To OpenAI: source-code chunks (text) for embedding generation, plus short bug-signature summaries (derived from extracted bug fields — title, description, severity, repro steps) which are embedded to enable cross-ticket clustering. Raw ticket conversation threads are not sent to OpenAI.

We do notsend the following to any AI provider: your OAuth tokens, your end users' payment information, your billing data, or any data you have configured the Service to exclude.

4. No training on your data

Watari does not use Customer Content to train any AI model. Anthropic and OpenAI are contractually prohibited from using Watari's API traffic to train their foundation models, consistent with Anthropic's Commercial Terms and OpenAI's API Data Usage Policy.

5. Provider retention

• Anthropic Claude API: content is not retained at rest beyond what is necessary to return the API response, except for safety-classifier results retained for abuse monitoring.
• OpenAI embeddings: content is retained by OpenAI for up to 30 days for abuse monitoring and then deleted.

Inside Watari, the outputs of these calls are stored against your account for as long as your subscription is active and for 30 days thereafter (see Privacy Policy §7).

6. Human-in-the-loop

Watari is a human-in-the-loop system:

• Pull requests generated by Watari are drafts and are never auto-merged. A human on your engineering team must explicitly review, edit, and merge them.
• Root-cause analyses are drafts and are never auto-published to your end users without an explicit approval action by your authorised user.
• You retain full control over which Outputs are accepted, edited, or rejected.

7. Hallucinations

Large language models can produce inaccurate, incomplete, or misleading outputs (“hallucinations”). Common failure modes for Watari include: misclassifying an unrelated user message as a bug, mapping a bug to the wrong file or function, drafting a fix for a symptom rather than a root cause, and drafting an RCA that omits relevant context. You must independently verify every Output before relying on it.

We surface AI confidence scores alongside extracted bugs and mapped code locations, and we bill only for “Mapped Bugs” that exceed the confidence thresholds described in our Terms of Service. If a billed Mapped Bug is misclassified, you have a 7-day window to dispute it — see Refund Policy.

8. EU AI Act — deployer transparency

Watari is a deployer of general-purpose AI systems within the meaning of Regulation (EU) 2024/1689 (“EU AI Act”). Outputs that may be visible to your end users (such as customer-facing root-cause analyses) are clearly labelled as AI-generated. Watari does not classify as a “high-risk” AI system under Annex III of the EU AI Act.

We comply with Article 50 transparency obligations effective from 2 August 2026. If you have questions about the specific mitigations and documentation we maintain, email privacy@watari.ai.

8a. AI-drafted Watari blog posts

Some posts on the Watari Blog may originate from an internal automated drafter. The Watari team reviews and edits every auto-drafted post before merge; nothing is published without human approval. Auto-drafted posts are attributed to the “Watari Curator” byline at /blog/author/watari-curator. Customer data is never used to seed or train the drafter.

9. Opt-out

The Watari Service is fundamentally an AI service. Opting out of AI processing is equivalent to discontinuing use of the Service.

You may delete your organisation and all associated Customer Content at any time from Settings → Organization → Delete (organization owner only) or by emailing privacy@watari.ai from any account.

10. Changes

We may update this page from time to time. Material changes to our AI model selection, processing flow, or training stance are notified by email to the registered owner of each Organisation at least 30 days before they take effect.