Qualflare offers AI-assisted features inside the application — for example, AI-suggested test cases, defect clustering, and natural-language analysis of test results.
You are interacting with an artificial intelligence system. When you use these features, your input and the resulting output are processed by an AI model and may not be accurate, complete, or appropriate for every situation. Treat the output as a draft that needs human review.
Which models. We use three model providers, all reached through Cloudflare AI Gateway for caching, rate limiting and observability. Google's Gemini API is our main one: it powers search embeddings, failure clustering, launch analysis, natural-language query generation and the default conversational agent. OpenAI powers the Quo conversational agent on workspaces we have enabled for it. Clef, Cloudflare's own decision model on Workers AI, answers specific typed questions — how to route a request, and a failure cluster's category, severity, membership and redundancy — returning a category and a probability rather than generated text. Until 21 October 2026 those questions went to TypeSafe's Jev model instead; TypeSafe no longer receives any of your data.
Where the data goes. Prompts and outputs sent to Google are processed in accordance with the Google Generative AI terms and may be processed in the United States or other regions in which Google operates. Requests to OpenAI are processed in the United States, on OpenAI's standard API endpoint, with retention disabled since 28 September 2026. Clef requests run on Cloudflare's own Workers AI network, and we redact credentials, email addresses and identifiers we recognise from them before they are sent. Every AI request passes through Cloudflare AI Gateway, which retains the request and the response in its logs. All of these transfers rely on the safeguards described in section 9.
Our AI providers do not train on your data. Your prompts, outputs, test artifacts and workspace content are not used to train, fine-tune, or evaluate their foundation models. Google Gemini's enterprise terms exclude API inputs and outputs from being used to improve Google's models, and OpenAI's API terms exclude data submitted through the API from training — our OpenAI Data Processing Agreement is in progress and not yet executed. For Clef, Cloudflare commits that it does not use content processed by Workers AI to train AI models; we hold that commitment directly as a Cloudflare customer. The content sent there is redacted first.
We use records of AI requests to improve our own features. So that Qualflare's defect clustering and analysis get better over time, we keep a record of AI requests and their results — the prompt, the output, and the failure messages and test names involved. Before anything is stored we remove credentials, API keys and email addresses. The record is stored in Cloudflare D1, so Cloudflare is the sub-processor for it (see section 8) rather than it sitting only on our own servers. We use these records only to improve Qualflare's own features; we do not sell them, and we do not share them with our AI providers for their training.
The records carry no identifier linking them to you — no workspace identifier, no project identifier, and no launch or cluster identifier. There is nothing in them to match back to your workspace, so they are anonymous rather than merely pseudonymised.
That was not true until recently, and we would rather say so than quietly correct it. Until 29 September 2026 some rows kept a launch or cluster identifier, and launch-analysis rows kept workspace, project and launch identifiers in their metadata — any of which could resolve a row to the workspace it came from. Those records were pseudonymised, not anonymous, while this notice described them as anonymous and outside the scope of data-protection law. On 2 October 2026 we deleted every record written before the fix — 1,984 AI-request records and 51,895 text pairs, the entire set collected up to that point — rather than only stripping the identifiers out of them, because stripping could not reach an identifier quoted inside a stored failure message. Collection started again from the fixed code, so the record set is now genuinely anonymous.
Opt out. Contributing to that record is on by default, and any workspace admin can turn it off under workspace settings → privacy. Turning it off stops all future collection for that workspace, normally within five minutes.
It cannot remove records already collected: they carry no identifier, so we cannot locate the ones that came from your workspace in order to delete them individually — which is the same property that makes them anonymous. Everything collected before 29 September 2026, when identifiers were still present, was deleted outright on 2 October 2026. You can also disable AI features entirely at any time from your workspace settings.
Risk classification. Qualflare's AI features are not, in our assessment, “high-risk AI systems” under Annex III of the EU AI Act (Regulation (EU) 2024/1689). They do not make eligibility, employment, credit, biometric, or law-enforcement decisions. They assist QA engineers with test management and analysis. We will reassess this classification if we change the scope of our AI features.
Synthetic content. Where AI generates substantial new content (for example, suggested test cases or defect summaries), we label that content as AI-generated in the user interface in line with Article 50(2) of the EU AI Act.
