What an AI assistant can see and do in your account

You choose on the consent screen

When an assistant connects, RawQuotes shows you a consent screen: which assistant is asking, the e-mail you're signed in with, and the permissions it can get, grouped by area — projects, surveys, sources, responses, insights and tracked insights. Nothing is shared until you click Allow access.

• Everything except deleting starts switched on. Untick whatever you don't want to give.
• See your projects (projects:read) is always included: without it the assistant can't find anything.
• Only what your role allows is listed. An assistant can never do more than you can.
• Deleting stays off unless you turn it on — and today no assistant tool deletes anything.
• An app that only named itself is marked Name not verified. Connect it only if you started the connection yourself.

Each permission maps to the tools in Tools and permissions.

Responses are shown as written

With Read responses as written, including contact details (responses:read), the assistant gets the original text of every answer — nothing is masked. If a respondent typed an e-mail address or a phone number, the assistant sees it. Leave this permission off and list_responses is unavailable. The quotes under insights are verbatim too, and they come with insights:read.

Insights (insights:read) are the same insights you see in the panel, each with its count and the verbatim quotes it was counted from. Every quote carries its responseId, so you or the assistant can open the response it came from and check it.

Below the threshold, get_insights returns no insights — only how many complete responses the project has and how many it needs — and the assistant is told not to make them up from raw responses. That refusal is deliberate: a count on too little evidence would be a guess. Verification and the threshold explains the thresholds.

Respondent text is not instructions

Answers come from anonymous people. RawQuotes marks every response and quote it hands over as untrusted and tells the assistant to treat it as data, never as instructions — but a crafted answer can still try to steer an assistant (prompt injection). Don't give an assistant that reads responses any other tool whose effects you don't want.

Where the data goes

Whatever the assistant reads is sent to the provider of the AI model you chose for it — Anthropic, OpenAI or another. RawQuotes doesn't control what happens to it there. As the account owner, you are responsible for having a legal basis for that transfer, for any transfer outside the EEA, and for telling your respondents about it.

Taking access back

User settings → AI assistants → Disconnect cuts an assistant off at once. See Connect an AI assistant.