Ethics & security
Useful AI should earn your trust.
AI-R is built to support independent businesses while keeping people responsible for important choices. Here is how the assistant is designed, what protects the workspace, and which safeguards still need to be completed before a full production launch.
Responsible use
People and judgement come first.
AI should help with repeatable work and clear information. It should not hide uncertainty or remove a person from decisions that need care.
01People stay in control
Customer assistants return answers the business has approved. When AI access is available, matching runs automatically and can choose an approved answer or an excerpt from an approved reference file. When an answer is unclear or unsupported, the assistant points the customer to a person. AI-R does not make consequential decisions about people.
02Clear limits on AI
AI-R’s website assistant can draft general replies using approved AI-R information. Customer assistants use AI to select existing wording from business-approved answers and reference excerpts. AI-generated social content should be checked by a person before it is published.
03Privacy with restraint
Chats with AI-R’s own website assistant are not stored. Published customer assistant conversations are saved in the owning business’s private workspace so the owner can review them; they are deleted automatically 90 days after the last message. When AI is enabled, a visitor’s question may be sent to OpenAI with the relevant approved facts or answers needed for a response. Provider handling depends on the provider’s terms and account settings.
04Honest and inclusive service
We aim to explain what a tool does, where it has limits, and when a person should take over. AI-R is not designed for children or for high-impact automated decisions such as decisions about eligibility, employment, lending, housing or access to services.
Security controls
What the app protects today.
These controls are implemented in the application and database setup. They still need to be checked against the live AWS and Supabase configuration before launch.
01Account access and data boundaries
The app uses Supabase Auth, server-side session checks and database policies that scope customer records to their account. The production and staging projects still need their deployed migration and policy state verified as part of launch checks.
02Private customer files
The Asset Library uses a private Supabase Storage bucket with account-scoped access. A customer can explicitly attach a JPG, PNG or WebP to a product/service listing or approved answer; only selected images are served through the app to visitors. Those assistant display images are not sent to AI. Separately, the Social Content Studio can send a browser-resized photo to OpenAI for analysis only when the owner chooses that action; AI-R does not save the photo and disables API storage for that request. Other unselected files remain private. Uploads have a 20 MB library limit and are not malware-scanned.
03Protected billing and secrets
Stripe webhook signatures and configured prices are checked before billing updates are applied. Payment processing is handled by Stripe; AI-R does not store full card numbers in its application database. Server secrets are read by server-side code and must not be placed in browser-visible settings.
04Request and browser protections
The app checks the origin of browser changes, validates request types and sizes, applies basic chat rate limits, and configures security response headers. Published customer chat limits are shared across app instances. For each bot, an IP address can send up to 20 messages per minute and 100 per UTC day. The shared counters use a keyed hash of the address.
When a question uses AI
AI-R’s Responses API calls set store: false. The AI-R website assistant can send a visitor question and approved public facts to OpenAI for a reply, but AI-R does not save those chats. For customer assistants with AI access, an unmatched question, up to eight relevant approved question-and-answer pairs, and up to three approved reference excerpts totalling 3,200 source characters may be sent to OpenAI to select one source. The visitor’s question and assistant reply are saved in the owning business’s private workspace and deleted automatically 90 days after the last message. Social-content briefs and voice profiles are sent only when the owner requests draft ideas. Review the provider’s current terms, usage limits, retention and regional settings before using live customer data.
Current limits and launch checks
We do not claim security is finished.
- Two-factor authentication is not currently available in the AI-R app.
- No independent penetration test or third-party security certification is claimed.
- The project deployment notes say the AWS app and WAF were not yet deployed or associated. Confirm the current AWS state, then activate shared edge rate limits before public multi-instance traffic.
- Confirm production and staging database policies, backups, monitoring, alert ownership and incident response before inviting customers.
- Uploaded files are not checked for malware. Keep passwords, payment details and sensitive personal information out of the Asset Library and assistant chats.