Safety, privacy, transparency by default
AI that works for you, not on you
Kiaro is an AI-native product. That means we thought about safety, privacy, and transparency before we shipped, not after. Here is exactly how AI works in Kiaro and what it can and cannot do with your data.
AI proposes, you confirm
When you use the AI command bar, read operations (searching, viewing data) execute immediately. But write operations (creating a client, sending an invoice, updating a status) are always staged as a plan for you to review.
Nothing touches your database until you click “Confirm.” You can review every proposed action, remove individual items, or cancel entirely. This is not an AI guardrail added after the fact. It is the core architecture.
Your data is never used for training
Data Processing Agreements are in place with every AI provider behind Kiaro's built-in AI: Anthropic (Claude), and OpenAI (GPT) and Google (Gemini) as backups. All three contractually commit that your data is never used to train, fine-tune or improve their models. AI apps you connect yourself follow your own agreement with their provider.
Providers keep requests for a limited time to detect abuse: Anthropic deletes them within 30 days, OpenAI keeps abuse logs for up to 30 days, and Google keeps abuse logs for a limited period.
Data minimization by design
AI features only activate when you use them. Your data is never sent to AI providers in the background. When you use the AI command bar, only the data relevant to your request is sent to the provider, scoped to your workspace.
Kiaro includes a data sanitization layer that strips sensitive fields (Stripe IDs, authentication tokens, file URLs, and banking details) from AI context. We are continuously expanding the set of fields excluded from AI processing to minimize data exposure.
Tenant isolation is absolute
Every AI query is scoped to your workspace. The AI can only access data belonging to your tenant. There is no cross-tenant data leakage because every database query filters by your tenant ID at the query layer. This is enforced in every Server Action and API route, not just the AI features.
Multi-model routing with full transparency
Kiaro routes AI requests to the most appropriate model for the task. Complex generation (proposals, contract analysis) uses a powerful model like Claude Sonnet. Simple lookups and classification use a fast, cost-effective model like Claude Haiku. If Anthropic is unavailable, requests fall back to OpenAI, or to Google for the weekly digest, so your work isn't blocked by an outage.
You pay a single subscription price with unlimited AI commands on Solo. You never need to think about which model is used or worry about per-token costs. After very heavy use (3,000 commands in a month on Solo), complex requests start on the faster model until the count resets on the 1st.
AI assists, never acts autonomously
There is no background AI processing. Kiaro does not send emails, create invoices, modify projects, or perform any action without your explicit instruction and confirmation. Smart suggestions on your Home page are derived from your workspace data on each request and surface the highest-leverage next actions, but they are read-only and only take action if you click them.
Where your data is processed
AI providers process requests outside the EU. Anthropic and OpenAI process in the United States, and Google in the United States and other Google regions. Transfers are covered by Standard Contractual Clauses. Your database is hosted on Neon PostgreSQL in the EU (Frankfurt), and files are stored encrypted on Cloudflare R2 in the EU. See the full subprocessor list.
Rate limiting and abuse prevention
AI endpoints are rate-limited to prevent abuse. The AI command bar is limited to 20 requests per minute per workspace. Read-only AI operations count as half a command. All AI usage is logged for billing and auditing purposes, with logs automatically purged after 90 days.
Talk to us
Questions about our AI practices?
We are happy to discuss our approach in detail. Reach out at hello@kiaro.io or visit our Security and Privacy pages for more.