It reads with your permissions.
SAI uses the login of the person asking. Postgres, not the model, decides what it can see. At a permission wall it declines politely instead of guessing.
It drafts, you approve.
Tasks, invoices and memory entries from SAI stay drafts until a person approves them. Draft invoices never reach a client: no email, no proforma, no reminder.
Amounts come from code.
When a meeting reveals extra work, the model estimates hours. The amount is hours times your project rate, calculated in code, never by the model.
It refuses when data is thin.
Brand analysis needs three analyzed meetings or two completed projects. Below that, SAI says so in your language without calling a model.
Every call is logged with its model and cost.
Each model call is recorded with the model, the tokens and the dollar cost. Settings shows your agency's last 30 days.
AI included, metered openly.
SAI runs on SmoothOS's own Anthropic and Google accounts. Each plan includes monthly AI credits, and every call is logged with its model and cost.
| Job | Tier | Provider |
|---|---|---|
| Reading long material: full meeting transcripts and documents | bulk | Google, Gemini 3.6 Flash |
| Judging scope, money and tools: answers built from your records | reason | Anthropic, Claude Sonnet 5.5 |
| Short questions with no record on screen | quick | Anthropic, Claude Haiku 5.5 |
| Meeting scope check: in or out of scope, with hours | reason | Anthropic, Claude Sonnet 5.5. It sees the extracted requests, the summary and your signed proposal, never the transcript. |
| Monday Brief | none | No model. Written by rules from your numbers. If your agency has a key, Claude Haiku 5.5 adds a short first-things-first note. |
| Brand analysis refusal when data is thin | none | No model is called. |
Meeting data
Today, analysis runs on transcripts delivered to SmoothOS. A recording bot is planned and isn't available yet. Transcripts, summaries, decisions and draft tasks are stored in your agency's database under the same row-level security as everything else. Only the reading model sees the full transcript; the model that judges scope never does. Transcripts are deleted with your agency 60 days after a subscription ends. A retention period for transcripts while you're subscribed hasn't been set yet, and we'll publish it here before general launch.
