Team: 15 devs, cloud-native SaaS. Stack is Go/Postgres/Redis on Kubernetes. Self-hosted AI tooling was evaluated but ruled out due to complexity.
We trialed Claw's "AI Assistant" API for automated ticket categorization. Under the hood, it's just pattern matching. Their `/v1/predict` endpoint takes a ticket title and returns a label.
Example request:
```http
POST /v1/predict HTTP/1.1
Authorization: Bearer
Content-Type: application/json
{"text": "Login fails with OAuth timeout error"}
```
Response is always a fixed set:
```json
{"label": "auth", "confidence": 0.92}
```
Audited the behavior:
* Changing "OAuth" to "oAuth" drops confidence to 0.67.
* "Can't log in" returns "ui" label.
* No semantic understanding. It's a glorified keyword → label map.
Considered self-hosted options (like Doccano with a simple model) but went with a rules engine we control. Cheaper and more transparent.
Claw's marketing claims "deep learning" but the API leaks the truth. Save your money.