# Check what the AI did and what it cost

The AI runs log shows every answer, draft, hand-off and test the assistant made in the last 30 days, with its outcome and cost.

Every time the AI does something for your business, Ha Tang records it: the customer message it answered, what it decided, how long it took and what it cost. Use the log to understand a strange answer, to see why the assistant handed off, and to keep an eye on spending.

## The summary on the Assistant tab

The top of **Customer Support → AI assistant** shows four figures:

| Card | What it counts |
|---|---|
| **Spent today / month** | What the AI cost today and this month, with your daily and monthly caps underneath |
| **Runs, 30 days** | Every AI call in the last 30 days, split into replies and drafts |
| **Hand-offs** | Conversations handed to your team, with how many runs were stopped by the caps and how many failed |
| **Cost, 30 days** | The total of the last 30 days and the average time per answer |

## The AI runs log

Open the **AI runs** tab. It lists the latest runs, newest first. Runs are kept for 30 days.

| Column | What it shows |
|---|---|
| **When** | Date and time |
| **Kind** | What the run was for: `respond` (an answer to a customer), `draft` (a suggestion for an agent), `test` (the **Chat with the AI** panel), `eval` (a test on the **Tests** tab), `followup`, `workflow`, `comment` or `tagging` |
| **Outcome** | **Replied**, **Draft**, **Email request** (it showed the email form), **Handoff**, **Skipped**, **Capped** (the budget was used up) or **Error**, with the error underneath |
| **Conversation** | A link to the conversation, or "test" for runs made from the dashboard |
| **Reply** | The beginning of what it wrote. Actions appear in brackets, for example `[handoff_to_human: not_in_knowledge]` or `[show_products]` |
| **Tokens** | The size of the request and of the answer, and how much was reused from cache (which costs less) |
| **Cost** | The cost of this run in US dollars |
| **Latency** | How long the model took |

## Use it to fix problems

- **A wrong answer**: open the conversation from the log, check **Knowledge used** under the reply, and correct it with **Improve this reply**. See [Answer the questions the AI could not](https://hatang.dev/help/ai/knowledge-gaps).
- **Many hand-offs with `not_in_knowledge`**: answer the questions in **Unanswered questions**.
- **Many hand-offs with `max_turns`**: customers need more back and forth than the limit allows; raise **Max assistant replies before hand-off**.
- **Runs marked Capped**: the daily or monthly cap was reached and the assistant stopped answering. Raise the caps in **Team budget and key**, or wait for the next day or month.
- **Runs marked Error**: the message under the outcome says why. "no API key" means neither the platform nor the business has a key for the model.

## Keep costs predictable

- Set a **Daily cap** that matches a busy day, and a **Monthly cap** that matches your budget. Owners and admins get an email the first time a cap is hit each day.
- Tests and the **Chat with the AI** panel cost money too. They appear in the log with the kinds `eval` and `test`.
- AI follow-ups, AI comment replies and the **AI writes a message** workflow step count toward the same caps. If spending jumps, the **Kind** column shows which one grew.
- Answers reuse the fixed part of the request from cache, which is why the **Tokens** column often shows a cached share.

For charts of AI activity over time, add the **AI assistant** and **AI by day** widgets to a dashboard. See [Custom dashboards](https://hatang.dev/help/reports/custom-dashboards).