# AI features
Triage, summaries, sentiment, duplicate detection, drafted replies, and transcription — what runs automatically, what you invoke, and what's gated.
Source: https://docs.annot8.app/docs/ai/overview
Annot8's AI routes through Vercel AI Gateway. Text, structured output, vision,
and agent work use OpenAI GPT-5.6 Luna; audio uses the dedicated OpenAI
transcription model. AI features are available on **Pro** and above.
## Automatic: triage on arrival [#automatic-triage-on-arrival]
When a thread arrives without a priority or type, AI triage fills them in and
proposes a title. The result is written onto the thread and marked as
AI-triaged, so you can tell a machine guess from a human decision.
Triage looks at the comment text, the page, the element, and the diagnostics
already attached — not just the sentence someone typed.
## On demand: the thread toolbar [#on-demand-the-thread-toolbar]
Open a thread and you get:
| Action | Does |
| --------------------- | ------------------------------------------------------------------------------------------- |
| **Summarise** | Condenses a long thread into the issue, the specific problems, and any action items |
| **Sentiment** | Reads the thread as positive, neutral, or negative |
| **Suggest priority** | Proposes a priority with reasoning |
| **Categorise** | Proposes an issue type |
| **Detect duplicates** | Checks the thread against existing feedback so the same bug reported five times becomes one |
| **Draft reply** | Writes a first-pass response for you to edit |
| **Extract actions** | Pulls out the concrete action items |
| **Rewrite** | Tightens a comment you're about to post |
Nothing here posts on its own. Every AI output is a draft you accept, edit, or
throw away.
## The dashboard assistant [#the-dashboard-assistant]
Press Ctrl/Cmd + J anywhere in the dashboard to open the
assistant. It has context on your projects and feedback, so you can ask
questions like "what's the most complained-about page this week?" instead of
building the filter by hand.
## Transcription [#transcription]
Voice notes and screen recordings are transcribed to searchable text. A
transcript carries a status — pending, done, or failed — so a thread never looks
like it's missing audio when transcription is simply still running.
Transcription of **visitor-submitted** media is gated. It only runs when the
project's audience is classified as **Business users aged 18+** and visitor AI
is enabled on the deployment. See [Visitor-facing
AI](#visitor-facing-ai-is-gated) below.
## Heatmaps [#heatmaps]
Annot8 can analyse a page screenshot and generate an attention overlay,
highlighting the zones a visitor's eye is drawn to. It's a design-review aid, not
a measurement of real visitor behaviour — for that, use
[Analytics](/docs/analytics/overview).
## Model routing [#model-routing]
Model routing is operator-owned so workspace data cannot switch providers or
change data-processing terms. There is no per-project model picker.
| Workload | Gateway model |
| ----------------------------------- | ------------------------------- |
| N8, triage, summaries, and vision | `openai/gpt-5.6-luna` |
| Durable Eve draft-PR Agent | `openai/gpt-5.6-luna` |
| Voice notes and recording narration | `openai/gpt-4o-mini-transcribe` |
If Gateway authentication is missing, the request fails clearly. Annot8 does
not silently fall back to a direct provider.
## Usage limits [#usage-limits]
AI triage operations are metered per workspace, per calendar month.
| Plan | AI triage / month |
| ---------- | ----------------- |
| Free | Not included |
| Pro | 500 |
| Business | Unlimited |
| Enterprise | Unlimited |
Hitting the limit disables AI actions until the month rolls over; everything
else keeps working exactly as before.
### Agent Credits [#agent-credits]
Resolve with Agent uses a separate workspace pool of **Agent Credits**. One
credit is one autonomous run on frontier models: the Agent reads the thread,
finds the responsible code, and can open one draft pull request. Included
credits reset each UTC calendar month; purchased credits never expire.
| Plan | Included Agent Credits / month |
| ---------- | -----------------------------: |
| Free | 0 |
| Pro | 10 |
| Business | 20 |
| Enterprise | Contracted pool |
Paid workspaces can top up with 15 credits for $19, 45 for $49, or 120 for $119.
Annot8 reserves a credit when a run starts and automatically returns it when
the platform fails or Eve cannot produce a safe patch.
## Visitor-facing AI is gated [#visitor-facing-ai-is-gated]
There's an important asymmetry: AI that helps *your team* is available on any
paid plan, but AI that talks to *your visitors* — live chat auto-replies and
transcription of visitor-submitted media — is fail-closed and requires **both**:
1. The project's **audience** in **Settings → Configuration** is set to
**Business users aged 18+**.
2. Visitor AI is enabled on the deployment.
This exists because the model provider's terms prohibit services likely to be
accessed by under-18s. A project whose audience is Unknown, General public, or
Under 18 will not get visitor-facing AI, whatever the plan says — and that is
deliberate, not a bug.
If your AI agent isn't replying in live chat, this is nearly always why. See
[the AI agent](/docs/chat/ai-agent).
## What we don't do [#what-we-dont-do]
* No AI output is published to a visitor without a human in the loop, other than
live chat auto-replies you explicitly turn on.
* Model choice is controlled by the Annot8 operator and audited in Gateway
usage; projects cannot inject a provider or model id.
* AI usage is logged per workspace for metering. See
[Privacy](/docs/security/privacy) for what's sent where.