AI Control
AI Control is the one place to manage AI for a server. Open the dashboard sidebar and pick AI Control under your server. It brings together your token usage and budget, how the assistant sounds, what it is allowed to answer from, how often it may reply on its own, what gets captured for review, the Memory Library, and the results your team is getting.
Pro capability
AI is part of the Pro plan. On Free and Community, AI Control opens in a read-only preview so you can see exactly what Pro does for you. Nothing is saved and the Memory Library stays empty until you upgrade.
You need the Manage Settings permission (an admin role) to change anything here. Read-only staff can view the page, but the controls are disabled for them.
Usage and tokens
The usage panel shows what your server has spent on AI and what is left to spend. AI Assist, the AI flow nodes, and the automation AI reply action all meter against the same budget.
That budget is two separate pools, and the panel gives each one its own bar:
Monthly allowance
The tokens your plan includes each month, and how many of them you have used, so you always know where you stand before the cycle resets.
Top-up tokens
A separate balance that never expires. Nothing touches it while the monthly allowance still has room, and it is used automatically once that runs out.
Bring your own key
Add your own AI provider key to run on your account. Usage is still metered and shown, and your own key removes the monthly cap.
Top up on the usage panel is where you buy more: pick an amount, pay, and the tokens land in the top-up balance. It needs permission to manage billing, and it only shows on plans that meter AI usage, so it is not offered while you are running on your own provider key. A server that belongs to an organization is metered against the organization's shared pool, so its top-ups are bought on the organization's billing page instead.
If a plan includes no monthly allowance at all, the panel drops the monthly bar and the server simply runs on whatever top-up tokens it has.
Here is the AI token budget each server plan includes per month:
| AI budget | Free | Community | Pro |
|---|---|---|---|
| AI tokens / month | 0 | 0 | 5,000,000 |
Staff tools
The staff tools panel controls which AI tools your team can reach, in the ticket viewer and in Discord.
| Setting | What it does |
|---|---|
| AI summary | Shows the Summary tool in the ticket viewer's AI panel |
| AI suggested reply | Shows the Suggested reply tool in the ticket viewer's AI panel |
| AI commands in Discord | Keeps the staff AI slash commands, such as /ask, /draft, and /summarize, and the message right-click AI actions registered in your server. Turn it off and the bot removes them |
The AI commands only exist on the Pro plan: on other plans the bot does not register them at all, so there is nothing to hide. See AI Assist for what each command does.
Voice
The voice panel sets how the assistant sounds when it writes for your server. Everything here applies to auto-sent replies, staff drafts, and AI Assist suggestions alike.
| Setting | What it does |
|---|---|
| Persona | Custom instructions that shape every AI reply: brand voice, sign-offs, phrases to avoid, house rules |
| Reply language | What language replies are written in. Auto by default, or pin one |
| Reply style | Concise, friendly, formal, or apologetic tone for AI replies |
| Keep replies brief | Nudges the model toward shorter answers by default |
The persona is a short block of your own words, for example "You are the support voice of Acme. Be warm, use first names, and never promise a refund without a staff member confirming." Drafts in the ticket viewer follow it too, so a staff member starts from text that already sounds like you.
Reply language is a separate setting from the persona, and it is the one to use if you want answers in a particular language. On Auto, the default, the assistant answers each customer in the language they wrote to you in, so a server with customers in several languages needs no configuration. Pick a language from the list to pin every reply to it regardless of what the customer wrote, or choose Other and type one. That box takes regional and register variants such as "Latin American Spanish", "Brazilian Portuguese", or "formal Japanese".
Asking for a language in the persona box instead is not reliable: the persona shapes tone, and the language setting is applied after it, so the two can contradict each other. Set the language here.
Grounding
Grounding is what keeps answers honest: instead of writing from general knowledge, the assistant retrieves the most relevant approved answers (Memory) and knowledge base articles first, and writes from those. The grounding panel controls the server-wide default and how strict the assistant is when nothing relevant turns up.
| Setting | What it does |
|---|---|
| Ground AI replies | Server-wide default: pull from Memory and the knowledge base before writing. Reply nodes inherit this unless set per node |
| Answer strictness | How the assistant behaves when retrieval comes back thin: strict, balanced, or lenient |
| Escalation role | A role to ping when a strict-mode auto-reply declines, so a human picks it up |
The strictness dial has three positions:
Strict
Only replies when an approved answer or article genuinely matches. If nothing grounds the reply, the bot stays quiet instead of guessing, and can ping your escalation role so a person takes over. Needs knowledge to match against, so read the note below before you pick it.
Balanced
The default. Replies when relevant material is found and leans on the conversation when the match is partial.
Lenient
No relevance cutoff. The assistant always drafts from the closest matches it can find. Best when your knowledge is thin but you still want a first attempt on every ticket.
Strict needs something to be strict about
Strict only applies once the server has knowledge to match: at least one published knowledge base article, or one approved memory entry, counting entries shared with you by your organization. Until then the assistant answers as if it were set to balanced, because strict with an empty library can only ever refuse and page a human. Your setting is left alone, the dial still reads Strict, and strict answers begin once the first article or approved entry exists.
Strict mode plus an escalation role
Once your library has answers in it, strict with an escalation role is the tightest way to run auto-replies: customers with a covered question get an instant grounded answer, and everyone else gets a human ping instead of a guess. On a brand new server, fill the library first.
Auto-reply limits
These dials bound how much the assistant does on its own, per ticket and per category.
| Setting | What it does |
|---|---|
| Max auto-replies per ticket | Caps how many automated AI replies a single ticket can receive, so a back-and-forth never turns into an AI loop. Set to 0 for no cap |
| Auto-reply categories | Restrict the first-message auto-reply to specific ticket categories. Leave empty to cover all of them |
| Reply to first customer message | Posts a grounded reply the moment a customer opens a ticket |
The Reply to first customer message toggle is the one-click way to answer new tickets. When it is on, a customer's first message gets a reply grounded in your approved knowledge, with no flow or rule to build. For more control, build a rule on the On customer's first message trigger instead. See Automations.
Start in draft, then send
If you want a human to approve first replies before they go out, build the reply with an AI Auto Reply in draft mode rather than using the toggle. Drafts post a private, staff-only suggestion with Send, Edit, and Discard buttons. See AI Assist.
Memory capture policy
The capture policy decides which moments in your support work are saved as memory candidates. Everything captured still lands in the review queue: nothing is reused in a reply until an admin approves it.
| Setting | What it does |
|---|---|
| Capture resolved tickets | When a ticket closes, distill its question and resolving answer into a candidate entry |
| Capture categories | Restrict resolved-ticket capture to specific ticket categories. Leave empty to cover all of them |
| Capture staff feedback | A thumbs-up on an AI reply, or a draft a staff member edited before sending, becomes a candidate too |
There is also a capture path that needs no dial at all: staff can save any message in a ticket straight into the library as an approved entry, because a staff member chose it deliberately.
In Discord, right-click the message, open Apps, and pick Save to AI Memory. A short form opens with the title, question, and answer filled in so you can correct them before saving. Clear the question and the message is saved as a fact instead.
In the dashboard ticket view, right-click the message and pick Save as AI fact to store it on its own, or Save as AI Q&A to open the same form, where you can pick which earlier message the answer belongs to.
Two kinds of memory
Every entry is one of two shapes, and you choose which when you save it:
| Kind | What it holds | Good for |
|---|---|---|
| Question & answer | A customer question paired with the answer that resolved it | Recurring questions, where the phrasing customers use helps the match |
| Fact | A single statement that stands on its own | Policies, prices, limits, and anything true regardless of who is asking |
Both are retrieved the same way and ground replies the same way. A fact simply avoids inventing a question that nobody asked, which used to be the only way to record one.
Memory Library
The Memory Library is where you curate everything the assistant is allowed to reuse. It holds every entry, whatever its status: approved answers that ground replies today, candidates waiting for review, and rejected entries kept for the record.
Beyond approve, edit, and reject, the library gives you the tools to manage a collection that grows on its own:
Bulk actions
Select many entries at once and approve, reject, or delete them together, so a backlog of candidates clears in one pass.
Tags
Label entries with your own tags and filter by them, so billing answers, onboarding answers, and product answers stay organized.
Kind, source and unused filters
Filter to just Q&A pairs or just facts, by where an entry came from (resolved ticket, feedback, manual, saved message), or to entries that have never grounded a reply, so dead weight is easy to spot and retire.
Import and export
Export the library as JSON or CSV, and import entries the same way, so you can seed a new server from an existing answer set or keep an offline copy. Imported rows with no question are stored as facts.
When you add an entry by hand, the library checks it against what you already have and warns you if a near-duplicate exists, so the same answer does not pile up twice.
Organization-shared memory
Servers in an organization also see entries shared at the org level. Shared entries are managed on the organization page and are retrieved into AI replies in every member server, so one curated answer set can serve your whole fleet. In each server's library they appear read-only with an org badge.
Retrieval of approved answers, nothing else
Memory is scoped to your server (or, for shared entries, your organization) and none of it is ever used to train an AI model. The assistant only reuses answers you approved. Nothing a customer said appears in a reply until an admin has approved it here, and capture is a cheap field save, not a fresh AI write, so building your memory costs you nothing extra.
Results
The results panel shows how AI is performing for your server over the last 30 days, built from real usage and staff feedback. Values with no basis yet read as no data rather than an invented number.
Staff shape these numbers as they work: a thumbs up on an AI reply, or an edit to a draft before sending, feeds the results shown here and, with feedback capture on, becomes a memory candidate for review. It never trains an AI model, and none of it costs extra tokens. See AI Assist for how the feedback loop works.
Related
- AI Assist: the AI capabilities in full, from grounding to memory
- Automations: the first-message trigger and AI reply action
- Knowledge Base: the articles AI answers ground in
- How Billing Works: plans and AI token budgets