AI assistant
AI assistant
Modes, earning trust, reviews, info gathering, guidance and budget.
What it does#
- Drafts answers to tickets from your knowledge sources and solutions your staff approved.
- Asks players for missing details (proof, order IDs, times) before staff take over.
- Sorts tickets into the right category and flags urgency.
- Suggests next steps to staff, like a ban when a report has proof. A person always runs them.
It never bans, refunds, promises outcomes or asks for passwords or payment details. Those always go to your team.
Turning it on#
The assistant is off by default. Turn it on in Settings → AI assistant or /config area: AI assistant. If that step says the AI is not available, the bot host has not added an AI key yet; everything else works without it.
It answers under the name in AI persona name (default “Support Assistant”), set under branding.
Allow, Ask or Block#
You choose per category what the AI may do:
| Access | Meaning |
|---|---|
| Block | The AI is never used for answers in this category. |
| Ask | The AI drafts answers, staff decide whether each one is sent. (Default.) |
| Allow | The AI may answer on its own, but only once it has proven itself (see below). Until then it behaves like Ask. |
How it earns trust#
Your Allow/Ask/Block choice is the limit. What the AI actually does in a category depends on how much it has learned there:
- 1
Learning (practice answers)
Until a category has 25 approved solutions (the activation floor), the AI only practises. For each new ticket it posts a private practice answer in the staff channel. Staff rate it Good answer or Needs work; the player sees nothing. When the floor is reached, admins get an “AI unlocked” message.
- 2
Review (staff approve every reply)
With Ask, or Allow while it is still proving itself, each draft goes to staff as a review card before the player sees it.
- 3
Auto (answers on its own)
With Allow, after 20 reviewed drafts with an approval rate of at least 80%, the AI answers on its own. If the rate drops, it goes back to review automatically.
/status and the dashboard overview show each category’s current mode, its approved solutions and its approval rate. The numbers are adjustable: Solutions needed (0–1000) and Approval rate for Allow (0–100%).
Reviewing answers#
Review cards appear in the staff channel and ping the ticket’s support roles. Anyone with the Support tier can handle them:
- Send: posts the answer to the player as-is.
- Edit & send: fix it first, then send.
- Don’t send: the ticket waits for staff instead.
Every decision counts toward the category’s approval rate. A card nobody handles within 60 minutes (the review timeout, 5–1440 minutes) is handed to staff the normal way, so no player waits forever on a draft.
Checks before an auto reply#
Even in auto mode, every answer is checked first. If any check fails, it goes to review instead, and the card says why:
- It must cite at least one knowledge source.
- The best knowledge match must be strong enough.
- The assistant must be confident, and a second check of the reply must agree.
- If the category was picked by the classifier, that guess must be confident too.
- Answers quoting rules, legal or payment terms always go to staff.
Tickets the classifier thinks need a human (for example an angry player or a sensitive request) skip the AI and go straight to staff.
During a conversation#
- Quiet period. The AI waits until the player has been quiet for 15 seconds (0–120) before replying, so a question typed over several messages gets one answer.
- Follow-ups. When the player replies, the AI drafts again (practice answers are only made for the opening message). If a draft was waiting for review, it is marked outdated and replaced.
- Talk to a human. AI messages have a button the player can press at any time. The AI stops replying and staff take over.
- Staff take over. As soon as a ticket is claimed, the AI steps back. With the default “first staff reply claims” setting, that happens when a staff member answers.
- Hand-off. When the AI can’t answer from what it knows, it hands the ticket to staff with a reason.
Info gathering#
Some tickets can’t be solved by an answer, only by staff, but staff need details first. A player report without proof, or a missing purchase without an order ID. For these categories, turn on Gather details first. The AI then asks for what is missing instead of trying to answer.
| Mode | When staff are notified |
|---|---|
| Off | Right away. The AI may answer, depending on access. (Default.) |
| Assist | Right away, and the AI asks for details in the meantime. When it is done, a summary is posted to the staff channel without pings. |
| Gate | Only once the details are in, or the player says they don’t have them. Saves staff from chasing basics. |
How it asks
- List what staff need under Required info, one per line (up to 10), e.g. “Proof (screenshots or video)”. Leave it empty and the AI works it out from similar past tickets.
- It asks for at most three things at a time, in a short friendly message.
- It accepts reasonable answers (“yesterday around 20:00” is a time) and never asks twice for something the player said they don’t have.
- After 3 follow-up questions (1–5) staff get the ticket anyway, with whatever was collected. A gated ticket is never left stuck.
- The first staff reply, or the player pressing Talk to a human, ends gathering.
Staff see a checklist with each item marked provided ✅, missing ❔ or not available ➖, plus a one-line summary. The AI never judges the case or promises an outcome.
Learned solutions#
With Learn from closed tickets on (the default once the AI is on), every closed ticket where someone actually helped is summarised into a short problem → resolution note, with usernames, order IDs and other personal details removed. It appears in the staff channel for approval:
- Approve saves it to the knowledge base.
- Edit & approve lets you fix the title, problem or resolution first.
- Discard throws it away.
Approving needs the Staff tier, because approved solutions shape future answers. Only approved solutions are used, and they count toward the activation floor.
To get going faster, use Learn from past tickets in the AI step of /config: it reads your last 50 closed tickets. You can also import tickets from before Tikkora.
Suggested actions#
Turn on Suggest actions for a category and, after a ticket is sorted, the AI proposes next steps to staff based on how your team handled similar tickets. For example: “ban the reported player: screenshot proof in message X”. It only proposes tools from your integrations, cites the evidence, and proposes nothing when evidence is missing. Each step has a button; a staff member with the right tier runs it.
This is staff-only, so it works even when the category’s access is Block.
Sorting tickets#
The classifier reads each new ticket and estimates its category, urgency and whether it needs a human. Staff see the result on the notification. With auto-categorise on, tickets opened without a category get one when the classifier is at least 60% sure.
When staff move a ticket to another category, that correction is remembered and shown to the classifier as an example next time, so it gets better at your server’s categories.
Guidance#
Tell the assistant about your server and how your team likes things done, in Settings → AI assistant or /config area: AI assistant → Guidance.
| Section | Default | Use it for |
|---|---|---|
| About this server | Empty | Server IP, version, store link, how whitelisting works… (up to 3000 characters) |
| Tone and style | Ours: friendly, calm, concise, replies in the player’s language | Your voice (up to 2000) |
| Ticket workflow | Ours: ask once for missing details, solve from knowledge, hand off when unsure | How your team works (up to 3000) |
| Moderation and actions policy | Ours: act on evidence only, mildest fitting action, punishments are staff decisions | Your moderation rules (up to 3000) |
Each category can add its own instructions too, e.g. “always ask for the transaction ID” (up to 1500 characters). Clear a section to go back to our default.
Budget & costs#
The assistant needs a plan, which includes AI usage every month. Every AI request is logged per ticket. Set a monthly cap in US dollars and when it is reached, new tickets simply go to staff until the next month (months run in UTC). 0 means no cap.
/costsor the dashboard’s Costs page: totals, by use (decisions, writing, search), average per ticket and the most expensive tickets./ticket cost: what one ticket cost.
All AI settings#
| Setting | Default | Range |
|---|---|---|
| AI assistant on | Off | |
| Learn from closed tickets | On | |
| Solutions needed before replying (activation floor) | 25 | 0–1000 per category |
| Approval rate for Allow | 80% | 0–100% |
| Drafts counted for the approval rate | 20 | 5–200 |
| Quiet period | 15 seconds | 0–120 |
| Follow-up questions while gathering details | 3 | 1–5 |
| Review timeout | 60 minutes | 5–1440 |
| Monthly budget | $0 (no cap) | 0 or more |
| Classifier | On | |
| Auto-categorise | On, at 60% confidence | 0–100% |