AI Copilot

When an agent opens a ticket, the AI Copilot panel loads a set of suggested replies ranked by predicted helpfulness. Suggestions draw from your historical resolved tickets, your configured response templates, and the Keme knowledge base for common gaming support scenarios.

Agents can insert a suggestion directly into the reply box with one click, or use it as a starting point and edit before sending. Accepted and edited suggestions both feed back into the model as implicit training signal.

Suggestions are generated on-demand and cached for 60 seconds. If the player adds a new message while the agent is drafting, the panel refreshes automatically with updated context.

Each AI suggestion displays a confidence score from 0–100. A score above 80 indicates the model has high certainty the suggestion is relevant. Scores below 50 are shown with a caution indicator and should be reviewed carefully before sending.

Confidence is influenced by: similarity to resolved tickets, availability of matching templates, recency of training data, and ticket complexity (single-issue tickets score higher than multi-issue threads).

On the Growth plan, agents can filter the suggestion panel to show only high-confidence results (80+). Enterprise customers can configure the minimum confidence threshold globally from Settings → AI.

Pattern detection runs continuously across all open tickets in your workspace. When a cluster of tickets shares a common root cause, the AI surfaces a Pattern Alert in the Copilot sidebar and the Home dashboard.

Pattern alerts include: the identified topic, a count of affected tickets, the first occurrence timestamp, and a suggested bulk reply you can send to all affected players with one click.

Pattern data is also available via the API so you can pipe it to your internal tooling or incident management system. See the API Reference for the /patterns endpoint documentation.

Risk signals identify players who are at elevated risk of churning based on their support interaction history, sentiment trends, and in-game spending patterns (when connected via the data integration).

A risk signal appears as a coloured badge on the ticket header: amber for moderate risk, red for high risk. Clicking the badge opens a player context panel showing their history, spend tier, and the factors that triggered the signal.

Risk signals can trigger automations — for example, automatically routing a high-risk player's ticket to a senior agent or applying a VIP priority tag. Configure risk-based automations under Automations → New Rule → Trigger: Risk Signal.

Keme's AI is pre-trained on a broad gaming support corpus. You improve it for your specific titles by consistently resolving tickets with quality replies and using the thumbs-up / thumbs-down feedback on AI suggestions.

On Enterprise plans, you can upload a training document set (PDF, plain text, or Markdown) under Settings → AI → Knowledge Base. Uploaded documents are chunked, embedded, and searchable by the suggestion engine within 15 minutes.

Custom model fine-tuning (training a model on your ticket history only) is available as an Enterprise add-on. Contact your Customer Success Manager to enable it. Fine-tuned models take 24–48 hours to deploy after the training job completes.

Every AI action — suggestion generated, suggestion accepted, suggestion edited, pattern detected, risk signal triggered — is recorded in the AI Audit Log. Access it under Analytics → AI Audit.

The audit log shows the originating ticket, the suggestion or signal content, the agent who acted on it, and the outcome (reply sent, ticket resolved, etc.). Logs are retained for 12 months.

Enterprise customers can export the audit log via the API for compliance reporting. The export endpoint supports date-range filtering and outputs JSONL format for easy ingestion into data warehouses.