AI
Every AI feature uses credits from your shared OfficeBats wallet: the monthly allowance of the workspace owner’s plan first, then purchased credits (Enterprise teams share one pool). Failed actions cost nothing. When AI is off or credits run out, everything else keeps working.
For agents
- Draft a reply grounded in your published and internal articles and similar solved tickets, with the sources listed. You edit and send it; nothing goes out by itself.
- Summarise a long thread: what the customer needs, what was tried, what to do next.
- Change the tone: friendlier, more formal, shorter, more empathetic, simpler, or just fix the grammar.
- Translate: HelpBat detects the customer’s language. Write in yours and send it in theirs; the original stays on the ticket.
- Suggestions: macros and articles that fit the ticket.
Triage
New tickets get a category, sentiment, intent, language and a suggested priority. Triggers can use them (for example, route negative billing tickets to a senior group). Applying the suggested priority automatically is optional.
The AI agent
Turn it on per channel (chat, email, forms, portal) in Settings, AI, with a confidence threshold (75% by default). For each customer message it:
- Hands off at once if the customer asks for a person.
- Looks for published public articles that match. None: hand off.
- Answers only from those articles, saying how confident it is and whether it needs a person.
- Sends the answer if confidence is at or above the threshold and no person is needed; otherwise hands off.
- Hands off after two answers that did not solve it.
A handoff adds an internal note with the reason and the ai_handoff tag, and in chat tells the visitor a person is coming. When a visitor says the answer solved it, the ticket is solved and counts toward the AI deflection rate in reports.
Quality and knowledge gaps
QA scoring rates the agent side of a conversation on empathy, accuracy, tone, resolution and process, with coaching tips. Knowledge gaps groups recent tickets by topic and lists topics no article answers well; the AI can draft an article from those tickets, saved as a draft for review.
Review and logs
Every AI result is logged: what was asked, the answer, the credits, and the person’s decision (accepted, edited, rejected or marked wrong). Supervisors review them in AI, Activity. Customer text is always treated as data, never as instructions.