Skip to content
WhatsTIG
Get started

AI assistant

The AI assistant answers customers on WhatsApp from the business knowledge your team writes, and hands the conversation to your team whenever it is unsure. It works inside the team inbox: every answer appears in the conversation, and your agents can take over at any time.

Find it in the dashboard under AI assistant (owners and admins configure it; the menu entry appears once the team inbox is on).

Set it up

  1. Turn on the team inbox for the project (Inbox → Settings). The assistant answers inbox conversations.
  2. Write business knowledge (Knowledge tab): services, prices, timings, admission rules, policies, FAQs. Write it the way you would brief a new colleague. One topic per entry works best; inside long entries, start sections with a heading line such as Fees: or ## Transport.
  3. Test it (Test tab): ask what your customers ask. Nothing is sent on WhatsApp. Each test answer costs budget like a real one and shows its confidence, the knowledge it used, tokens and cost.
  4. Turn it on (Settings tab) and choose when it answers:
    • Always
    • Outside business hours only (uses the inbox business hours)
    • Until an agent is assigned

How it answers

  • Only facts from your knowledge and the conversation are used. It never invents prices, dates, links or promises.
  • It replies in the customer's language, briefly, in WhatsApp style.
  • Several quick messages ("Hi" … "do you have" … "a science lab?") get one answer: it waits a few seconds for follow-ups (configurable).
  • Each answer is rated by the model for confidence. Answers below your minimum confidence (default 60%) are not sent.

When it hands over to your team

The assistant stops answering and leaves the conversation to your team when:

  • the answer is not in the knowledge, or it is not confident enough;
  • the customer asks for a person, complains, or needs something only a person can do (payments, refunds, bookings);
  • your organization's monthly AI budget is used up;
  • the AI provider is failing (after retries).

It sends your handoff message once (if you wrote one, e.g. "A team member will reply shortly"), records why in the conversation timeline, and auto-assignment picks an agent if it is on. It stays quiet for the rest of that conversation; when a resolved conversation is reopened by a new message, the assistant helps again.

It also stays quiet:

  • for 12 hours after a teammate replies (inbox or phone; configurable, 0 = never pause);
  • when an agent clicks Pause here in the contact panel of a conversation (until someone clicks Let the assistant answer);
  • after 15 replies in 24 hours in one conversation (configurable) — protection against loops and abuse.

The away message is not sent while the assistant is on: its handoff message replaces it.

Budget and usage

AI usage is billed by the tokens the model reads and writes, at the platform's price per million tokens. Each organization has a monthly budget (USD, calendar month in UTC) shared by its projects:

  • Before every model call its worst-case cost is reserved; calls that do not fit are not made, so the budget is never exceeded by more than rounding.
  • After the call, the cost is computed from the token counts the provider reports and the prices at that moment, and recorded in a usage ledger.
  • When the budget is used up, the assistant hands new conversations to your team until the 1st of next month.

The Usage & budget tab shows spend against the budget, cost per day, model calls, replies, handoffs, tokens, and spend per project. Test-console answers count too.

Your budget comes from your plan or an agreement with WhatsTIG; contact support to change it.

Good knowledge

  • Be specific: "Class 1–5: 3,000 PKR per month, due by the 10th" beats "fees are affordable".
  • Include what customers really ask: timings, location, documents needed, how to pay, who to contact.
  • Say what you do not offer, so the assistant can answer "no" instead of handing over.
  • Keep it current: the assistant uses only enabled entries, immediately after you save.
  • Do not put secrets, personal data or internal notes in knowledge: it can be quoted to customers.

Privacy

Recent messages of the conversation (text only), the contact's WhatsApp name and the relevant knowledge are sent to the platform's AI provider to generate an answer. Phone numbers are not sent. Projects that do not retain message content cannot use the assistant (there is nothing to read).