Agent Assist

1. 1

   Listen to the live thread

   Each inbound message that reaches the inbox through the messages webhook is also passed to the assist service, together with the recent thread, the customer's profile and any open tickets.

2. 2

   Summarise on pickup

   When an agent opens or is assigned a conversation — especially after a bot handoff — assist writes a three- or four-line summary: what the customer wants, what has been tried and what is still open.

3. 3

   Retrieve the facts

   It looks up the order, delivery, payment or appointment record and the matching knowledge-base article, and shows them as cards with their source and date, so the agent can check rather than trust.

4. 4

   Draft a suggested reply

   A draft appears in the composer in the customer's language and the brand's tone. If the 24-hour customer service window has closed, a good tool proposes a suitable approved template instead of free text.

5. 5

   The agent decides and sends

   The agent accepts, edits or discards the draft. Whether each suggestion was used as-is, changed or rejected is logged — it is the main signal for improving prompts and content.

6. 6

   Wrap up

   After the chat, assist proposes disposition tags, a CRM note and any follow-up task, which the agent confirms or corrects.

Who sends

Always the human agent

Typical features

Suggested replies, summaries, knowledge cards, next-best actions, wrap-up notes

Inputs

Live thread, order/CRM data, knowledge base

Window rule

Free-form drafts inside the 24-hour window; templates outside it

Sent replies (India)

Billed as service messages, ₹0.145 each since 1 October 2026

Key quality metric

Suggestions used as-is / edited / rejected

“Most of an agent's time on WhatsApp is not typing — it is reading back through a thread, switching tabs to find an order and searching for the right policy wording. Agent assist removes that overhead, so first response time and handling time fall without letting an unsupervised model speak for the brand. It also levels up new hires: a suggested answer grounded in the current refund policy is live training, and consistent wording keeps the team's promises consistent.

• Agents rubber-stamping drafts

  When suggestions are usually right, agents stop reading them. Show the source beside every fact, highlight numbers, dates and amounts the model inserted, and review a sample of accepted-unedited replies each week.

• Suggesting from stale content

  Assist is only as good as the policies and product data it retrieves. Version the knowledge base, retire old articles and show each card's last-updated date.

• Leaking personal data into prompts

  Threads contain phone numbers, addresses and order details. Mask what the model does not need, keep logs within your retention policy and check where your AI provider processes data.

• Measuring only speed

  Faster handling with more reopened tickets is not a win. Track edit rate, first-contact resolution and CSAT alongside handling time.

• Drafting free text after the window closes

  A perfect free-form draft fails to send once 24 hours have passed since the customer's last message. Make the tool aware of window status and suggest the right template category instead.

• Start with summaries on handoff — the lowest-risk, most immediately useful feature.
• Put a one-tap "bad suggestion" control in the composer and read what agents flag.
• Draft in the customer's language, but keep the agent's interface in theirs.
• Feed agents' edits back as examples when you improve prompts and articles.
• Show the window countdown right next to the suggested reply.