Unified Conversations Inbox

The unified inbox is the one place a human goes to talk to a contact. SMS, email, web chat, the connected social and messaging channels, call recordings and voicemails, and anything an AI agent said on the account's behalf all land in the same thread against the same contact record. The team stops asking which tab a conversation happened in, because there is only one.

The organizing idea is that the thread belongs to the contact, not to the channel. A person who fills in a form, replies to a text, then emails a week later has one history, read top to bottom. That history is also what the AI agents read before they reply and what a rep reads before they call, so the quality of the inbox is the quality of every conversation that follows it.

The second idea is that the inbox is a work queue, not an archive. Conversations are unread or read, open or closed, assigned to someone or unowned, starred, snoozed to a later time. Treated as a queue with an agreed definition of done, it tells the team exactly what is outstanding. Treated as a feed to scroll, it quietly loses people who asked a question on a Friday.

Channels in the thread

• SMS and MMS from the account's numbers, with the number used shown on each message
• Email, sent from the account's verified sending domain and threaded against the same contact
• Web chat from the widget embedded on funnels, websites and blog pages
• Connected messaging channels: Facebook Messenger, Instagram direct messages, Google Business Profile messages, WhatsApp where enabled
• Calls, voicemails and recordings, with transcripts where transcription is enabled
• AI agent turns, marked so a human can see what BO said before they reply
• Internal notes that are visible to the team and never sent to the contact

Queue management

• Unread, read, open and closed states, and the filters built on them
• Assignment to a user, reassignment, and an unassigned queue for conversations nobody owns yet
• Starring, snoozing until a chosen time, and manual closure as the definition of done
• Saved filters and views by channel, owner, tag, unread state or last activity
• Search across message bodies, contact details and tags
• Sorting by most recent activity or by longest waiting, which are very different work orders

Composing and replying

• Channel selection per reply, including replying on a different channel to the one the contact used
• Saved replies and snippets for the answers the team gives every day
• Merge fields for contact, appointment and opportunity values inside a reply
• Attachments, images and documents, subject to each channel's own limits
• Scheduled sends, so a reply written at midnight arrives at a civil hour
• Trigger links and tracked links inserted into a reply
• Drafting assistance from the AI content assistant, with the human sending the final text

Context beside the conversation

• The contact panel: fields, tags, source, owner, consent and opt-out state
• Open opportunities and their stage, editable without leaving the thread
• Upcoming and past appointments, with the booking link to hand
• Tasks and notes against the contact, and the ability to create either from the thread
• Payment and invoice history where the commerce modules are in use

Handover, routing and automation

• Which conversations an AI agent may answer, and the conditions that hand the thread to a human
• Workflow triggers on inbound message received, outbound message replied to, and conversation unread for a period
• Internal notifications and task creation when a conversation needs a person
• Business-hours rules governing automatic replies and after-hours acknowledgements
• Per-user permissions covering which conversations a team member can see and send on

1. Connect the channels before you train anyone. Verify the sending domain for email, provision and verify the numbers for SMS, install the chat widget, and connect the social and messaging accounts the client actually monitors.
2. Send one test message on every connected channel and confirm each lands in the thread against the right contact. A channel that posts to the wrong record is worse than one that is switched off.
3. Agree the definition of done with the client in plain words, then write it down: a conversation is closed when the question is answered or the next step is booked, not when someone has read it.
4. Decide the ownership model. Either everything lands unassigned and the team claims it, or a workflow assigns on arrival by source, service or territory. Pick one; a mixture means both habits fail.
5. Build the standing views the team will live in: my open conversations, unassigned, waiting longest, and one per channel if volumes justify it. Order them so the first view is the one that should be worked first.
6. Set up the after-hours behavior. Configure business hours, then decide whether an out-of-hours message gets an automatic acknowledgement, a hand to an AI agent, or simply waits until morning with a clear expectation set.
7. Write the saved replies for the ten questions the business answers every week, with merge fields and fallbacks, and name them so a new starter can find the right one under pressure.
8. Define the AI boundary explicitly if an agent is answering: what it may handle, what it must never discuss, and the exact conditions that hand the conversation to a human with a notification.
9. Add the escalation safety net: a workflow triggered by a conversation left unread beyond the agreed response time, creating a task and notifying a named person rather than the whole team.
10. Configure per-user permissions so each team member sees the conversations they are responsible for, and confirm by logging in as a restricted user rather than trusting the setting.
11. Make the contact panel useful. Put the fields the team actually needs (service interest, source, consent state, last appointment) where they are visible without scrolling.
12. Train the team on tagging and internal notes before go-live, because the notes written this month are the context the AI agents and the next rep rely on next month.
13. Rehearse the handover both ways: AI to human, and human back to automation. Confirm that a human reply pauses the chase sequence instead of racing it.
14. Check opt-out handling end to end. Send a stop message from a test handset, confirm the contact is marked, and confirm the inbox blocks further sends on that channel.
15. Agree a daily routine and name an owner for it: clear unassigned, work longest-waiting, close what is finished, snooze what is genuinely waiting on someone else.
16. Review the queue weekly for the first month. Conversations left open, snoozed indefinitely or closed without a reply are the honest signals that the routing or the staffing is wrong.

The inbox is where automation meets a person. Workflows, calendars, forms, funnels, the phone system and both AI agents all push into it; pipelines, tasks and payments sit beside it as context. In most builds it is the module the client's staff spend their day in, which makes it the one that decides whether they believe the rest of the system works.

Missed call turned into a conversation

An unanswered call triggers a text back from the phone system. The reply arrives in the inbox as a new thread with the call recording above it, BO answers the opening question from the knowledge base, and the moment the contact asks about availability the conversation is handed to the front desk with a task and a notification. The rep reads the call, the AI turns and the contact's history in one screen before replying.

Campaign reply triage

A broadcast goes out to a segment and replies land in the unassigned queue. A workflow classifies each inbound message, honors opt-outs immediately, tags the interested, assigns them to the rep who owns that territory, and leaves the rest in nurture. The team only opens threads that are genuinely live, and the longest-waiting view keeps anyone from being forgotten.

Appointment day questions

A reminder goes out the evening before and the replies (parking, paperwork, running late) arrive in the inbox with the appointment and the booking link already in the side panel. Saved replies answer the routine ones in seconds, a reschedule is handled from the same screen, and the no-show rate falls without a single extra workflow.

• Duplicate contacts split the thread. The same person with two records has two histories, and whichever one the team opens is missing half the story. Deduplication is an inbox problem before it is a database problem.
• An unverified channel fails quietly. Email without a verified sending domain, or SMS from a number that has not completed registration, can appear to send from inside the thread and never arrive.
• Each channel keeps its own rules. Message length, attachment size, and the window in which you may reply to a social or messaging channel are set by that provider, not by the platform. A reply outside the window is rejected.
• Replying on a different channel confuses people. Answering a web chat by email is technically easy and usually unwelcome. Match the channel unless the contact asked otherwise.
• Automation and humans will talk over each other. If a chase sequence is running while a rep is replying, the contact gets both. Exit the sequence on human reply, not on conversation closure.
• Closed does not mean resolved. Closing is a manual act with no validation, so a team under pressure closes threads to clear a number. Audit closed conversations with no outbound reply.
• Snooze hides work. A conversation snoozed for a month leaves the queue and nobody thinks about it again. Cap snooze lengths by convention and review snoozed threads weekly.
• Opt-out is per channel, not per person. Someone who stops your texts may still be emailable, and the inbox will let a rep send the thing the contact asked not to receive. Keep consent visible in the contact panel.
• Internal notes are not private. They are invisible to the contact but visible to every user who can see the conversation, and they persist. Write them as if the client will read them, because one day they will.
• AI turns are indistinguishable to the contact. If an agent answered badly, the business owns it. Review AI conversations in the inbox for the first weeks rather than trusting the configuration.
• Call recordings and transcripts carry consent and retention obligations. Recording is a setting; lawful recording is a decision. Confirm the client's position before enabling it, especially across jurisdictions.
• Attachments and media sit in account storage. High-volume image and document exchange accumulates, and nothing prompts a clear-out.
• Permission changes are retroactive. Restricting a user's visibility hides history they may need mid-conversation, and widening it exposes threads the client assumed were private.
• Search finds messages, not intent. Tags and notes are what make a conversation findable six months later. If nobody tags, the history is effectively write-only.