What you can read is decided by your account, not by anything the
browser sends.
Answer a message
Paste what the customer wrote. You get the answer from your own
documents with a citation you can check, a reply drafted in their language using
your organisation's approved wording, and actions you can take — nothing is sent
without you.
Leave both on Auto and the reply is
shaped from the channel plus what the message actually is — a complaint is
acknowledged, an SMS stays short, a multi-part question gets the conditions
spelled out. You will be shown the reasoning and can override it.
Support work is a thread. Reopen one to
see what was said and continue it.
loading...
Operations
The parts of the desk that run without anyone watching, and the
one control that had no counter-control.
Follow-ups, offer expiry, presence
and SLA warnings. Exactly one instance does this work however many processes
are running — it holds a database lock — so a follow-up is never sent twice.
Before this existed the follow-up timer rode the visitor's own polling,
which meant it fired only while they still had the page open.
loading...
Ten wrong passwords locks an
account for fifteen minutes. That is worth having, but until now nobody
could undo it — the person was simply unreachable. Clearing a throttle is
recorded in the audit log with who did it and for whom.
loading...
Keep the Library in step with where
the documents actually live. Content inherits the connector's scopes — the
default is internal, because a mailbox synced into the public space would be
readable from the chat widget.
loading...
Provisioning tokens let Okta, Entra
or Google create and — the part that matters — deactivate accounts
here. A leaver keeping a session into a desk full of customer conversations
is the window this closes. A token is shown once and stored only as a hash.
loading...
Live Ops Board
What is happening on the desk right now. Four panels for the
four questions actually being asked: who is waiting and for how long, what is
mine to act on, who else is here, and whether today is unusual. The warning
and breach thresholds are yours to set in Settings — a number compiled into a
dashboard is a number nobody can argue with.
Longest wait first. Time is counted
only while the desk owes a reply — a conversation waiting on the customer
does not accrue against you.
loading...
Presence expires on a stale
heartbeat, so an agent who closed their laptop reads offline rather than
staying online forever.
loading...
Inbox
Live conversations. The assistant answers from public knowledge
only; it hands over when it cannot answer, when the visitor is unhappy, or when
they ask for a person. You can take a conversation, reply yourself, and hand it
back. Open the visitor widget in another tab
to try it.
Oldest waiting first. Every row says
why it escalated.
loading...
Conversation
Select a conversation.
Select a conversation from the queue.
This conversation
Why you have this
Select a conversation.
How it is going
Time on this conversation
What was asked, and read
Sources it read
This conversation
What happened, and when
0
Tickets
People who could not wait and left a message instead. They are
still in the Inbox — this is the same queue, in the same order, narrowed to the
ones who are owed a reply by email. Each carries the reference the customer was
given to quote back.
Oldest first. Replying or resolving
the conversation marks the ticket answered.
loading...
Opening one takes you to the conversation in the Inbox.
There is one transcript view, deliberately: a second copy of it is a
second thing to keep in step with the first.
Settings
What this desk does, and who decides. Values shown with their
source: your setting, a deployment default, or the built-in default.
Some settings are deliberately not editable — they are tied to the database
schema or to a security guarantee, and the reason is given for each.
loading...
A settings change is often the
proximate cause of an incident, so the history outlives the next change.
Quality
How the assistant and the team are actually doing. The number
worth watching is how much agents change an AI draft before sending it —
"assist used" counts a draft accepted untouched and one rewritten from scratch
the same way, and only the edit distance tells them apart.
loading...
Where a human sent something
materially different from the assistant. A real question, what the bot said,
and what a person said instead — the most honest signal this system produces.
loading...
Library
Your organisation's documents. Everything answered at the Desk is
grounded in these and nothing else. Each document carries the permission needed
to read it.
PDF, DOCX, TXT or Markdown. A PDF page becomes one chunk, so a
citation points at a real page. Scans with no text layer are refused rather
than indexed empty.
Added text is chunked, embedded and indexed immediately.
loading...
Glossary
Words your organisation has already decided about. A
locked term must appear exactly as written — if the model renders it any
other way the reply is withheld rather than sent with the wrong wording.
Your documents already contain the
vocabulary. SenseBridge can read them and propose terms — but it cannot lock
one. Locking is you saying "we have already had this argument", and that is
not a machine's call to make.
loading...
Usage
What this has cost, and the ceiling. When the ceiling is reached
requests are refused before they reach a provider — you are not billed and then
told.
Now
live · refreshes while this page is open
Over the window
What the assistant is doing
every assistant reply in the window, by how well it was supported
Where the work is going
Sorted by how much they look like they need a person. Each
reason is shown on the row.
What the assistant searched for, what it read, and how well
the answer was supported.
