Customer support
A support chatbot that can actually look things up
Most support bots fail on the second message. That's the one where the answer depends on this customer's account, not on your help centre. Mopus runs the conversation, searches your own docs, calls your systems for the specifics, and escalates when it's out of its depth.
A bot that can only quote a help centre deflects the easy half of your volume and irritates people with the rest. The hard part was never the writing. It's giving the model safe, scoped access to the order, the subscription and the account, and knowing when to stop and fetch a human.
of deployed AI customer service agents are rolled back or shut down.
Sinch, 2026
The flow
What runs when the message arrives.
Every step below is a real node type on the canvas, and the log underneath is the execution it produces. A branch the run didn't take stays dark, same as it would in the console.
- 01agent.invokesupport-chat · ses_4d10run-sync
- 02messages.loadmanaged · 9 turns14ms
- 03text.analyzesentiment · neutral 0.78300 chars
- 04llm.call2 tools offered398ms
- 05bucket.searchhelp-centre · semantic9 hits · 26ms
- 06tool.httpGET /v1/accounts/8812200 · 104ms
- 07run_jscompose answer6ms
- 08agent.returnanswered · no handoff1.08s
What it uses
What this leans on.
Your help content, searchable by meaning
Import your help centre or policy documents into a semantic bucket. PDF, DOCX, Markdown, HTML, or a crawl of your own site, up to 25 MB per file. The agent searches it mid-conversation.
Conversation that persists
Managed history keeps the thread under a session id and hands the model the recent turns. Or leave your own system as the source of truth and let Mopus map its response into turns.
Tone read before the model sees it
A Text Analysis step returns sentiment, entities and key phrases in one call, so an angry message can route straight to a person instead of getting three cheerful replies first.
Escalation that resumes
A `human_input` step parks the run and puts an item in the Inbox. When someone answers, the run continues from that step. It doesn't count as a second execution.
What it meters
One execution per conversation turn, plus one semantic search, whatever bucket reads the run makes, and the characters it analyses. A typical support message bills at the 300-character floor. A run that resumes after a human answers is still the same execution.
More
Other things people build.
Ecommerce
Order status
Look up the order, check the carrier, work out whether it's actually late, and either answer or raise a reship for approval.
ReadMoney movement
Refunds and returns
Read your refund policy, check the order against it, settle the small ones automatically, and route the rest to a human with the context already attached.
ReadRevenue
Cancellation and retention
Screen the message, read the account, and make a retention offer worth making, with a human approving anything that costs real money.
Read