10-20 → 50-100tickets an hour, per agent
4 channelsone queue
PII-safeevery draft
Home / Real systems / Customer Support Copilot
Case study · Customer Support Copilot
One calm, on-brand answer across every channel.
Email, Microsoft Teams, Facebook and Instagram pull into one queue. For each message the copilot investigates your systems, retrieves the right answer from your knowledge base and drafts an on-brand reply - cited, with a confidence score. Your agent reviews and sends. A person approves every reply by default; you choose which low-risk categories may send on their own.
10-20 → 50-100 tickets/hour4 channelsone queueagent approves
The problem
Four inboxes, one team, quality that swings with the shift
Customers message you everywhere - email, Microsoft Teams, Facebook, Instagram - and your team hops between inboxes, re-writing the same answers and hoping nothing slips through. Quality swings with whoever is on shift, and response times stretch when it is busy.
Investigating the issue takes time before a reply can be written. Your team checks documents, past tickets, customer records and live system logs to understand what happened. That's the work we handed to the copilot.
How it works
From every channel to one approval queue
Support agents spend most of their time investigating, not replying. The copilot does the digging - then hands your team a finished, cited draft to approve.
1
Read & triage
Every message - email, Microsoft Teams, Facebook or Instagram - lands in one queue, classified by intent, account, urgency, sentiment and language. Spam and duplicates are filtered; threads are understood so it never re-answers.
2
Investigate your systems
It searches documents and past tickets, checks Sentry logs and audit trails, and reads the customer record in your database and CRM. It also requests current account details from your connected services. Access rules limit each lookup to that customer’s records.
3
Decide & draft
It recommends a response, drafts a reply in your voice and cites its sources, with a confidence score. It also prepares any refund, resend, ticket update or escalation.
4
Human approves & sends
The draft and its evidence land in your agent's queue. One click to send, edit or reject. Nothing reaches the customer or writes to a system until a person says go.
Connected to your real stack
It answers from your systems, not a guess
✓
Knowledge / RAG
Azure AI Search, Onyx and Amazon Bedrock KBs - docs, past tickets, runbooks and policies, retrieved with search + rerank so the answer is genuinely relevant.
✓
Production signals
Sentry errors, application & audit logs, incident and status data - so it knows what actually broke.
✓
Systems of record
Your DB and CRM - read for context, and it proposes writes for a human to approve.
✓
Microservices
Live state via your internal APIs - orders, billing, entitlements, shipping and more.
Message #4821 · proposed reply
✓ KB · "Resetting a device token"cited
✓ Sentry · issue TM-2261 linkedresolved
✓ Order #A-5567 · API lookupshipped
✓ CRM · plan Business, since 2024read
! Refund $49 · staged actiongated
✅ Draft ready · 96% confidence · awaiting approval
Security & control
Built for a team that has to be right
The whole point is control: the copilot gathers everything and proposes the answer, and a person signs off on the send.
🔒
Row-level security everywhere
The copilot only ever sees what that customer's conversation is entitled to. RLS is enforced at the source, not bolted on afterwards.
✋
Human-in-the-loop by design
No auto-send, no silent writes. Autonomy is configurable per action type - start fully approved, graduate low-risk categories like FAQ answers to auto-send later.
📎
Grounded, not guessed
Answers are cited to your sources and carry a confidence score. If it can't ground an answer, it escalates instead of hallucinating.
🛡️
PII & content screening
Every draft is screened for private data and unsafe content before a human even sees it - nothing sensitive is surfaced or logged in the clear.
🧾
Full audit trail
Every draft records what it read, why it decided and who approved - so "why did we tell them that?" is always answerable.
🇦🇺
Onshore, your tenancy
Runs in Australia and in your cloud. Your data and knowledge base stay yours.
What changed
The outcomes, in plain numbers
- One agent moves 50-100 tickets an hour instead of 10-20 - the investigation is finished before they open the ticket
- The reply is already drafted when the agent opens the ticket
- Four channels, one queue - nothing falls between them
- New staff sound like your best staff from day one
- Every draft is cited and carries a confidence score
- Every draft screened for private data before a human sees it
- It gets sharper over time - approved and edited replies feed back into retrieval and tone
Built from real systems
This isn't a concept
The Support Copilot is assembled from components we already run in production: reading an inbound message, checking the systems behind it and returning a structured reply.
📧
Invoice Email Analyzer
Reads inbound invoice emails, extracts the data and matches them to bills automatically.
📨
RatesMailParser
Parses inbound rate emails into clean, structured data ready for downstream systems.
📚
Onyx
Self-hosted RAG knowledge system - the retrieval layer that grounds every answer in your sources.
🎙️
voice-stt
Live speech pipeline - the same real-time, tool-calling backbone the copilot runs on.
FAQ
Common questions
Does it send replies on its own?
No - human-in-the-loop by default. Every draft lands in your agent's queue to send, edit or reject. You choose which (if any) low-risk categories graduate to auto-send later.
Which channels does it pull into one queue?
Email, Microsoft Teams, Facebook and Instagram today, with the same pattern extending to help desks like Zendesk and Freshdesk. Every message lands in one queue so nothing falls between inboxes.
What if it doesn't know the answer?
It escalates with everything it found attached, instead of guessing. Answers are grounded in your sources, cited, and carry a confidence score, so a person can verify at a glance.
Which knowledge base and systems does it work with?
Azure AI Search, Onyx or Bedrock knowledge bases; and your own DB, CRM and internal APIs - every lookup scoped by row-level security to that customer.
How do you keep customer data safe?
Row-level security scopes every lookup to the customer, every draft is screened for PII and unsafe content, it runs in your tenancy and onshore in Australia, and every action is written to a full audit log.
What does it cost?
Every build is scoped and quoted up front. Start with the AI Opportunity Audit - a fixed A$2,000, and we'll hand you a costed plan that's yours to keep. You only pay once you approve a build. The audit fee is credited in full against your first build.
Why it matters
Could this be your workflow?
If support quality depends on who is on shift, a copilot grounded in your own knowledge base makes every shift your best one.
Best for · support teams, e-commerce, high-volume services
Have a similar workflow? Let's map it.
We start with the AI Opportunity Audit - a fixed A$2,000: we map where the hours and the risk actually are, then show you the one or two systems worth building first. Here, one agent moves 50-100 tickets an hour instead of 10-20.
Book a 60-minute meetingOr reach us right here, right now - talk, type or call, whichever suits you:
ChatType here - a founder replies
Voice AITalk to our assistant
Emailsergey@twinmind.au
Microsoft TeamsChat with Sergii Diatchenko · sergey@twinmind.au
Live video & voice calls run 9am-5pm Brisbane; chat, the AI assistant and email are open any time. Calls are transcribed live into your chat, so everything you discuss stays right here in your messages.