CDA customers can reach the same assistant wherever they already are. Every channel ends up at one agent, Ellie, on ElevenLabs Agents, with the same instructions, the same knowledge of CDA and the same memory of the customer. When a person is needed, CDA staff take over in an Aida room, or check Ellie's email answers first.

Website chat / voice / avatar ─────┐
Phone (Twilio) ─────────────────────┤
Telegram ───────────────────────────┤
Email ─► Gmail ─► web app ──────────┤──►  Ellie (ElevenLabs agent)  ──►  answer on the same channel
Instagram / Messenger ─► web app ───┤         │  knowledge base (RAG) · customer_lookup tool
Alexa ─► web app ───────────────────┤         │
Intercom / hosted page / widget ────┘         ▼
                                   web app (Next.js on Vercel)  ─  Supabase database
                                   customer memory · Aida rooms · admin · learning · mood

The demo has three pages:

• Customer site (/, site password): chat, voice, avatar, the Aida tab, the customer account and links to every channel.
• Staff page (/admin, staff password): Aida rooms, customers, mood, call list, knowledge and the email switch.
• This documentation (/docs, open to everyone).

Ellie is CDA's virtual assistant. She answers questions about CDA appliances, warranties, spare parts, repairs and where to buy, collects the details CDA needs for a repair or a complaint, and passes the conversation to a person when needed.

Rules she follows

• She answers only from CDA's knowledge; when she does not know, she says so and points to the right place. Those questions are collected for staff (see Ellie learns).
• Channel rules: short spoken answers on the phone, Alexa and the avatar; plain text on Instagram; one plain-text reply by email; she ignores robots and newsletters.
• Safety first: a smell of gas or any danger gets safety advice and the emergency number straight away.
• She never asks for card details or passwords, and never claims something is booked when it is not.
• When a customer is upset she apologises once, slows down, and offers a person (see Customer mood).
• She never asks customers to identify themselves: if she recognises them, she uses what she knows (see Customer memory).

The customer site (password protected for the demo) has four tabs, all talking to the same Ellie:

• Customer account: sign up with an email, then link Telegram, Instagram, Messenger or another email with a short code, and add a phone number, so Ellie recognises the customer everywhere.
• Email me this conversation: the transcript of a chat, voice or avatar conversation by email.
• Channel links: buttons that open email, the phone line, Telegram, Instagram and Messenger.
• Intercom bubble: a second way to chat, answered by Ellie through Intercom.

A real UK mobile number, bought on Twilio and connected natively to ElevenLabs. Customers call it and talk to Ellie; she recognises them by the number saved in their account and greets them by name. She hangs up politely at the end and, when she calls out and reaches voicemail, leaves a short message.

Ellie also calls customers herself, from the staff call list, with a greeting that says why she is calling.

1. A customer emails the CDA demo mailbox (Gmail)
2. Google notifies the web app at once (Gmail watch + Pub/Sub)
3. Robots, codes, alerts and newsletters are skipped (no cost)
4. Claude rates the email's mood (see Customer mood)
5. The email goes to Ellie through an ElevenLabs Custom Channel
6. Ellie's answer is sent in the customer's thread (auto)
   or saved as a Gmail draft for staff (draft); an upset customer always gets a draft
7. The email is labelled in Gmail: Replied · Draft ready · Skipped · Failed · Upset customer

• Auto or draft: staff switch it on /admin → Email. In draft mode staff open the email, check Ellie's draft and press Send.
• What staff change in a draft is compared with what Ellie wrote, and a real correction becomes a lesson (see Ellie learns).
• Ellie reads up to 6,000 characters of an email; she cannot open attachments. Mail older than 24 hours is never answered.
• The daily job renews Gmail's watch, so nothing is missed.

On Instagram and Messenger one person's messages stay in one conversation for 10 minutes, and “typing…” shows while Ellie writes. Photos and files are not passed on: Ellie asks the customer to type the details.

"Alexa, ask cda assistant why my oven shows F3"
   → Amazon → web app (checks Amazon's signature)
   → Ellie through the "CDA Alexa" Custom Channel, marked [Alexa] so she answers in 1–3 short sentences
   → Alexa reads the answer out and the customer can ask a follow-up

• The skill “CDA Assistant” is in development mode (the developer's own Amazon account); publishing it needs Amazon's certification.
• A question starts with a question word (why, how, what, where, can, my, it's…), as Alexa requires for free speech.
• Alexa speaks in its own voice, in English. The same Alexa account is remembered between conversations.

• Intercom: the chat bubble on the customer page. Intercom sends the conversation to Ellie through ElevenLabs' native Intercom integration, and staff see the same conversations in Intercom's inbox.
• Hosted page and QR code: ElevenLabs' own talk-to page for Ellie (voice and text), and a QR code that opens it.
• Widget: ElevenLabs' chat bubble, which any website can add with two lines of HTML.
• Not built yet: Slack (waiting for a workspace) and WhatsApp (parked).

Ellie recognises the same person on every channel and remembers what they asked before. She never asks anyone to identify themselves: someone she cannot place is simply helped.

Start of every conversation → customer_lookup → known? greet by name, use the last notes and their appliances
End of every conversation   → post-call webhook  → one short note, the appliance model if mentioned, the mood

• Linking a channel: in their account the customer gets a short code (for example CDA-4F2K9M) and sends it from Telegram, Instagram, Messenger or another email address. One code works once, for 30 minutes.
• Appliances: when a model is known (a receipt, a photo of the rating plate, or what they said), it is saved, so Ellie never asks for the model number twice.
• What is stored: which conversation belongs to whom, one short note per conversation, their appliances and the mood scores. Full transcripts stay in ElevenLabs.

A live call between CDA staff and a customer. Everyone can talk or type, the call is transcribed live, and Aida, a second agent, drafts a reply to each thing the customer says. Only staff see the drafts; they approve, edit or decline each one, and an approved draft is sent to the customer in the chat.

• Voice and chat travel over LiveKit. Each browser transcribes only its own microphone (ElevenLabs Scribe), so every line is known to come from the person who said it.
• Aida runs in one staff browser (the first employee to join; the next takes over if they leave).
• A signed-in customer is recognised: Aida gets what CDA already knows about them, and the call is added to their memory.
• Live mood: each customer line gets a mood dot and the staff see a mood meter; when the customer is frustrated Aida's next draft opens with an apology (see Customer mood).
• Closing a room is final: its history stays readable and can be emailed. Rooms expire after 4 hours.
• What staff change in Aida's drafts is compared with what she wrote, and a real correction becomes a lesson (see Ellie learns).

The staff side, at /admin, behind its own staff password (the customer site password does not open it).

The Claude insights are written only when a staff member clicks, and nothing is stored.

Staff type phone numbers, an optional name and instructions for Ellie for each call (for example “their dishwasher warranty is not registered yet: explain the lifetime parts warranty and how to register”), then press Start calling.

• Ellie calls one number at a time from the demo line. Her greeting says why she is calling: “Hello Helmi, this is Ellie, the virtual assistant from CDA. I'm calling about your dishwasher warranty. Have you got a moment?”
• No answer, busy or voicemail: she tries again a minute later, up to 3 times, then moves to the next number. On voicemail she leaves a short message.
• The page shows each call's status live (calling, try 1 of 3, reached ✓ with a summary, not reached). Stop lets the current call finish.
• She learns the instructions through customer_lookup at the start of the call, so no other channel is affected.

Ellie gets better from real conversations, but nothing reaches her without a staff member approving it.

Questions she could not answer

• After every conversation, ElevenLabs' analysis lists the questions Ellie could not answer from CDA's knowledge. They appear on /admin → 📚 Knowledge with the channel.
• ✨ Group and suggest answers: Claude merges questions that ask the same thing and suggests wording, marking [check: …] wherever a CDA fact is needed; it never invents one.
• Approve and teach Ellie: the answer goes into the “CDA approved FAQ”, which is always in Ellie's (and Aida's) context, so it works from her very next conversation.

Feedback and corrections

A real correction (a changed fact, not just a new greeting) becomes a card; ✨ Make it a general answer turns it into a question and answer for everyone, without that customer's details. A “right first time” score counts how many of Ellie's and Aida's drafts were sent unchanged.

How customers feel, on every channel, measured and acted on. Five parts work together:

The 😊 Mood tab on /admin

• The share of positive, neutral and negative conversations and the average frustration, for 7 or 30 days.
• Mood by channel, and conversations per day.
• The upset emails (sender, subject, why, Open in Gmail) and the frustrated lines in Aida rooms, listed one by one.
• Unhappy conversations: frustration of 60% or more, a very negative moment, or a promised follow-up. Each shows the customer, the channel, the message where it turned, a small mood curve and a Mark followed up button.
• Import past conversations fills in the last 30 days; opening the tab and a daily job bring in new ones (reading from ElevenLabs is free, Claude's rating costs a fraction of a cent).

On a customer's page, each conversation gets a mood dot, and Claude's customer insight bases its mood on these measured scores. Moods are for staff only: a customer never sees one.

• Two separate passwords: the site password for the customer site, and the staff password for /admin. Being on the site never makes anyone staff.
• Everything that calls the web app from outside proves who it is: ElevenLabs' tools send a secret, its webhooks are signed, Google, Meta and Amazon send their own proof, and the daily job has its own secret.
• Keys and passwords live only in Vercel and the servers' settings, never in the code (the repository is public).
• The database is locked down (row level security) and only the server reads it. No message text is stored for memory: short notes and scores only; transcripts stay in ElevenLabs.
• The demo uses fake customers only. For real customers, the use of Anthropic (Claude) for insights and mood would belong in the privacy notice.