What runs and what we build
Part of this already runs on our platform, and part is written by our engineers for your line.
- Already runs: the shared team inbox, internal notes, the Needs reply filter, the 24-hour window banner, pause-on-reply, and the Android app with push notifications for your team.
- We build: the conversation itself (menus, questions, when to hand over), the prompts Claude works from, the facts it answers with, and any calls to your own systems. These are written as code on our platform, so there's no drag-and-drop bot builder for you to maintain.
We don't offer a knowledge base, document upload or website training. When you want the bot to answer from your own knowledge, we build that: the approved answers live in the build, or the bot looks them up in your system through its API. That keeps answers to what you've signed off.
How the conversation works
- The customer messages your numberText, a button tap or a voice note. Voice notes are transcribed first, with South African languages set as candidates, so the rest of the flow works on text.
- Claude sorts the messageA single Claude call classifies what the customer wants (a fault, a billing question, opening hours, something else) and pulls out details such as an account number, address or order number.
- The bot asks for what's missingIf a detail is needed, it asks for it, one question at a time, instead of sending a form.
- Routine questions get answeredClaude drafts the reply from the facts we've given it, or from a live lookup in your system. If the answer isn't in those facts, it doesn't guess.
- Everything else is handed overThe bot tells the customer a person will reply, sums up what it has gathered in the chat, and leaves the conversation for your team in the shared inbox, where it shows as unread until someone opens it. There's no separate alert at the moment of handover, so if you need one we build it, for example a webhook into a tool your team already uses. We can also add a "Talk to a person" button to the menu.
- A person replies and automation pausesWhen someone on your team replies, automation stops for that chat for 2 hours by default, or for a set time, or until it's switched back on. While it's paused, new messages from that customer push to your team's Android app.
What your team sees
One shared inbox for everyone, with the bot's messages and the customer's in the same thread, so nobody asks the customer to repeat themselves.
- TMBot answeredThandi M.What time do you close on Saturday?
- JSUnreadJohan S.Fault, account 40-118-223, line down
- NKTranscriptNomsa K.Voice note, 0:38: "my bill is higher than…"
The inbox has filters for All, Unread, Needs reply (customers who messaged in the last hour) and Archived, transcripts shown under voice notes the bot has transcribed, internal notes your customers never see, media, template replies, and a banner showing whether the 24-hour window is still open. There's no assignment to named agents, routing rules, queues, SLAs or canned replies. If you need those, see using a helpdesk instead.
Four ways businesses use it
These are common patterns in the market, not Euphony clients. Details are illustrative.
1. Utility or telco: bot first, warm handover
A fibre provider, ISP or utility gets the same handful of questions all day: is there an outage, why is my bill higher, how do I move my line. The bot checks the customer's account or area against your system, answers what it can, and captures fault reports with the account number and a description. Anything it can't settle goes to the support team with the details already in the chat, and the team gets push notifications on the Android app while they're handling it.
2. After-hours reply with a top-questions menu
A business that closes at 17:00 still gets messages at 21:00. Outside your hours, the bot replies with your opening times and a short menu of the questions people ask most. Claude matches typed questions to the answers you've approved. Everything else stays unread in the shared inbox for the morning, so your team starts the day with a list instead of a scroll. If the customer messaged within the last 24 hours, your morning reply is still a normal service message.
3. Specialist advice line that asks follow-up questions
Good advice needs context. A pet-food brand's advice line, for example, asks for the breed, age and current diet one question at a time before Claude drafts guidance, and only from the brand's own feeding guidelines written into the build. Anything that sounds like a health problem is never answered by the bot. It goes straight to a specialist with the details gathered so far. The same pattern suits nutrition, skincare or product-advice lines.
4. Keep your in-house bot and add WhatsApp
If you already run a chatbot on your website or app, you don't have to replace it. We relay WhatsApp messages to your bot's API and send its replies back to the customer, or your bot calls our platform when it has something to send. When your bot can't help, the chat lands in the shared inbox. Your bot needs to answer within a set time (10 seconds by default), so we check that during scoping.
Systems connected
Claude and transcription already run on our platform and are used inside the conversation we build for you. See how we use Claude. Lookups in your own software go through its API (your own software), and outside automations can be triggered through n8n webhooks.
What Meta charges for it
Almost everything in this use case is a service message: a reply sent while the customer's 24-hour window is open. That covers the bot's replies and your team's replies alike.
| Message | Meta category | Meta's rate |
|---|---|---|
| Bot or team reply within 24 hours of the customer's last message | Service | US$0.0095, after the first 1,000 a month per number, which are free |
| Customer messages to you | Inbound | Free |
| Following up after the 24-hour window has closed (for example "we've fixed your line") | Utility template (if Meta approves it as utility) | US$0.0095 |
On top of Meta's charge there's the Claude model cost for each message the AI handles. Sorting with Haiku is cheaper than drafting a long answer with Sonnet. We size this with you during scoping.
For comparison, Meta's own Meta Business Agent is charged per token instead: US$2.00 per 1 million tokens, which Meta says typically works out to about 4 to 5 US cents a message. Our replies aren't sent through the Meta Business Agent, so they're charged as service messages. The WhatsApp AI agent cost guide has worked examples.
Source: Meta, Pricing on the WhatsApp Business Platform (updated 30 Sep 2026, checked 26 Oct 2026). See also our pricing page.
One more Meta rule: since 15 January 2026, Meta restricts general-purpose AI assistants on the platform. Our reading is that a support assistant for your own business is a different thing, and that's what we build. See Meta's AI Providers policy.
If you need agent routing
Our inbox suits a team that works one queue together. If you need tickets assigned to named agents, team queues or SLA reports, we can put the bot in front and pass the conversations it can't handle into Zendesk or Freshdesk as tickets, so the helpdesk does the routing.
Variants by industry
| Industry | What the bot handles | What always goes to a person |
|---|---|---|
| Fibre, ISP, telco | Outage checks, fault reports with account number, moving a line | Faults with no known outage, cancellations |
| Municipal or utility services | Fault and meter reports, billing questions answered from your system | Disputes and complaints |
| Retail and e-commerce | Opening hours, returns policy, order look-ups | Damaged goods, refunds |
| Schools and colleges | Fees, dates and requirements from the facts you supply | Serious admissions enquiries |
| Advice lines (pet food, nutrition, skincare) | Follow-up questions, guidance from your own guidelines | Anything health-related |
What we need from you
- Your WhatsApp Business number, connected through Meta's own sign-up, and a payment method on your Meta account, because Meta bills you directly. Our platform doesn't support using the WhatsApp Business app on the same number, so a number connected here can't stay in the app.
- Your top questions and the approved answers, plus your opening hours.
- The rules for handover: which topics the bot must never answer, and what counts as urgent.
- API access to any system the bot should look things up in, or your existing bot's API.
- A sample of real messages and voice notes (with personal details removed) so we can test sorting and transcription before go-live.
- Who on your team uses the inbox and the Android app.