<a href="https://www.devstree.com/"><img src="assets/brand/devstree-logo.svg" alt="DevsTree · AI Powered Digital Excellence" height="40" class="dt-logo"></a>

# AI Call Assistant: Client Guide

**One AI assistant that chats, talks and calls for your business, in every language.**

Prepared by **DevsTree IT Services** · [devstree.com](https://www.devstree.com/) · [info@devstree.com](mailto:info@devstree.com)

---

## Contents

1. [The product in one minute](#1-the-product-in-one-minute)
2. [Who is involved](#2-who-is-involved)
3. [The end-to-end flow](#3-the-end-to-end-flow)
4. [Company setup: ten steps to go live](#4-company-setup-ten-steps-to-go-live)
5. [Where your leads come from](#5-where-your-leads-come-from)
6. [What the customer experiences](#6-what-the-customer-experiences)
7. [How the assistant talks](#7-how-the-assistant-talks)
8. [Multi-language: the core feature](#8-multi-language-the-core-feature)
9. [After every conversation](#9-after-every-conversation)
10. [The company dashboard](#10-the-company-dashboard)
11. [Safety, privacy and trust](#11-safety-privacy-and-trust)
12. [How we build it: the roadmap](#12-how-we-build-it-the-roadmap)
13. [Decisions we need from you](#13-decisions-we-need-from-you)
14. [Plain-language glossary](#14-plain-language-glossary)

---

## 1. The product in one minute

Any company, in any industry (clinic, real estate, online shop, school, agency, software), can sign up, upload its information, and get an AI assistant that answers its customers.

| Channel | Who starts it | How it works |
|---|---|---|
| 💬 **Website chat** | Customer | A chat bubble on the company website. Type a question, get an instant answer. |
| 🖥️ **Web call** | Customer | A “Talk to us” button. Speak with the assistant in the browser. No phone needed. |
| 📞 **Phone: customer calls** | Customer | Customers ring the company's number. The assistant picks up, day or night. |
| 📲 **Phone: AI calls back** | AI assistant | A customer fills an enquiry form, and the assistant calls them within minutes. |

The assistant:

- answers in **near real time** (about one second on calls) with a **natural, human-like voice**,
- uses **only the company's own information**, and never makes things up,
- speaks the **customer's language**, and switches when they switch,
- **hands over to a real person** when needed.

```mermaid
mindmap
  root((AI Call Assistant))
    Channels
      Website chat
      Web call
      Phone in
      AI calls back
    Knows your business
      Documents & FAQs
      Website pages
      Company profile
    Gets things done
      Answers questions
      Books appointments
      Qualifies leads
      Follows up forms
    Every language
      Detects language
      Switches mid-call
      Local voices
    Reports back
      Transcripts
      Summaries
      Leads
      Dashboard
```

---

## 2. Who is involved

```mermaid
flowchart TB
    PA["🛠️ Platform admin (us)<br/>onboards companies, plans, numbers, languages"]
    CO["🏢 Company (our client)<br/>sets up the assistant, reviews calls & leads"]
    AI(["✨ AI assistant"])
    EU["🙋 End user (the company's customer)<br/>chat · web call · phone, in their language"]
    HU["👩‍💼 Company staff"]
    PA -->|creates & manages| CO
    CO -->|documents, settings, goals| AI
    EU <-->|asks · answers · calls back| AI
    AI -->|transcripts, summaries, leads| CO
    AI -.->|transfers when needed| HU
    style AI fill:#4f46e5,color:#fff,stroke:#4338ca
```

| Role | Who | What they do |
|---|---|---|
| **Platform admin** | Us | Onboard companies, manage plans and billing, the phone number pool, supported languages and global safety settings. |
| **Company admin & staff** | Our client | Set up the assistant, pick languages and voices, upload knowledge, connect phone numbers and forms, review conversations and leads. |
| **End user** | The client's customer | Chats, talks in the browser, calls in, or receives a call after filling a form, in their own language. |
| **AI assistant** | The product | Talks to the end user, looks up answers, collects details, books or escalates, writes the summary. |

---

## 3. The end-to-end flow

This is the whole story on one page: the company teaches the assistant, customers talk to it, and every conversation flows back into the company's dashboard, where the company fills gaps and the assistant gets better.

```mermaid
flowchart TB
    A["🏢 1–4 · Company sets up<br/>template · knowledge · assistant · leads · channels"]
    C["🙋 5 · Customer reaches out<br/>💬 chat · 🖥️ web call · 📞 phone · 📝 form → call back"]
    B(["✨ AI assistant<br/>one brain · every channel · every language"])
    D["✅ 6 · Goal done, or a human takes over"]
    E["📬 7 · Automatic follow-up<br/>transcript · summary · lead · notifications"]
    F["📈 8 · Company dashboard<br/>review and improve"]
    A --> B
    C --> B
    B --> D --> E --> F
    F -. "adds missing answers" .-> A
    style B fill:#4f46e5,color:#fff,stroke:#4338ca
```

### The eight steps, in words

| # | Step | Who | What happens |
|---|---|---|---|
| 1 | **Company joins** | Company | Account created; industry template pre-fills tone, typical questions, details to collect and goals. |
| 2 | **Teaches the assistant** | Company | Uploads documents in any language; sets the assistant's name, voice, languages, rules and goals. |
| 3 | **Brings in leads, connects channels** | Company | Imports a spreadsheet or connects Meta lead ads / the CRM; one line of code on the website; buys or forwards a phone number; sets up forms; adds a staff number for handoff. |
| 4 | **Customer reaches out** | Customer | Chats, talks in the browser, phones in, or fills a form. Never has to pick a language. |
| 5 | **Assistant talks** | AI | Detects language, finds the answer in the company's documents, works toward the goal, confirms details. |
| 6 | **Goal done, or a human takes over** | AI | Books / confirms, or transfers to a person (with a summary), or books a callback. |
| 7 | **Automatic follow-up** | Platform | Transcript, recording, summary, lead, team email, CRM push, confirmation to the customer. |
| 8 | **Company reviews and improves** | Company | Sees everything in the dashboard; answers the questions the assistant couldn't. The loop closes. |

---

## 4. Company setup: ten steps to go live

One-time setup. No custom code per company: the industry template fills in sensible defaults, and the company adjusts them.

```mermaid
flowchart TB
    subgraph S1["Basics"]
        direction LR
        A["1 · Sign up"] --> B["2 · Industry<br/>template"] --> C["3 · Company<br/>profile"]
    end
    subgraph S2["Teach & configure"]
        direction LR
        D["4 · Upload<br/>knowledge"] --> E["5 · Configure<br/>assistant"] --> F["6 · Choose<br/>goals"]
    end
    subgraph S3["Leads & channels"]
        direction LR
        L["7 · Bring in<br/>your leads"] --> G["8 · Connect<br/>channels"]
    end
    subgraph S4["Launch"]
        direction LR
        H["9 · Test in<br/>playground"] --> I["10 · 🚀 Go live"]
    end
    S1 --> S2 --> S3 --> S4
    style L fill:#e0e7ff,stroke:#4f46e5,color:#1e1b4b
    style I fill:#10b981,color:#fff,stroke:#059669
```

### Step 1 · Sign up
The company account is created (by the company or by us) under an organisation, with an admin user and staff roles.

### Step 2 · Pick an industry template
Clinic, Real estate, E-commerce, Education, Software agency or Generic. The template pre-fills tone, typical questions, details to collect and goals. **The industry is a setting, not code**, which is how one platform serves any kind of company.

```text
┌─ console · Templates ─────────────────────────────────┐
│  What kind of business are you?                        │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐                   │
│  │ 🏥 Clinic│ │🏠 Real   │ │🛒 E-com- │                   │
│  │  ✓       │ │  estate │ │  merce  │                   │
│  └─────────┘ └─────────┘ └─────────┘                   │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐                   │
│  │🎓 Educa- │ │💻 Agency │ │⚙️ Generic│                   │
│  │  tion   │ │         │ │         │                   │
│  └─────────┘ └─────────┘ └─────────┘                   │
│  Includes: Book a consultation · Answer FAQs           │
│  Collects: name, phone, reason, preferred time         │
└────────────────────────────────────────────────────────┘
```

### Step 3 · Company profile
Name, branches, working hours, services and products, pricing notes, escalation contacts, and **things the assistant must never say** (for example “no medical diagnoses”).

### Step 4 · Upload knowledge
PDFs, Word files, text, FAQs and website links (auto-read), **in any language**. No translation needed. Each document shows a status.

```text
┌─ console · Knowledge ─────────────────────────────────┐
│  ┌ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┐   │
│    ⬆  Drop files or paste a website link              │
│  └ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ┘   │
│  📄 Price-list-2026.pdf     EN            [ Ready ]    │
│  📄 Servicios.docx          ES            [ Ready ]    │
│  📊 FAQ.csv                 EN · 48 Q&As  [ ⟳ Processing ]│
│  🌐 abcclinic.com           34 pages      [ ⟳ Reading ]   │
│  📄 old-brochure.pdf        unreadable    [ Failed ]   │
└────────────────────────────────────────────────────────┘
```

### Step 5 · Configure the assistant
- **Name and persona**, e.g. “Maya from ABC Clinic, warm and professional”.
- **Languages**: which ones to speak, and a default (see [section 8](#8-multi-language-the-core-feature)).
- **Voice**: one per language, or one multilingual voice so the persona sounds the same everywhere. Preview each.
- **Greetings** for chat and calls, per language (auto-translated drafts the company edits).
- **Guardrails**: “don't give medical advice”, “don't quote final prices”, “always offer a callback”.

```text
┌─ console · Assistant ─────────────────────────────────┐
│  Name [ Maya ]          Tone [ Warm, professional ]    │
│  Languages & voices                                    │
│   🇬🇧 English  (default)   ▶ Maya            [■■ on ]  │
│   🇪🇸 Spanish              ▶ Maya (multi)    [■■ on ]  │
│   🇮🇳 Hindi                ▶ Maya (multi)    [■■ on ]  │
│   🇸🇦 Arabic   beta voice                    [□□ off]  │
│  Guardrails: Always offer a callback   [+ add]         │
└────────────────────────────────────────────────────────┘
```

### Step 6 · Choose goals (playbooks)
What the assistant should get done in a conversation. Each goal lists the details to collect and what happens at the end. **Written once, works in every enabled language.**

| Goal | Details it collects | What happens at the end |
|---|---|---|
| Answer questions (default) | – | Logs anything it couldn't answer |
| Qualify a lead | Name, need, budget, timeline | Email team, push to CRM |
| Book an appointment | Name, phone (read back), reason, date/time | Confirms slot, SMS/email confirmation |
| Follow up an enquiry form | Only what the form didn't already capture | Used for AI call-backs |

### Step 7 · Bring in your leads
The people the assistant will call back. Leads can come from a **spreadsheet** (CSV, Excel or Google Sheets), **Meta lead ads** (Facebook and Instagram), the company's **website forms**, its **CRM**, and from chats and calls. The company matches the spreadsheet's columns once; every row is then checked, and only people who gave consent are called. Full detail in [section 5](#5-where-your-leads-come-from).

```text
┌─ console · Leads · Import ───────────────────────────────┐
│  clinic-leads-nov.csv · 248 rows                          │
│  Full Name   → Name          ✓                            │
│  Mobile      → Phone         ✓   07700 900123 → +44 …     │
│  Opted in    → Consent       ✓                            │
│  Lang        → Language      ✓                            │
│  ───────────────────────────────────────────────────────  │
│  ✅ 231 ready to call                                      │
│  ⚠️  12 no consent → saved as contacts, never called        │
│  ⛔   5 duplicates / do-not-call → skipped                  │
│  [ Add to campaign: Website enquiry follow-up ]           │
└───────────────────────────────────────────────────────────┘
```

### Step 8 · Connect channels
- **Website chat + web call:** paste one script snippet on the website (only allowed domains work).
- **Phone:** buy a number from the platform, or forward the company's existing number. Optionally one number per language.
- **Forms / surveys:** use the platform's hosted form (in each enabled language), embed it, or connect an existing form tool.
- **Human handoff:** a staff phone number or email for transfers and callbacks, optionally per language.

```html
<script src="https://cdn.yourplatform.com/widget.js" data-company="abc-clinic"></script>
```

### Step 9 · Test in the playground
Chat and talk to the assistant from the dashboard in each language, and see **which documents it used** for every answer.

### Step 10 · Go live
Turn it on. Customers can now chat, call and receive call-backs.

---

## 5. Where your leads come from

The assistant only calls people who **asked to hear from you**. This chapter shows where those people come from, how they get into the platform, and what happens before the first call.

```mermaid
flowchart TB
    subgraph SRC["1 · Leads come in"]
        direction LR
        S1["📄 Spreadsheet<br/>CSV · Excel"] ~~~ S2["📊 Google<br/>Sheets"] ~~~ S3["📣 Meta lead ads<br/>Facebook · Instagram"] ~~~ S4["📝 Website<br/>forms"] ~~~ S5["🔄 Your<br/>CRM"]
    end
    SRC --> CHK["2 · Matched & checked<br/>columns · phone format · duplicates<br/>consent · do-not-call · language · time zone"]
    CHK --> R["✅ Ready to call"]
    CHK --> C["⚠️ Contact only<br/>no consent: never called"]
    CHK --> X["⛔ Skipped<br/>duplicate / do-not-call"]
    R --> CALL["3 · The AI calls<br/>calling hours · their language · retries"]
    CALL --> RES["4 · Results come back<br/>dashboard · CRM · email/SMS · export"]
    style R fill:#d1fae5,stroke:#10b981,color:#065f46
    style C fill:#fef3c7,stroke:#f59e0b,color:#78350f
    style X fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
```

People who chat or call you are captured as leads automatically; they started the conversation, so no list is needed for them.

### 5.1 The ways leads arrive

| Source | How it connects | Good for | Where consent comes from |
|---|---|---|---|
| 📄 **Spreadsheet** (CSV, Excel) | Upload in the console; match the columns once | Existing lists, event sign-ups, older enquiries | A consent column and date in the file |
| 📊 **Google Sheets** | Paste the sheet link, or keep the sheet in sync | Teams that already track leads in Sheets | Same as above |
| 📣 **Meta lead ads** | Connected once through Zapier or Make (a direct connection later) | Facebook and Instagram campaigns | The “you may call me” question on the ad's form |
| 📝 **Website forms** | Hosted or embedded form | Enquiries on your website | The consent tick box |
| 🔄 **Your CRM** | A workflow in the CRM sends new leads | Leads already in your CRM | Consent recorded in the CRM |
| 💬 **Chats and calls** | Automatic | People who contacted you | They started the conversation |

### 5.2 Importing a spreadsheet, step by step

```mermaid
sequenceDiagram
    actor S as Clinic staff
    participant P as Platform
    participant AI as AI assistant
    S->>P: Uploads clinic-leads-nov.csv (248 rows)
    P->>S: Suggests which column is name, phone, consent, language
    S->>P: Confirms the matching
    P->>P: Checks every row: phone format, duplicates,<br/>consent, do-not-call, language, time zone
    P->>S: 231 ready · 12 contact only · 5 skipped
    S->>P: Adds the ready leads to a campaign and starts it
    loop Inside calling hours, in each lead's time zone
        P->>AI: Next lead, with their details and language
        AI->>AI: Calls, completes the goal, retries if no answer
    end
    AI->>P: Outcomes, summaries, bookings
    P->>S: Dashboard updated · CRM · export
```

### 5.3 What the file needs

Column names don't have to match: the company picks which column is which during the import. A sample file can be downloaded from the interactive walkthrough.

| Column | Needed? | Example |
|---|---|---|
| **Name** | Required | Alex Martin |
| **Phone** | Required | +44 7700 900123 (local formats are converted automatically) |
| **Consent** + **date** | Required to be called | yes · 2026-11-02 |
| **Language** | Optional (otherwise the company default) | English |
| **Email** | Optional | alex@example.com |
| **What they need** | Optional (the assistant mentions it on the call) | Knee pain |

### 5.4 Meta lead ads in three steps

```mermaid
flowchart LR
    A["📣 Lead form ad<br/>Facebook / Instagram<br/>with a consent question"] --> B["🔗 Zapier or Make<br/>connected once"]
    B --> C["🛡️ Platform<br/>checks the lead"]
    C --> D["📞 AI calls<br/>within minutes"]
```

1. **Create a lead form ad** on Facebook or Instagram, with a “you may call me” question.
2. **Connect it once**, through Zapier or Make (or a direct connection later). Each ad field is matched to a platform field, the same way as a spreadsheet's columns.
3. **Every new lead arrives in seconds**, is checked, and the assistant calls within minutes, inside calling hours.

### 5.5 The consent rule

- The assistant **only calls people who agreed to be called**. The wording, date and language of that consent are kept with the lead.
- Imported rows **without consent** are saved as contacts, so the team can still reach them another way, but the assistant never calls them.
- Do-not-call lists are always checked, for every source.

### 5.6 Getting your data out

- **Download as CSV** at any time: leads, conversations and call results.
- **Google Sheets sync** keeps a sheet up to date automatically.
- **CRM push** sends every new or qualified lead to the CRM as it happens.

> **Note for the build plan:** the system design covers website forms and webhook sources (Meta via Zapier/Make, CRM). Spreadsheet import, Google Sheets and export complete this lead flow and fit in the *Forms + AI call-backs* phase. See decisions 10–12.

---

## 6. What the customer experiences

### 6.1 Website chat

```text
┌──────────────────────────────┐
│ (M) Maya · ABC Clinic        │
│     AI assistant             │
├──────────────────────────────┤
│ ┌──────────────────────────┐ │
│ │Hi! I'm Maya. How can I   │ │
│ │help?                     │ │
│ │[Opening hours] [Prices]  │ │
│ └──────────────────────────┘ │
│        ┌───────────────────┐ │
│        │Do you open Saturday?│ │
│        └───────────────────┘ │
│ ┌──────────────────────────┐ │
│ │Yes! 9:00 to 14:00 at City│ │
│ │Centre. Book a visit?     │ │
│ │📄 Opening-hours.pdf       │ │
│ └──────────────────────────┘ │
├──────────────────────────────┤
│ [Type a message…] [📞 Call]  │
└──────────────────────────────┘
```

1. Customer opens the chat widget on the company website.
2. Sees the company's branding, welcome message and suggested questions, in the website / browser language.
3. Asks questions in any enabled language; the assistant answers from the company's knowledge in that language.
4. When needed, the assistant collects details (with an email code check where the goal requires it) and completes the goal.
5. Can switch to a web call from the same widget; the call continues in the same language and remembers the chat.
6. The transcript can be emailed to the customer.

```mermaid
sequenceDiagram
    actor U as Customer
    participant W as Chat widget
    participant AI as AI assistant
    participant C as Company
    U->>W: Opens chat on website
    W->>U: Welcome + suggested questions (page language)
    U->>AI: "Do you open on Saturdays?"
    AI->>AI: Finds the answer in company documents
    AI->>U: "Yes, 9:00 to 14:00. Book a visit?"
    U->>AI: "Yes, for knee pain"
    loop One detail at a time
        AI->>U: Asks name / phone / time
        U->>AI: Answers
    end
    AI->>U: Reads back details, confirms booking ✅
    opt Customer prefers to talk
        U->>W: Taps "Call instead"
        W->>AI: Voice call continues same chat & language
    end
    AI->>C: Summary + lead + transcript
```

### 6.2 Web call (voice in the browser)

```text
┌──────────────────────────────┐
│ (M) Maya · ABC Clinic   [ES] │
├──────────────────────────────┤
│            ◉◉◉               │
│         (speaking)           │
│      ▁▃▅▇▅▃▁▃▅▇▅▃▁            │
├──────────────────────────────┤
│ Customer: ¿Tienen fisio los  │
│           sábados?           │
│ 🌐 Spanish detected → voice  │
│ Maya: ¡Sí! De 9:00 a 14:00.  │
├──────────────────────────────┤
│          (🎙)   (✕)           │
└──────────────────────────────┘
```

```mermaid
sequenceDiagram
    actor U as Customer
    participant W as Website widget
    participant P as Platform
    participant AI as AI assistant
    U->>W: Clicks "Talk to us"
    W->>U: Browser asks for microphone
    W->>P: Request a call session (company, website, page language)
    P-->>W: Call session ready
    W->>AI: Audio connected
    AI->>U: Greeting + "I'm an AI, this call may be recorded"
    U->>AI: Speaks (any enabled language)
    AI->>AI: Detects language, switches voice if different
    loop Conversation
        U->>AI: Speaks
        AI->>U: Replies in the same language, in about a second
    end
    Note over U,AI: Customer can interrupt anytime, the assistant stops and listens
    AI->>U: Wraps up, confirms next step
    AI->>P: Transcript + summary + collected details
```

- Works in any modern browser; no phone needed.
- The customer can interrupt mid-sentence; the assistant stops and listens, like a real person.
- Live captions can be shown in the widget.

### 6.3 Phone call (customer calls the company)

```mermaid
sequenceDiagram
    actor U as Caller
    participant T as Phone network
    participant P as Platform
    participant AI as AI assistant
    participant H as Company staff
    U->>T: Dials company number
    T->>P: Incoming call for number X
    P->>P: Which company owns X?<br/>Is the caller known?<br/>Pick language: number, history or default
    P->>AI: Start assistant with company setup
    AI->>U: Greeting + "this call may be recorded"
    U->>AI: First sentence
    AI->>AI: Detect language, switch if needed
    loop Conversation
        U->>AI: Question
        AI->>U: Answer from company knowledge, same language
    end
    alt Needs a human
        AI->>H: Transfers call with a short summary<br/>(to a staff member who speaks the language, if configured)
    else Done
        AI->>U: Confirms next step, says goodbye
    end
    AI->>P: Transcript, recording, summary, lead
```

- **Returning callers** are greeted by name, in their preferred language, and the assistant remembers the last conversation.
- **Out of hours**, the assistant still answers and takes a callback request.

### 6.4 Form, then the AI calls the customer (outbound)

```text
┌─ abcclinic.com/consulta ─────┐        ┌───────────────┐
│ Solicitar información         │        │  ABC Clinic   │
│ Nombre   [ Alex Martín      ] │  ───▶  │  calling…     │
│ Teléfono [ +34 600 000 000  ] │ checks │     (M)       │
│ Motivo   [ Dolor de rodilla ] │  ✓✓✓✓  │ "Hola Alex,   │
│ Idioma   [ Español          ] │        │  soy Maya…"   │
│ ☑ Acepto que me llamen        │        │  (✕)   (🎙)    │
│ [ Enviar ]                    │        └───────────────┘
└───────────────────────────────┘
```

```mermaid
sequenceDiagram
    actor U as Customer
    participant F as Company form
    participant P as Platform
    participant AI as AI assistant
    U->>F: Fills enquiry form (name, phone, need,<br/>preferred language) + ticks consent
    F->>P: New lead
    P->>P: Check consent, duplicates,<br/>do-not-call list, calling hours
    P->>AI: Schedule call in the customer's language<br/>(within minutes, or next allowed time)
    AI->>U: Calls customer
    alt Customer answers
        AI->>U: "Hi Alex, this is Maya from ABC Clinic.<br/>You just asked about knee pain. Is now a good time?"
        loop Conversation
            U->>AI: Questions / details
            AI->>U: Answers, collects only missing info
        end
        AI->>U: Books / confirms next step
    else Busy / "call me later"
        AI->>P: Schedule callback at the time asked
    else No answer / voicemail
        AI->>P: Retry later (up to 3 tries),<br/>then SMS / email with a link, in the customer's language
    end
    AI->>P: Outcome, summary, lead updated
    P->>F: Company notified (email / CRM)
```

- The assistant already knows what the customer wrote in the form, so it **does not ask again**.
- The call language comes from the **form's language** or the customer's chosen preference.
- Calls happen **only during allowed hours** in the customer's time zone, **only with consent**, and the assistant **always says it is an AI**.

**If the customer doesn't answer** (default rules, each company can change them):

```mermaid
flowchart TD
    A["📞 First call, within calling hours"] -->|no answer| B["🔁 Retry after 1 hour"]
    B -->|no answer| C["🔁 Retry after 4 hours"]
    C -->|no answer| D["🔁 Retry next day"]
    D -->|no answer| E["💬 SMS / WhatsApp / email with a link,<br/>in the customer's language"]
    A -.->|"“call me later”"| CB["🕑 Callback at the time asked"]
```

**Default outbound rules** (each company can change them):

| Rule | Default |
|---|---|
| First call | Immediately, within calling hours |
| Calling hours | 10:00–19:00 in the customer's time zone, not on holidays |
| Call language | Form language or preferred language, else company default; still switches if the customer answers in another language |
| Retries | 3 attempts: +1 hour, +4 hours, next day. Stops on “don't call me” |
| Callback requests | The time the customer asks for becomes a new call |
| After the last failed attempt | SMS / WhatsApp / email in the customer's language with a link to chat or book |
| Consent | Required on the form; no consent means no call |
| Do-not-call | Platform and company lists, plus national registries where required |

---

## 7. How the assistant talks

### 7.1 What happens on every turn (voice)

```mermaid
flowchart TB
    A["🗣️ Customer speaks"] --> B["🎧 Listen: voice → text, live,<br/>language detected"]
    B --> C["⏱️ Notice the customer finished"]
    C --> D["🧠 Think: check company documents and the goal"]
    D --> E["✍️ Write the reply word by word,<br/>in the customer's language"]
    E --> F["🔊 Speak with a human-like voice,<br/>starting before the reply is finished"]
    F -->|next turn| A
    F -.->|customer interrupts| G["✋ Stop speaking within 0.3 s, listen"]
    G -.-> B
```

**Target: the assistant starts speaking about 1 second after the customer stops.** Everything runs as a stream, so it never waits for the full answer before starting to talk.

### 7.2 Where that one second goes

Targets from the system design; real numbers are measured per language once live. The chart uses the middle of each range (about 1.45 s in total); the goal is a typical reply under 1 second and 95% of replies under 1.6 seconds, in every Tier 1 language.

```mermaid
pie showData
    title Target per stage (ms)
    "Sound reaches us (network in)" : 125
    "Notice the customer finished" : 350
    "Final words turned into text" : 200
    "Think + first words of the reply" : 450
    "Voice starts" : 200
    "Sound reaches the customer (network out)" : 125
```

| Stage | Target | How we keep it fast |
|---|---|---|
| Sound reaches us (network in) | 0.05–0.2 s | Voice servers in the callers' region |
| Notice the customer finished | 0.2–0.5 s | Understands the sentence is complete, rather than waiting for a long silence |
| Turn speech into text | 0.1–0.3 s | Text arrives live while the customer is still talking |
| Look up the answer | mostly hidden | Starts searching while the customer is still speaking |
| Think and write the first words | 0.3–0.6 s | Fast model for live calls; company basics (hours, address) already loaded |
| Start the voice | 0.1–0.3 s | Starts speaking from the first sentence |
| Sound reaches the customer (network out) | 0.05–0.2 s | Same region |
| **Total** | **≈ 0.9–1.5 s** | Everything streams; nothing waits for a full response |

### 7.3 How it behaves

| Behaviour | Example |
|---|---|
| Speaks the customer's language | Replies in the language used; switches when the customer switches or asks (“Can we speak in Spanish?”). |
| Answers only from company knowledge | If the answer isn't in the documents, it says so and offers a callback or a person. Works even if documents are in another language. |
| Short, spoken-style replies on calls | One or two sentences, then a question. Long lists go by SMS/email instead. |
| Formatted replies in chat | Bullets, links, buttons. |
| Works toward the goal | Collects the required details one question at a time and skips what it already knows. |
| Confirms important details | Reads back phones, emails and dates: “That's alex at gmail dot com, right?” |
| Local formats | Says dates, times, numbers and prices the way people say them in that language. |
| Handles interruptions | Stops talking as soon as the customer speaks (in under 0.3 s). |
| Fills silence naturally | “Let me check that for you…” while a lookup takes a moment. |
| Knows when to hand off | Upset customer, sensitive topic, or “can I speak to a person?” → transfer or callback. |
| Same assistant everywhere | Same knowledge, persona and rules in chat, web call and phone, in every language. |
| Says it is an AI | At the start of every call, in the call's language. |

---

## 8. Multi-language: the core feature

Multi-language is a launch requirement, not an add-on. It applies to every channel.

### 8.1 The idea in one picture

```mermaid
flowchart LR
    D["📄 Your documents<br/>e.g. price list in English"] --> AI["✨ Assistant"]
    Q["🙋 Customer asks<br/>in Spanish"] --> AI
    AI --> R["💬 Answers in Spanish<br/>using the English file"]
```

- **Customers never pick a language.** They just speak or type.
- **Your documents can be in any language.** No translation work.
- **Your staff read everything in their own language.**

### 8.2 Company side

| What | How |
|---|---|
| Choose languages | Pick the enabled languages and one default. The platform shows which languages support full voice and which are chat-only for now. |
| Voices | One voice per language, or one multilingual voice for a consistent persona. Preview each. |
| Knowledge | Upload in any language. Optional per-language FAQ answers where exact wording matters (legal, medical, pricing). |
| Greetings, disclosures, messages | Per-language greetings, AI/recording notice, SMS/email templates. Auto-translated drafts the company reviews. |
| Forms | Hosted forms in each enabled language; the form language becomes the customer's preferred language. |
| Phone numbers | One number that detects the language, or one number per language. |
| Human handoff | Optional: send transfers to staff who speak the caller's language. |
| Reviewing | Staff read each transcript in the original or translated; summaries are written in the staff's language. |

### 8.3 Customer side: how the language is chosen

```mermaid
flowchart TD
    S["Conversation starts"] --> K{"Language<br/>already known?"}
    K -->|"yes: form, past chat,<br/>or language number"| L1["Start in that language"]
    K -->|no| L2["Start in website language<br/>or company default"]
    L1 --> T["Customer speaks or types"]
    L2 --> T
    T --> D{"Same<br/>language?"}
    D -->|yes| R["Reply in that language"]
    D -->|"no, enabled"| SW["Switch language + voice"]
    D -->|"no, not enabled"| FB["Apologise, offer<br/>person or callback"]
    SW --> R
```

**Switching without flip-flopping.** One borrowed word or a name does not trigger a switch. The assistant switches when the customer clearly continues in the new language, or asks for it:

```mermaid
flowchart TD
    A["Speaking the current language"] -->|"one sentence in another<br/>enabled language"| B{"Just a word<br/>or a name?"}
    B -->|"yes: next sentence is back<br/>in the current language"| A
    B -->|"no: second sentence in the new language,<br/>very clear, or the customer asks"| C["Switch to the new language<br/>and its voice"]
    C --> A
    A -->|"language not enabled<br/>for this company"| D["Apologise, offer a person,<br/>callback or chat"]
```

- Mixed-language sentences are understood.
- Follow-up SMS, emails and confirmations arrive in the language the customer used.

### 8.4 Assistant side

- Detects the language on every turn and keeps track of the current one.
- Searches knowledge across languages, then answers in the customer's language, keeping names, prices and product terms exact.
- Uses the voice, greeting and disclosure text for that language.
- Formats numbers, dates, times and currency for that language when speaking or writing.
- Stores the transcript in the original language and records which languages were used.

### 8.5 Language tiers

Voice quality differs by language and provider, so each language gets a published tier.

| Tier | Chat | Voice calls | When a language gets this tier |
|---|---|---|---|
| **Tier 1** | ✅ | ✅ Full quality, ~1 s replies | Passes our test calls for accuracy, natural voice and speed |
| **Tier 2** | ✅ | ✅ Marked “beta”, slightly slower (~1.5 s) | Passes with lower thresholds |
| **Tier 3** | ✅ | ❌ Offers a chat link or callback instead | Chat quality check only |

Tiers are reviewed as speech providers improve. Adding a language is a settings change plus testing, not new development.

---

## 9. After every conversation

All automatic, the moment a chat or call ends.

```mermaid
flowchart LR
    END(["Conversation ends"]) --> T["📝 Transcript saved<br/>original language + recording"]
    END --> S["✨ AI summary<br/>in staff's language:<br/>details, outcome, mood"]
    END --> A["📬 Goal actions<br/>team email · CRM · confirmation<br/>to customer in their language"]
    END --> F["🔔 Follow-up message<br/>if left unfinished"]
    END --> Q["❓ Unanswered questions<br/>to 'improve knowledge' list"]
    END --> U["📊 Usage recorded<br/>for billing"]
```

**Example summary card** (sample data):

| Field | Value |
|---|---|
| Summary | Alex called about knee pain treatment. Asked about Saturday hours and price. Booked a consultation with Dr. Rao for Saturday 10:30. Prefers SMS reminders. |
| Channel | AI called back (form) |
| Language | Spanish → summary written in English |
| Collected | Name, phone (confirmed), need, slot |
| Outcome | ✅ Booked |
| Mood | 😊 Positive |
| Sent to | Team email · CRM · SMS to Alex (Spanish) |

Possible outcomes: **resolved**, **lead captured**, **booked**, **transferred**, **callback needed**.

---

## 10. The company dashboard

```mermaid
flowchart LR
    D["📈 Dashboard"] --> C1["💬 Conversations<br/>transcripts, recordings, summaries"]
    D --> C2["🧲 Leads<br/>collected details, outcome, next step"]
    D --> C3["❓ Unanswered questions"]
    D --> C4["📞 Campaigns<br/>outbound call status"]
    D --> C5["🧾 Usage & billing"]
    C3 -->|company adds the answer| K["📚 Knowledge updated"]
    K -->|assistant improves in every language| D
```

Screen sketch (numbers are placeholders):

```text
┌─ ABC Clinic · Overview ───────────────────────── Last 7 days ─┐
│ ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐   │
│ │Conversations│ │Resolved byAI│ │Leads        │ │Avg reply   │   │
│ │   1,284    │ │    82%     │ │    217     │ │   1.1 s    │   │
│ └────────────┘ └────────────┘ └────────────┘ └────────────┘   │
│  Conversations per day           Languages used               │
│   ▇ ▇ ▇ █ █ ▅ ▃                   EN 52% · ES 26%              │
│   M T W T F S S                   HI 14% · AR 8%               │
│  Unanswered: “Do you accept XYZ insurance?” (23×) [Add answer] │
└───────────────────────────────────────────────────────────────┘
```

| View | What the company sees |
|---|---|
| **Conversations** | Every chat and call: transcript, recording, AI summary, mood, outcome and language. Other-language transcripts can be shown translated. |
| **Leads** | Details collected by the assistant: name, phone, email, need, preferred language, custom fields. |
| **Unanswered questions** | Questions the assistant couldn't answer, grouped across languages. Answer once; it works in every language. |
| **Campaigns** | Outbound calls from forms: queued, calling, connected, no answer, retry scheduled, done. |
| **Notifications** | Email / webhook / CRM push when a lead is qualified or a person is needed. |
| **Usage & billing** | Chat messages, web call minutes, phone minutes, numbers; breakdown by language. |


---

## 11. Safety, privacy and trust

| Topic | How it works |
|---|---|
| 🤖 **Honest it's an AI** | Every call starts with “I'm an AI assistant and this call may be recorded”, in the call's language. |
| ✅ **Calls only with consent** | The AI only calls people who ticked “you may call me”. The wording, time and language of consent are stored. |
| 🚫 **Do-not-call respected** | Platform and company do-not-call lists; national registries where required. |
| 🕙 **Calling hours** | Only inside allowed hours in the customer's own time zone. |
| 🔒 **Your data stays yours** | Each company's data is fully separated. One company can never see another's documents, calls or customers. |
| 📚 **Answers only from your facts** | The assistant uses your documents and rules. If it doesn't know, it says so and offers a person. |
| 🎙️ **Recordings protected** | Encrypted, viewable only by your authorised staff via short-lived links, deleted on your schedule or on request. Recording can be switched off. |
| ⚖️ **Legal text reviewed by you** | Disclosures and consent wording are never machine-translated without your team approving them. |
| 💸 **Abuse protection** | Rate limits and bot checks on the web call button, a maximum call length, monthly minute caps and daily outbound call caps. |

**Regional rules** (confirmed with legal for each launch country): in the US, AI voice calls need prior express consent and must follow calling-hour rules; in the EU/UK, GDPR applies (lawful basis, deletion requests, data location). Other markets are checked before outbound calling is switched on.

---

## 12. How we build it: the roadmap

Each phase ends with something you can see and test. Order matters more than dates; timing depends on team size. **Multi-language is built in from Phase 1, not added at the end.**

```mermaid
timeline
    title Build phases
    Phase 0 : Foundations
            : Accounts, roles, secure storage, company data separation
    Phase 1 : Chat + languages
            : Website chat, document upload, goals, templates, language detection
    Phase 2 : Web call
            : Voice in the browser, voices per language, playground, ~1 s replies
    Phase 3 : Phone in
            : Phone numbers, transfers to staff, recordings
    Phase 4 : Forms + AI call-backs
            : Hosted forms, spreadsheet import, scheduler, retries, consent, fallback messages
    Phase 5 : Product depth
            : Analytics, billing, more languages, CRM integrations
```

| Phase | What you get | How we know it's done |
|---|---|---|
| **0 · Foundations** | Accounts and roles, secure storage, company data separation | Data separation tests pass |
| **1 · Chat + languages** | Website chat, document upload, goals, templates, language detection, staff translation | A company in any industry goes live from settings alone, chatting in every launch language |
| **2 · Web call** | Voice in the browser, voice per language, recordings, test playground | ~1 s replies for each Tier 1 language; interruptions and mid-call language switches work |
| **3 · Phone in** | Buy/forward numbers, recognise callers, language-matched transfers, keypad input | A real number answers in the right language and transfers; recordings in the console |
| **4 · Forms + AI call-backs** | Multilingual hosted forms, spreadsheet import (CSV/Excel/Sheets), Meta lead ads via Zapier/Make, call scheduler, calling hours, retries, voicemail, consent, fallback messages, CSV export | A form submission produces an AI call in the customer's language within the allowed hours |
| **5 · Product depth** | Analytics by language, billing, more languages, provider fallback, CRM integrations | Self-serve onboarding without our help; new languages added via settings + testing only |

---

## 13. Decisions we need from you

Each has our suggested answer. Accept it or change it.

| # | Question | Why it matters | Our suggestion |
|---|---|---|---|
| 1 | **Which languages at launch?** | Decides voice providers, testing and translation work | The languages of your first customers. Full voice only after test calls pass. |
| 2 | **Which country / region first?** | Phone providers, speed and legal rules | One region first, chosen by your first customers |
| 3 | **Our AI service accounts, or each client's own?** | Asking each company for 4 provider accounts is heavy | We provide and meter them; own accounts as an option |
| 4 | **Managed voice hosting or our own servers?** | Speed to launch vs. cost and data location | Managed hosting for launch; self-host later if needed |
| 5 | **Record calls by default? How long to keep?** | Storage cost, privacy, industry rules | On, kept 90 days, each company can change it |
| 6 | **How does handoff to humans work?** | Not every company has a call centre or multilingual staff | Live transfer if a (language-matched) number is set, otherwise a callback request |
| 7 | **Translate every transcript?** | Translation costs money per conversation | Translate when staff open it; summaries always in staff language |
| 8 | **How do we price it? Same price for every language?** | Some languages cost more to run | Monthly plan with included minutes/messages + overage; review per-language margins after launch |
| 9 | **Calendar / CRM at launch?** | Booking goals need one connected | Simple webhook first, then one calendar provider |
| 10 | **Spreadsheet import at launch?** | Most clients already have leads in a sheet | Yes: CSV/Excel import with column matching, in the forms + call-backs phase |
| 11 | **Meta lead ads: Zapier or direct?** | A direct connection takes longer to build | Start through Zapier/Make (works today); add a direct connection later |
| 12 | **Data export?** | Clients want their data in their own tools | CSV download from day one; live Google Sheets sync as an option |

---

## 14. Plain-language glossary

| Term | Meaning |
|---|---|
| **Tenant** | One company using the platform. Each tenant's data is kept completely separate. |
| **Industry template** | A ready-made starting setup for a type of business, with typical questions and goals. |
| **Knowledge** | The documents, FAQs and website pages the assistant uses to answer. |
| **Playbook / goal** | What the assistant should get done in a conversation, and which details to collect. |
| **Persona** | The assistant's name and personality, e.g. “Maya from ABC Clinic, warm and professional”. |
| **Inbound / outbound call** | Inbound: the customer calls you. Outbound: the assistant calls the customer. |
| **Handoff / warm transfer** | Passing the call to a real person. “Warm” means the staff member first gets a short summary. |
| **Barge-in** | When the customer talks over the assistant. It stops and listens. |
| **Latency** | The pause between the customer finishing and the assistant replying. Target: about one second. |
| **Speech-to-text / text-to-speech** | Turning the caller's voice into text, and the assistant's reply back into a natural voice. |
| **Lead** | A potential customer whose details the assistant collected. |
| **Webhook / CRM push** | Automatically sending a lead or summary into the company's other systems. |
| **Voicemail detection** | Recognising an answering machine on outbound calls and leaving a short message instead. |
| **Consent** | The customer's permission to be called, captured on the form with time and wording. |
| **Lead import (CSV)** | Uploading a spreadsheet of people who asked to be contacted. Columns are matched once; every row is checked before any call. |
| **Lead source** | Where a lead came from: a spreadsheet, a Meta ad, a website form, the CRM, or a chat or call. |
| **Playground** | A test area where the company can chat and talk to its assistant before going live. |

---

*For the technical design (components, data model, speed budgets and trade-offs), see the separate system design document.*

---

### About DevsTree

**DevsTree IT Services** builds AI-powered web and mobile solutions, process automation and data science products. 500+ projects, 13+ years, 160+ professionals and 7 offices worldwide (Ahmedabad HQ, Richmond TX, London, Gladesville NSW, Brampton ON, Gqeberha, Ho Chi Minh City).

[devstree.com](https://www.devstree.com/) · [info@devstree.com](mailto:info@devstree.com) · +91 92653 20742 · +1 (470) 791-9443 · [LinkedIn](https://www.linkedin.com/company/devstreeit) · [Clutch](https://clutch.co/profile/devstree-it-services)

© 2026 DevsTree IT Services Pvt. Ltd.
