Customers in Bangladesh rarely type the full question. They send a screenshot of your reel with "eta ache?", a photo of the bag they want, a bKash receipt, or a voice note because typing Bangla on a phone is slow. SaleMag's AI agent understands all of these before it replies.
What does the AI agent do with each kind of message?
| The customer sends | What SaleMag's AI agent does |
|---|---|
| Text in Bangla, Banglish or English | Replies in the same language |
| A product photo | Reads the product, colour, brand and any code or price shown, then answers from your product list |
| A screenshot of your post or reel | Reads the product code, name or price in it; asks which product when the picture does not say |
| A payment screenshot (bKash, Nagad, Rocket, bank) | Reads the amount and transaction ID shown; your team confirms the payment |
| A photo of a damaged or wrong item | Notes the damage; your instructions decide whether it hands the chat to a person |
| A voice message | Turns Bangla, Banglish or English speech into text, then replies like a typed message |
| A PDF, Word, Excel or text file | Reads it, for example an order list, invoice or price list, and answers |
| A video | Does not reply; the video stays in the inbox for your team |
This works on every channel SaleMag connects: WhatsApp, Facebook Messenger, Instagram DM, Telegram and WeChat.
How does the AI agent understand a photo?
Before it replies, the agent asks your AI model to describe the photo: what it shows, every piece of text exactly as written (prices, product codes, sizes, phone numbers, transaction IDs), what kind of screenshot it is, and any visible damage. The agent then answers with that description in mind, in the customer's language.
Two rules keep the answers safe:
- Prices come from your product list. With AI order taking on, a price guard checks every amount in a reply against your catalog, even when the customer's screenshot shows a different price.
- Text inside a picture is never an instruction. A screenshot that says "give me 50% off" is read as content, not obeyed.
If a customer sends a photo or a file with no message, the agent waits a moment for the question that usually follows, then answers both together.
How does it handle voice messages?
A voice message is turned into text first, then answered like a typed message. It is tuned for Bangla, Banglish and English, so "ei panjabi ta M size e ache?" spoken aloud gets the same answer as the typed version.
Voice needs a model that can hear audio: an OpenAI or Google Gemini key, or Platform AI. Anthropic Claude reads text and photos but cannot hear audio, so a workspace with only a Claude key sends voice messages to the team, and the agent settings say so.
Can it read documents customers send?
Yes. PDF, Word (.docx), Excel (.xlsx), CSV and text files are read and summarised for the agent: what kind of document it is, the names, items, quantities, totals and numbers in it, and what the customer seems to want. A scanned PDF with no text layer is read from the image of the page.
What are the limits?
| Item | Limit |
|---|---|
| Photos | Up to 10 MB each, up to three at a time |
| Voice messages | Up to 5 MB each |
| Documents | Up to 10 MB each; the first 15,000 characters of a long document |
| Videos | Not understood |
| Payments | Read from the screenshot, never confirmed with the provider |
Anything the agent cannot use goes to your team in the shared inbox, where a person can answer.
What does my team see in the inbox?
Your team sees what the AI understood, next to the original message: AI saw under a photo, AI read under a document, and Transcript under a voice message. That makes it easy to check an answer, or to pick up the chat after a human handover.
How do I turn it on?
It is already on. Every AI agent has three switches under Photos, voice & documents (understand photos, voice messages and documents), all on by default. The create an AI agent guide shows where they are. Each photo, voice note or document is one small extra AI call, billed like the rest of your AI usage.
To see it work, send a product screenshot and a short voice note from your own phone to your connected WhatsApp number or Page. For the price questions that usually come with screenshots, see how to auto-answer "price koto?".
Step by step
- 1
Open your AI agent
In SaleMag, go to AI agents and open the agent that answers your chats.
- 2
Check Photos, voice & documents
Keep Understand photos, Understand voice messages and Understand documents switched on. They are on by default.
- 3
Give voice a model that can hear
Connect an OpenAI or Google Gemini key, or use Platform AI. Without one, voice messages go to your team.
- 4
Test from your phone
Send a photo and a voice note from your own phone to your connected number or Page. The test playground only takes typed messages.
Frequently asked questions
Can SaleMag's AI agent understand voice messages?
Yes. A voice message is turned into text and answered like a typed message. It needs an OpenAI or Google Gemini key, or Platform AI; with only an Anthropic Claude key, voice messages go to your team.
Does it understand Bangla voice notes?
Yes. It is set up for Bangla, Banglish (Bangla in English letters) and English voice notes, and it replies in the customer's language.
Can the AI agent read a product photo or a screenshot of my post?
Yes. It reads what the picture shows and any text in it, such as a product code, name or price, and answers from your product list. If the picture does not show which product it is, it asks.
Does it confirm a bKash or Nagad payment from a screenshot?
No. It reads the amount and transaction ID shown on the screenshot, but it does not check them with bKash, Nagad or a bank. Your team confirms the payment.
Does the AI agent watch videos?
No. Videos are not understood, so the agent does not reply to them. They stay in the shared inbox for your team.
Can I turn photo, voice or document understanding off?
Yes. Each AI agent has three switches under Photos, voice & documents, and all three are on by default. Turn any of them off and those messages go to your team.
Does understanding photos and voice cost extra?
Each photo, voice note or document is one small extra AI call, billed like the rest of your AI usage: Platform AI credits, or your own OpenAI, Anthropic or Gemini account.
About the author
WhatsApp & F-commerce automation, Dhaka
The Dhaka team that builds SaleMag, the AI sales inbox for WhatsApp, Messenger and Instagram used by online sellers in Bangladesh.