Vardast

Persian AI Chatbot: What English-First Bots Get Wrong

Category: Guides & Tools
Persian AI Chatbot: What English-First Bots Get Wrong

Half the messages Iranian buyers send are Finglish, slang or voice notes. Here is what a Persian AI chatbot has to handle before it can sell anything.

Here is a message that lands in an Iranian seller's inbox a hundred times a day: salam in kife meshki mojode? gheymatesh chnd?

No Persian letters. No punctuation. "chnd" instead of "chand", because the customer typed it with one thumb on a bus. A human reads it in half a second and answers with a price and a photo.

A chatbot built for English reads it as noise. It replies with the most expensive sentence in commerce: "Sorry, I didn't quite get that. Could you rephrase?"

Nobody rephrases. They go to the next page.

That gap is the whole subject of this piece. Persian-speaking buyers do not write the way a support-ticket dataset thinks they do, and a bot that only understands tidy Persian is only slightly better than one that understands none.

Finglish is not a typo. It is half your inbox.

Finglish is Persian written in Latin letters, and it is not a mistake anyone is going to stop making. People type it on old keyboards, on work laptops, on phones where switching layouts is one tap too many. Sometimes they mix both scripts inside a single sentence: salam، رنگ آبیش موجوده؟

There is no spelling standard for it either. «چند» comes in as chand, chnd, cand, 4and. «قیمت» as gheymat, gheymate, qeymat, gheimat. A rules-based bot with a keyword list will catch maybe two of those and drop the rest.

So the first question to ask any vendor is not "do you support Persian". Everyone says yes to that. Ask: what happens when the message arrives in Latin letters, misspelled, with an emoji in the middle? Then send one and watch.

Three small things that quietly break an English-first bot

Persian has a few mechanical quirks that never show up in an English test suite, and each one turns into a silently ignored customer:

  • The half-space. «می‌شه» and «می شه» and «میشه» are the same word to a person and three different strings to a naive matcher. Same for «نمی‌دونم» and «نمیدونم». Your catalogue is probably written the careful way. Your customers are not.
  • Arabic letters in Persian text. Many keyboards send ي and ك instead of ی and ک. Visually almost identical, byte-wise different. A product called «كيف» will never match your «کیف» unless the system normalizes first.
  • Two sets of digits. A customer writes «۲ تا» or "2 ta" for quantity, and «۱۲۰» or "120" for price. If the bot only parses ASCII digits, it reads a quantity of zero and asks again. The customer already told it twice.

None of this is exotic. It is just work that only gets done if someone building the product actually reads Persian every day.

The buyer almost never uses the word "price"

This is my favourite one, because it breaks even bots that handle the script correctly.

Iranian buyers ask for a price sideways. «پی‌وی», «پیوی کن», «تو دایرکت بگو», «قابل نداره؟», «چنده عزیزم», or just the product photo re-sent with a question mark. Sometimes it is a single «؟». On a post with a price already in the caption, the question is often not about the price at all: it is a request for reassurance that the item is still there.

An assistant that only recognizes "how much is this" answers about a quarter of these and stares blankly at the rest. And these are the good messages. Our own numbers put the median order that starts in a DM at around 1.3 million toman, so each dropped price question is not a rounding error. Twenty of them a week is real money walking out.

Worth saying plainly: the fix is not a bigger keyword list. It is an assistant that understands intent, in Persian, and knows that «پیوی» in a comment and "how much" in a DM are the same human wanting the same thing.

Tone is the part nobody writes into the spec

Translate an English support script into Persian word for word and you get something that reads like a bank SMS. Correct, cold, and faintly insulting.

Persian selling runs on a register English does not have. «قربونت», «فدات شم», «جانم؟», «چشم»: none of it means what a dictionary says it means, and all of it signals that a person is on the other side. Meanwhile «شما» versus «تو» is a real decision. A 22-year-old buying a phone case and a 50-year-old buying a carpet do not want the same distance.

The other trap is over-formality. A lot of Persian bots default to office-memo language: «احتراماً به استحضار می‌رساند». On an Instagram page selling sneakers, that reads like a summons.

The bar I use is simple. Read the bot's reply out loud. If you would be embarrassed to send it yourself from your own phone, it is wrong, no matter how grammatical it is.

Then there is the voice note

Iranian customers send voice. Forty seconds, background traffic, half the sentence in dialect, the product name somewhere in the middle. Sellers I talk to treat these as a tax on the day: you cannot skim a voice note, you have to stop and listen.

A Persian assistant that transcribes and answers those is the difference between an inbox you dread and one you don't. We wrote about what happens when AI listens to voice messages and turns them into orders. The short version is that voice messages convert well, because someone who records a 40-second question is not window shopping.

The English-first workaround, and why it fails

Some vendors solve Persian by translating: your customer's message goes to English, a model answers in English, the answer goes back to Persian. It demos beautifully. It falls apart in two places.

First, the round trip flattens everything that mattered: slang, urgency, the customer's register. Second, product names do not survive translation. «شال نخی» comes back as a generic scarf and your assistant confidently offers the wrong item, at the wrong price, with total conviction. Wrong answers delivered fluently are worse than no answer, because the customer believes them.

This is also why feeding the assistant your real catalogue matters more in Persian than in English. Sizes, colours, materials and the nicknames your customers actually use for your products need to live in the assistant's knowledge in Persian, spelled the way people type them.

One habit nobody builds for: the message arrives in pieces

Persian-speaking buyers rarely send one complete sentence. You get «سلام», then the product photo, then «موجوده؟», then «سایز ۳۸ هم دارین», all inside twenty seconds.

A bot that treats each of those as a separate conversation turn will answer five times and be wrong five times: a long welcome for «سلام», an apology for the photo, a question it already had the answer to. By the time it reaches the size question, the customer has three useless notifications and the unmistakable feeling of talking to a machine.

The fix is unglamorous and rarely built: wait a few seconds, group the burst, then send one reply that covers all four things. That is what a good salesperson does anyway. They let you finish.

Ten minutes that tell you everything

Before you pay anyone, run these six messages through their demo. In this order:

  1. salam gheymate in kifo mishe befarmaeid? — pure Finglish, misspelled. Does it answer, or ask you to rephrase?
  2. «۲ تا از مشکیش می‌خوام» — Persian digits and a colloquial verb. Does it read quantity two?
  3. «پیوی» — one word, no context. Does it understand that this is a price request?
  4. A 30-second voice note asking about sizes. Does it transcribe and answer, or go quiet?
  5. «اینو دیروز سفارش دادم کی می‌رسه؟» — an order-status question with no order number attached. Does it ask the right follow-up, or invent a delivery date?
  6. Something hostile: «چرا جواب نمی‌دین؟». Does it de-escalate like a person, or reply with a cheerful template?

Any assistant that handles four of these six is ahead of most of the market. Any that fails the first one is not a Persian chatbot; it is an English chatbot wearing a Persian font.

This is the part we spent most of our time on at Vardast: an assistant that reads Finglish, slang, typos and voice notes the way a good salesperson in your own shop would, answers from your real products in your own tone across Instagram, WhatsApp, Telegram and your site, and hands the conversation to you the moment it should. You can connect a page and send it these six messages yourself before deciding anything.

Because the test was never whether a bot can speak Persian. It is whether the person on the other end forgets to wonder.

FAQs About Persian AI Chatbots for Online Sellers

What makes a Persian AI chatbot different from an English one?

The script, the spelling, and the tone. Persian text arrives with half-spaces, Arabic look-alike letters and two sets of digits, and a large share of messages are written in Latin letters with no fixed spelling. On top of that, Persian selling uses a politeness register that a literal translation destroys.

Can a chatbot understand Finglish?

A good one can. It needs to treat Latin-script Persian as a first-class input rather than a fallback, and tolerate misspellings like chnd, gheymat and qeymat. Test it yourself: send one misspelled Finglish sentence to the demo and see whether it answers or asks you to rephrase.

Is translating messages to English and back a workable approach?

It demos well and fails in production. The round trip flattens slang and urgency, and product names rarely survive it, so the assistant confidently offers the wrong item. Wrong answers delivered fluently cost more than no answer at all.

Will a Persian chatbot handle voice messages?

Some do, and for Iranian sellers it matters more than almost anything else, because a large share of serious buyers record instead of typing. Look for transcription plus an actual answer, not just a text version of the audio dropped into your inbox.

How do I stop a Persian assistant from sounding robotic?

Give it your own replies as examples rather than a translated support script, and pick a register on purpose: formal shoma or casual to, depending on who buys from you. Then read its answers out loud. If you would not send that sentence yourself, rewrite the instruction.

Does a Persian AI chatbot work across Instagram, WhatsApp and Telegram?

It should, and it should carry the same customer memory across all of them. Iranian buyers switch channel mid-conversation, and an assistant that forgets what was said on Instagram when the customer writes on WhatsApp will re-ask questions the buyer already answered.