Every business has the same five questions arriving every day: opening hours, delivery times, returns, pricing, location. Someone answers them twenty times, the customer waits, and nobody learns anything from the repetition. A chatbot can absorb that load, but the right starting point is not a tool subscription.
Pull one month of WhatsApp and email conversations and list the questions that repeat. Sort them into three buckets: fixed answers, answers that need the customer's own data, and answers that need a person to think. The first two are chatbot territory. Leave the third alone.
If your records are a mess, sit with the person who handles enquiries for half an hour and write down everything they get asked. This is the cheapest feasibility study available, and it usually shows that half the questions are the same three questions in different words.
Ten questions with answers you wrote and approved beat a hundred questions with answers the bot invented. A bot that tries to cover everything sounds vague about everything, and customers notice within two messages.
Write the answers yourself, in your own words, and keep each one short enough to read on a phone.
Your customers do not type formal Arabic. They write "shkathar el taqseel" in Arabic script, drop words, and misspell on purpose. A bot trained only on Modern Standard Arabic fails at the first message, and the customer closes the chat with an opinion about your whole service.
Test with real questions from your own logs, and load the bot with local variants: shkathar, kam, bkam, wain, mata. Allow for typos and run-together words, because that is normal usage, not an edge case. Q8DM has been doing software, web and mobile design, and training in Kuwait since 1998, and the Arabic-language work always comes down to this: test with real text, not clean text.
The most important button in any chatbot is the one that reaches a person. A bot that loops and repeats itself annoys customers more than a one-minute wait on the phone.
Define the trigger: after two failed attempts, or when the customer types words like "complaint," "manager," or "I want to talk to someone." Pass the full transcript to the agent so the customer does not repeat themselves, and say plainly that a human is joining.
Every field you collect needs a reason you can explain to the customer in one sentence. If you cannot, remove the field.
After a month, look at three numbers: how many conversations, how many closed without a human, and which questions came up with no good answer. That last list is next month's work.
Ask the staff too, not just the dashboard. They know which chats ended with an unhappy customer and which answers actually saved them time. A wrong answer about fees costs more than the hours saved.
If you are planning a first chatbot and want the scope, dialect handling, and handover rules set properly, talk to us: q8dm.com