Services
Industries
Hephon is an AI agency in Dubai (Hephon L.L.C-FZ) that builds AI customer support agents for businesses in Saudi Arabia, the UAE and Lebanon. Inside the Panda app in Saudi Arabia, our assistant reads Saudi dialect, answers from the retailer’s own data and passes the conversations it is unsure about to a human agent with the context attached.
An AI first line that reads each customer message in the dialect it was written in, answers the routine ones from your knowledge base and live data, and scores its own confidence. Anything below the threshold goes to a person with the question, the AI’s reasoning and the reason it stopped, so nobody starts over.
At Panda, help ran through a call centre and a generic web form, with no triage, no self-service and no way to get a fast answer without waiting for a person. Millions of users and thousands of daily interactions were all handled the hard way.
At a Lebanese ISP with roughly 2,000 customers, support simply meant a phone call, because customers had nothing they could do for themselves online.
The common failure in production AI is the opposite one: a bot that guesses. The customer asks something it cannot handle, it produces a plausible answer that is wrong, the customer acts on it, and the support team finds out three days later from an angry customer.
Arabic makes both problems harder. Customers write in dialect, in Arabizi and in a mix of Arabic and English. We have seen models that handle formal Arabic at 95 percent accuracy fall to 60 percent on Saudi dialect for the same intents.
Triage is three decisions on every message: what is being asked, can it be answered safely, and if not, who should take it.
At Panda we built a dataset of real customer queries from anonymised support history, labelled by intent, and tuned against it. The model works with the dialect as written instead of converting it to formal Arabic first.
Product and support questions are answered from the retailer’s own knowledge base and live catalog, not from what a model remembers.
Each response carries a calibrated score based on the type of question and the strength of what retrieval found. The score decides whether the AI answers or hands over.
Below the threshold the AI tells the customer it is connecting them to someone, logs the conversation and routes it through Genesys Cloud CX to a human agent, who sees the question, the reasoning and why the AI stopped.
For the Lebanese ISP, a Telegram bot lets customers report issues, share their location for service queries and handle basic account tasks. Telegram was chosen because it is what that customer base already used.
Panda’s team uses the assistant’s conversation logs to find product gaps and merchandising opportunities: a structured record of what customers ask for, in their own words.
The support team, not the model, decides where the line between AI and people sits.
Too high and the AI passes too much to people; too low and it answers things it should not. The threshold is tuned continuously from the support team’s feedback.
Low-confidence conversations go to a human with full context. The AI stops rather than guessing.
Handoffs are logged with the reason, so the team can see what the AI passed on and why, and adjust.
Panda
Eight weeks from start to live inside the app, with integration testing from week three.
Lebanese ISP
The Telegram bot shipped as part of a three-month platform build and has since expanded to more self-service flows.
Most of the time goes into the language and the handoff, not the model: real customer messages as test data, and a handoff your agents are happy to receive.
A focused first build usually lands between a few thousand and a few tens of thousands of dollars (AED or SAR equivalent), depending on how many systems it touches and how clean your data is. We scope a fixed price after a short discovery, so you know the number before you commit.
We build for the dialect your customers write. In production that is Saudi dialect at Panda; the same approach covers Gulf and Lebanese Arabic, Arabizi in Latin letters and messages that mix Arabic with English.
Every reply carries a confidence score. Below the threshold the AI stops, tells the customer it is connecting them, and routes the conversation to an agent with the full context.
At Panda the handoff runs through Genesys Cloud CX. We integrate with the support tools you already run rather than adding a separate inbox.
No. The agent sees the customer’s question, the AI’s reasoning, what it found and why it stopped, so the conversation continues where it left off.
The one your customers already use. Panda’s assistant is inside the Panda app; the Lebanese ISP’s bot is on Telegram; in the Gulf that is usually WhatsApp.
Read the logs: what was answered, what was handed over and why. At Panda a meaningful share of conversations are resolved end to end without a person; the exact figures stay private until the retailer clears them.
Tell us what you run today. We will say plainly whether this fits, what we would change for your setup, and what a fixed-price first build looks like.
