Skip to content
By Mishahal Palakuniyil, Co-Founder & CTO||5 min read

What We Learned Building AI Collections Calls for a Gulf Telecom

Lessons from building an AI agent that makes outbound collections calls in Arabic and English for a major Bahraini telecom: compliance, caller ID, handoffs, and what the numbers looked like.

Voice AIGulf & Middle EastCase Study
What We Learned Building AI Collections Calls for a Gulf Telecom

We built this for a major telecom in Bahrain. We are not naming them, and the numbers below are rounded. The lessons are the point.

Collections is the work nobody wants. A room of agents calling customers about overdue bills, nine to five, reading from a script, taking the same objections all day. Quality varies by agent and by hour. Nights and weekends are uncovered. And the manager can review maybe one call in twenty.

That was the starting point when a Bahraini telecom asked us whether an AI agent could make those calls. In Arabic and English. Politely. Without breaking any rules.

The answer was yes, and the way there taught us more than any other voice project we have done.

Start with the rules, not the model

Before we wrote a line of code, we sat with the collections head and asked questions that had nothing to do with AI:

  • What exactly can the agent say at each severity level, and what are the mandatory phrases?
  • Does Bahrain require telling the customer they are speaking to an AI? What is the company's own policy?
  • What hours are acceptable to call?
  • How many times can we call the same person in a week?
  • What happens when the customer disputes the amount?

Every one of those answers became a hard rule in the system. The agent cannot call outside the window. It cannot exceed the frequency limit. It discloses what it is. It never argues about an amount; it offers to connect the customer to a person.

The lesson: in regulated work, the system prompt is a compliance document first and a conversation script second. We wrote it with the collections team, not for them.

The caller ID problem nobody mentions

A customer in Bahrain who sees an unknown international number does not answer. If they do, they do not believe it is their telecom.

The fix was to verify a number the company already owned as the outbound caller ID, so customers saw a familiar local number. Callbacks to that number went to the company's regular lines, not to the AI. This one decision moved the answer rate more than any change to the voice or the script.

For a later phase, the plan was to connect the agent directly to the company's own telephony system over SIP, which removes the third-party carrier from the picture entirely. Worth knowing on day one, because it changes which tools you pick.

Arabic first, English when asked

The agent opens in Arabic. It identifies the department, states the purpose, and asks how the customer would like to proceed. If the customer answers in English, it switches and stays in English.

Small wording choices mattered more than we expected:

  • Saying the currency the way people say it, "دينار," not the full formal name every time
  • Dropping formal titles that sounded like a government letter being read aloud
  • A calm, even pace, slower than the English version of the same agent

We tested these with staff who take these calls for a living, and they caught things no benchmark would have.

The handoff is the product

The single most important feature was not the AI. It was the transfer.

When a customer says "I want to speak to someone," the agent says, in their language, that it is connecting them, and the call goes to a supervisor's line with the transcript attached. The call stays connected. The supervisor picks up already knowing what was said.

We spent more engineering time on making that transfer reliable than on the conversation itself. Early on, the transfer would ring the supervisor and then drop. Finding and fixing that race condition was unglamorous and necessary. A collections agent that strands customers during a handoff is worse than no agent.

What the numbers looked like

Rounded, because this was a proof of concept and not a full rollout:

  • Coverage: the agent could call during the full permitted window, including hours the human team did not cover.
  • Consistency: every call followed the script at the right severity level. No tired-agent variation at 4:45 PM.
  • Review: the human team could previously audit about one call in twenty. With the agent, every call had a transcript, a summary, and an outcome tag. The review rate went to one hundred percent.
  • Transfers: a meaningful share of customers asked for a person. That is fine. The agent's job was to handle the routine calls and get the hard ones to a human faster, not to replace the team.

The cost model came out well below the per-call cost of a human agent, but the honest headline is the review rate. Being able to read every call changed how the collections team managed quality.

What we would do differently

  • Build the test set from real call recordings before building the agent. We got there, but later than we should have.
  • Treat the telephony integration as a first-class part of the project, with its own timeline. It is not a detail.
  • Put a human in the loop for the first week of every new campaign. Let the team listen live and flag anything odd. Then let the agent run.
  • Decide the AI-disclosure wording with legal on day one. We had to revisit it.

Where this applies beyond telecom

The same shape fits any organization in the Gulf that makes a high volume of routine outbound calls: payment reminders, appointment confirmations, renewal notices, delivery scheduling. The rules differ by sector. The architecture does not.

If you run a team like that and want to see a call, talk to us. We can run one in Arabic or English while you listen.

Ready to Build Your AI System?

Let's discuss your use case and build something that works. AI solutions from Alabama to the Gulf.