Choosing the right AI model for customer support
Not all AI models are equal when the task is answering a frustrated customer writing in dialect who expects a reply in 30 seconds. Here is how we think about model selection.
The tradeoff that matters most
When you are routing a customer support message through an AI model, you are making a tradeoff between three things: response quality, latency, and cost. The right balance depends on the message — and no single model wins on all three.
For the MENA market, there is a fourth dimension: multilingual fidelity. A model that handles English well but produces stilted formal-MSA Arabic is not a good fit for a Levantine or Gulf support team.
How we categorise messages
At Orqelo, we use a routing layer that classifies each incoming message before sending it to a model. The categories:
Simple / transactional. "What is my order status?" "When does the store open?" These have a narrow answer that can be looked up. Fast, cheap models (Claude Haiku 4.5) handle these well. The goal is sub-two-second response time.
Nuanced / advisory. "I am not sure which plan is right for my team" or "I want to escalate a billing dispute." These need reasoning, context, and tone calibration. We route these to Claude Sonnet 4.6.
Complex / escalation-risk. High-stakes conversations: a customer about to churn, a legal query, a complaint that has already been mishandled. Claude Opus 4.7 handles these — the quality difference is worth the cost when the alternative is losing the customer.
What we learned about Arabic
Dialect matters more than we expected. A model that handles Modern Standard Arabic well does not automatically handle Levantine or Gulf dialect correctly. We evaluate model responses on a set of dialect-specific test prompts before updating the routing weights.
Transliteration is the other gotcha. Customers writing in Arabizi (Arabic in Latin characters) send prompts that look like noise to a model not trained on it. Our preprocessing layer normalises Arabizi before it hits the model, which improves response quality across all tiers.
The practical routing rule
We give every AI response a quality score (a lightweight evaluator model runs asynchronously). When the score falls below a threshold two messages in a row, the conversation is flagged for human handoff. The AI is the first line; the human is the last line.
The goal is not to replace human support agents. It is to make sure agents only handle conversations that genuinely need them.