Most support automation is sold on deflection: the share of conversations that never reach a human. It is the wrong number to optimise, because the cheapest way to raise it is to make escalation difficult, and a customer who could not reach a person does not experience that as efficiency.
What genuinely works
- Status questions. Where is my order, when does my booking start, what is my balance. These are lookups, the answer is unambiguous, and customers prefer an instant reply to a queue.
- Filling in the form. An agent that collects the details, identifies the account and hands a human a complete case removes the tedious half of the job and is invisible to the customer in a good way.
- After hours and second language. An accurate answer in Arabic at midnight beats no answer, and beats a human who is guessing.
- Answering from your own documentation. Grounded in your actual policies, with a citation, so an agent can verify it.
What backfires
Complaints. Anything involving money moving in the wrong direction. Anything where the customer has already said they are frustrated. In each case the automation is not solving the problem, it is delaying the conversation that solves it, and the delay is the injury.
Also: pretending. A bot presented as a person is a short-term trick that costs trust permanently when discovered, and it is always discovered.
The escalation rule
One clear route to a human, visible from the first message, that works. Not buried, not conditional on repeating yourself, not a form. If the automation is good, few people will use it. If it is bad, you want them to.
Measure the right things
Resolution without recontact within seventy-two hours, not deflection. Escalation rate is a diagnostic, not a failure: a rising rate on one topic tells you exactly which article to write or which process to fix. And read the transcripts. Weekly, by a person who can change something. The transcripts are the most useful customer research most companies own and never look at.
The Gulf specifics
Bilingual is a baseline, and code-switching mid-sentence between Arabic and English is normal here and breaks naive language detection. WhatsApp is frequently the primary channel rather than a secondary one, which changes the expected response time and the tone. And the phone call is still the escalation people expect, so the route to a human should end at a number that someone answers.
