Every few months a client shows us a support bot that promised to “handle 80% of tickets” and instead sends customers in circles. The technology has improved enormously since the old decision-tree bots, yet the failures look the same. The difference between an AI agent that delights customers and one that drives them to Instagram complaints is mostly design, not model choice.
What a modern support agent actually is
An AI support agent is a language model connected to three things: your knowledge (policies, FAQs, product information), your systems (order management, CRM, booking engine) and a set of rules about what it may and may not do. Without the first, it guesses. Without the second, it can only talk. Without the third, it will cheerfully promise refunds you never approved.
The useful mental model is a capable new hire on their first week. They are quick, polite and tireless, they read every document you give them, and they have never met your customers. You would let them answer common questions and look up orders, but you would not let them waive a cancellation fee for a VIP without asking a supervisor first.
That is why the quality of your knowledge base matters more than the model you pick. If your return policy lives in three slightly different versions across the website, a PDF and an old email template, the agent will find all three and confidently quote whichever it reads first. Cleaning up and consolidating that content is usually the first fortnight of any support project we run.
Where AI agents genuinely help
The best results we see come from high-volume, low-risk conversations where the answer already exists somewhere in your systems. In the Gulf, a large share of those arrive on WhatsApp in the evening and at weekends, when your team is offline and customers still expect a reply within minutes rather than the next working morning.
- Order and booking status — the agent looks up the order by phone number, reads the courier status from Aramex, SMSA or your own fleet, and explains it in plain language.
- Policy questions — return windows, delivery areas, warranty terms and opening hours, answered consistently in Arabic or English.
- Simple account changes like updating an address, rescheduling an appointment or resending an invoice, within limits you define.
- Triage and summarising — when a person is needed, the agent collects the order number, photos and the problem first, so your team starts with context.
- Drafting replies for staff — on email and complex tickets, the agent proposes a response that a human edits and sends.
That last point is underrated. Many teams get more value from an assistant that sits beside their agents in Zendesk, Freshdesk or Zoho Desk than from one facing customers directly. Response times fall, tone becomes more consistent, and nobody worries about the bot saying something odd in public, because a person reviews every message before it goes out.
Where they struggle or shouldn’t be used
AI agents fail predictably in a few places. Knowing them in advance lets you route those conversations to people from the start, instead of discovering the problem through an angry review or a screenshot shared in a family WhatsApp group.
Emotional and high-stakes conversations
A customer whose wedding cake arrived crushed, or whose medical appointment was cancelled, does not want an efficient summary of your policy. They want to feel heard by someone with the authority to fix it. Agents can detect frustration reasonably well, so use that signal to hand over early rather than attempting empathy at scale.
Exceptions, refunds and judgement calls
Anything involving money leaving your business, policy exceptions or goodwill gestures should be approved by a person, at least until you have months of data showing how the agent behaves. A model can be persuaded by a clever argument; your refund policy should not be. Give the agent the ability to request an exception, never to grant one.
Regulated advice and dialect-heavy voice
Banks, insurers, clinics and property firms face rules about what advice can be given and how it is recorded. Keep agents to factual information there. Voice is another weak spot: written Modern Standard Arabic works well, but phone calls mixing Emirati, Saudi or Egyptian dialects with English still trip up speech recognition more often than text does.
Let the agent request an exception. Never let it grant one.
Designing a handover customers don’t hate
The handover from agent to human is where most projects succeed or fail. Customers forgive a bot that says “I’ll pass you to a colleague” quickly. They do not forgive one that asks them to repeat their order number three times, or that pretends to understand when it clearly doesn’t. Make the route to a person obvious, and never hide it.
Be open about what customers are talking to. A short line such as “I’m the store’s virtual assistant, and a colleague can step in at any time” sets expectations without spoiling the experience. Customers in our projects react far better to an honest assistant than to one with a human name and a stock photo that is obviously not a person.
- Always offer a visible “talk to a person” option, and honour it immediately.
- Pass the full transcript, customer details and the agent’s summary to whoever picks it up.
- Tell customers honestly when a human will reply if it is outside working hours.
- Hand over automatically after two failed attempts or any sign of strong frustration.
How to measure it and what it costs
Containment rate, meaning the share of conversations resolved without a person, is the metric vendors love. On its own it is dangerous, because a bot that refuses to escalate scores brilliantly. Track it alongside customer satisfaction after bot conversations, repeat contacts within 48 hours and how often staff reverse the agent’s actions. Review a sample of transcripts weekly, at least for the first three months.
On cost, a focused agent for one channel, connected to your order system and knowledge base, typically takes six to ten weeks to build and test properly. Running costs combine model usage, which is usually modest per conversation, with WhatsApp Business Platform conversation fees and whatever helpdesk you already pay for. Arabic testing deserves its own budget line, because it always finds things.
We strongly recommend a pilot before a full rollout. Switch the agent on for one channel, perhaps website chat, for a subset of topics such as order tracking, and leave everything else with your team. Two to four weeks of real conversations will tell you more about tone, gaps in your content and customer reactions than any amount of internal testing.
The bottom line
AI agents are now genuinely good at the repetitive half of customer support, and they can make a small team feel available around the clock in two languages. They are not a replacement for people who can listen, decide and apologise properly. Start with one channel and a narrow set of tasks, watch the transcripts closely, and widen the agent’s job only when the data says it has earned it.
