AI that cannot reach your systems is a demo. The work that makes it useful is integration: giving an agent permission to read a CRM record, send a WhatsApp reply, open a ticket or update a pipeline stage, with a clear boundary around what it may do unsupervised. This pillar covers the integrations that carry the most traffic in a typical UAE business, and where a human handoff still belongs.
The model is close to a commodity; the difficulty is everything around it. Connecting an AI layer to a CRM, a messaging channel, an ERP or a booking system means handling each one's authentication, its data model, its rate limits and its particular ways of failing. That work is unglamorous, it does not demonstrate well, and it is where the budget goes.
It is also why a provider who talks fluently about models and vaguely about your systems has told you which half of the problem they are comfortable with. The useful question in a first meeting is not which model they use but what they will do when two of your systems disagree about the same customer.
A great deal of commercial conversation in the UAE happens on WhatsApp while reporting is built on the CRM. Any integration that treats messaging as something to bolt on afterwards has misread how business is actually done in this market: the enquiry, the negotiation and often the commitment all happen in a thread nobody can report on.
Connecting it properly means deciding what gets written back to the CRM and when, who owns a conversation when it moves between people, and — since October 2026 — accounting for the fact that replies inside the customer service window are now charged per message.
Usually yes, and connecting it is almost always the right approach rather than replacing the CRM. Most established systems expose an API, and where one does not there is generally an integration platform that bridges it. The questions that matter are which system is authoritative when two disagree, what the AI is permitted to write rather than only read, and what happens when the CRM changes a field. A programme that opens with a platform migration has substituted an infrastructure project for an automation one.
Yes, through the WhatsApp Business Platform rather than the phone app, and within Meta's rules — a business using AI to serve its own customers is permitted, while general-purpose assistants are not. The design decisions that matter are what it will and will not discuss, how it hands over to a person, whether it can see the customer's history in the CRM, and how it handles Arabic and English in the same thread. Since 1 October 2026 replies inside the 24-hour service window are charged per message, so resolving in one message is now a cost decision as well as a service one.
Administrative access to the systems involved, a decision about which one is authoritative for each field, and someone who can describe the process including the exceptions everybody works around. On the content side, the answers the AI will give have to exist somewhere maintained — if the correct response lives in three people's heads, assembling that is the first phase of the project rather than an afterthought.
It can, which is why write access should be scoped narrowly and introduced after read access has been running for a while. The sound pattern is that the system proposes and a person approves for anything consequential, with full logging of what was written and by which action, so a bad entry can be found and reversed. Start by letting it read and draft; add writing one field at a time.
More often than not there is a route — an export, a database connection, an integration platform that already supports it, or a scheduled file drop. The question to settle early is how current the data needs to be, because a nightly export is a legitimate answer for reporting and a poor one for a customer-facing agent that must know whether an order shipped this morning.