Custom AI Assistants in the UAE 2026: What a Company-Trained Assistant Does That a Chatbot Cannot

"AI assistant" is used interchangeably for two products that share almost nothing. The first is the generic chatbot: a scripted decision tree wearing a friendly avatar, deployed to deflect questions, with a fixed tree of branches and a fallback "please email us" at every leaf it did not anticipate. The second is a trained assistant: a language model grounded on your company's actual knowledge — your services, prices, policies, documents — with guardrails defining what it may say and a handoff path to your people for everything that needs human judgement. The first answers from a script; the second answers from your business. The practical difference shows up within a week of launch: the scripted bot frustrates anyone whose question bends slightly, while the trained assistant handles the bent questions because it is reading, not matching. Disclosure for this guide: Sijil builds trained assistants as a service — our own Maryam, who consults on UAE company formation over WhatsApp, is the reference build — and every price here comes from our published service list.

The failure everyone fears from AI is the confident wrong answer — a price invented, a policy hallucinated. Grounding is how a serious build prevents it, and the mechanism is unglamorous: the assistant answers only from a knowledge base you control, and when the question falls outside that base, it says so and routes the conversation to a human instead of improvising. This has two consequences for buyers. First, the quality of the assistant is capped by the quality of your knowledge base: if your prices live in three outdated spreadsheets, no model fixes that — the content work is part of the project whether the vendor says so or not. Second, scope discipline is a feature: our starter scope deliberately includes a single knowledge base and a single channel, because an assistant that knows one domain thoroughly outperforms one that half-knows five. Maryam's knowledge base, for comparison, is Sijil's verified free-zone data — signed partner agreements and official price lists — which is why she can quote a zone's total fees precisely and cite where the number comes from. Your assistant should know your equivalent of that, and nothing else, just as precisely.

A trained assistant needs rules as much as knowledge. Guardrails define identity (how it introduces itself — an honestly-built assistant never claims to be human), boundaries (what it may promise, how it quotes prices, what it says about competitors), escalation (which triggers hand the conversation to a person), and language discipline (your register, your terms, your banned phrases). Maryam's guardrails, for instance, require her to answer plainly that she is an AI consultant when asked and to name the human Registrations Officer who takes over execution — a design we consider non-negotiable and one that aligns with UAE consumer-protection expectations and WhatsApp platform policy. The handoff is the other half of the design: every business has conversations that should never be handled by software — negotiation, complaints, judgement calls — and a well-built assistant recognizes them and transfers with the full context attached, so your team starts from the conversation, not from a summary phone call. An assistant that tries to close everything itself is not ambitious; it is mis-scoped.

Our flagship custom AI assistant starts at AED 18,000 for a starter scope, delivered in three to six weeks by Bloom Techno, the in-house engineering team behind sijil.ae and Maryam. Starter scope means, precisely: one channel (WhatsApp or web), one knowledge base, one language, lead capture with structured records, human handoff, and guardrails aligned to your brand. The weeks break down honestly: content and knowledge-base preparation first, guardrail design with you second, build and integration third, testing against real conversations fourth — the knowledge base is usually the schedule's critical path, not the code. What moves a project above starter scope: additional knowledge bases, additional languages, additional channels, CRM or booking integrations, and volume tiers for conversation traffic. Running costs exist and belong in your budget: model usage is metered by the AI provider, and WhatsApp's Business Platform prices conversations by category with rates that change — confirm current rates with the providers. Market observation, not a verified figure: custom AI builds in the region are quoted anywhere from the low tens of thousands of dirhams to multiples of that, and the quotes that omit content work and running costs are the expensive ones. If the company that will run this assistant does not exist yet, start with our AI company setup guide, the AI free-zone comparison, or the software company guide.

A trained assistant is a long-lived asset: it keeps answering, capturing and handing off as long as the knowledge base under it stays true — which makes maintenance part of the purchase decision, and we say so rather than pretend it is a one-time build. Start the conversation at /services/custom-ai-assistant — from AED 18,000, built in-house by Bloom Techno, the team behind Maryam herself. And because the assistant serves the company, the company's own standing still decides how far trust goes: licence and establishment-card expiry, visa dates, VAT quarters and the corporate-tax return are tracked as dated obligations with a health score in Sijil's client portal. Talk to the team on WhatsApp — you will be talking to Maryam first, which is the honest demo of the product you are considering.