Dubai Chambers launches Chamber AI Hub: 22 agents now, 100 planned for 295,000 companies

Agentic AI has just moved from a build decision to a buy decision. Dubai Chambers now fronts 22 ready-made agents to 295,000 registered companies, with sandbox trials, introductory credits and usage-based pricing. When a business leader can subscribe to an invoice-matching agent before lunch, the question at your next steering committee is no longer whether to use agents, but what is still worth building. The honest answer is narrower than most CTOs assume: buy commodity admin work (invoice-to-purchase-order matching, CV screening, lead scoring), and keep anything that touches your contracts, your customer data or your differentiating process on infrastructure you control.

Read the catalogue before you applaud it. The opening set covers seven functions (compliance, procurement, operations, finance, sales, HR, logistics) and does concrete work: supplier evaluation, contract review, invoice matching against purchase orders, proposal drafting, CV screening and first-line IT helpdesk. Per-agent dashboards report usage, cost and request performance, an admin portal shows how staff use the agents, and data is hosted in the UAE with each company's information kept separate. That is a sensible shape for SME and mid-market adoption, and the two-year target of 100 agents shows this is distribution policy rather than a demo. For a firm with no in-house AI team, it is now the cheapest way to test agentic automation against real admin load.

The boundary is where most teams will get it wrong. Subscribed agents plus internal pilots sprawl until nobody can say which system sent which document to which model, or who paid for it. Put one control point in front of every model call your organisation makes, subscribed or built: an OpenAI-compatible gateway such as our own FastLLM Proxy, which sits in front of your own LLM servers and 80 hosted providers and keeps routing, budgets and access control in one place. One endpoint, one key per application, per-model grants and a monthly spend budget. Route sensitive workflows (contract review, HR screening, anything carrying customer data) to models on your own GPUs running a serving engine such as vLLM or SGLang, and allow hosted providers only where you have decided, in writing, that failover is acceptable. That pairing of gateway and self-hosted models is the AI infrastructure work we design, build and run, with planning, deployment and support done by our team in the UAE.

Applications then never talk to a provider directly. The gateway is the only place routing, keys and spend live:

from openai import OpenAI
import os

One endpoint for every app and every agent: routing, budgets and access control live in the gateway. client = OpenAI( base_url="https://llm-gateway.internal/v1", api_key=os.environ["CONTRACTS_APP_KEY"], # per-app key: per-model grants, monthly spend cap )

resp = client.chat.completions.create( model="contract-review", # logical name; gateway policy picks the backend (own GPUs first) messages=[{"role": "user", "content": clause_text}], )

The GCC angle is distribution plus control. Dubai is the first market where a chamber of commerce is effectively reselling agentic automation, and within a year your board, your counterparties and your regulator will assume you hold a position on it. The real risk is governance debt: every subscribed agent that reads your ERP or your mailbox needs an owner, a data classification and an evaluation set before it runs unattended. If you want a scaffold for that, the voluntary NIST AI Risk Management Framework is published in Arabic as well as English and its generative AI profile, NIST AI 600-1, covers the risks that unattended agents introduce. Do not accept a marketplace's success rates as your own. Baseline the manual task first (minutes per invoice, hours per contract) and hold each agent to that number. Teams that buy commodity workflows from the Hub and run everything sensitive on a governed internal platform will shift Arabic and English admin load fastest, while peers are still staffing pilots.