An AI knowledge base for your Dubai agents: answers on process, forms and developer terms

The questions your agents ask you every week

Before any software, collect the evidence. Scroll back through your agents' WhatsApp (messaging app) group and your own messages for the last month and write down every question that was asked more than once. You will usually find the same families:

• Which contract or form do I use for this situation, and who signs what?
• How do I get a permit for this advert?
• What is our split on this type of deal, and when do I get paid?
• What is the payment plan on this developer's project, and where is the brochure?
• How do I register this on the portal, and why was my listing rejected?
• Which documents do I collect from a new client?

Count them. That count is your baseline, and it is the number you will use later to see whether the assistant works. A tool that cannot reduce this count is not worth paying for.

What goes into the knowledge base (and what must stay out)

An AI knowledge base is a collection of your documents plus an assistant that finds the relevant passage and drafts an answer from it. It answers from what you gave it. So the quality of the answers is the quality of the documents.

Put in:

• Your step-by-step processes (listing a property, onboarding a client, closing a deal), written as short procedures.
• Your commission and split policy, as the written policy, not individual deals.
• Guidance on forms, permits and portals, written by you from the official sources.
• Developer summaries: payment plans, handover dates, brochure links and the rules of each launch, each dated.
• Templates: emails, client messages, checklists.
• An FAQ of the questions you counted above, with the answers you have already given.

Keep out:

• Client passports, Emirates ID copies, bank details and anything else personal. See client data and the UAE data protection law and secure document storage.
• The pricing and negotiation details of individual deals.
• Agent salary or split information for named people.
• Anything you have not reviewed. A wrong document in the knowledge base becomes a wrong answer, repeated to every agent.

Dubai process content: forms, permits, listing rules

This is where a Dubai brokerage gets value that generic tools cannot give, because you are the one who knows how your office actually works. Write short documents for each recurring task, based on the official source and your own practice:

• Which contract applies. The Dubai Land Department's 2024 brokerage practice guide lists Contracts A, B and F, the standard forms for brokerage dealings. Summarise from that guide in your own words and link to it. I would be careful with any other form name circulating among agents; for example, I could not confirm that Form I appears in the official guide, so do not load unofficial form guidance into the assistant as if it were policy until you have checked it. For the paperwork around a sale, see the deal file checklist.
• Advertising permits. The practice guide says real estate advertising needs a permit through Trakheesi (the property advertising permit system) with the number shown. A one-page procedure: who applies, what is needed, where the number goes.
• Listing rules for each portal. What your agents must provide, in what order, and your internal checks before posting.

Every one of these documents will go out of date. Which brings me to the next section.

Commission policy and developer terms: documents with owners and dates

Developer terms change often, and an assistant that quotes last quarter's payment plan with confidence is worse than no assistant. The best practical advice I found on this was from a consultancy: give every item an owner, a last-reviewed date and an escalation contact. I would add a "valid until" date for anything time-sensitive.

For example, a developer entry might read: Project name; payment plan summary; handover date as stated by the developer; brochure link; owner of this entry: the off-plan manager; last reviewed: a date; valid until: a date; ask: the manager's name. If the valid-until date has passed, the assistant should say so in its answer and send the agent to the owner. See how to track developer terms and allocations for the ledger behind this.

Commission questions need the strictest rules. The assistant can quote the written policy. It must not interpret it for a special case, and it must not compute anyone's payout, because the arithmetic and exceptions belong to a person. For the policy side, see commission split models.

How the AI answers: with the source, or it says it does not know

This is the design decision that matters most. Insist on four behaviours:

1. Every answer shows its source (document name, section and last-reviewed date), so the agent can open it.
2. It answers only from your documents. If nothing relevant is found, it says "I do not have this, ask" followed by a named person.
3. It escalates by topic. Legal, compliance, commission exceptions and anything involving a client dispute go straight to a human.
4. It never guesses a figure. Fees, deadlines and percentages come with a source or not at all.

Test the "I do not know" behaviour on purpose. Ask it five questions you know are not in the documents. If it answers confidently, the setup is wrong. Also read the limits and consent risks of AI for UAE teams before rollout.

Who sees what

Not every document is for every person. A new agent should see processes and templates; a team leader may see commission policy; only owners and managers should see financials. Most tools let you set access by group or role. Ask this in any demo: can I limit documents by role, and can I see who asked what? Logging matters so you can find which questions the assistant could not answer, which is your list of documents to write next.

Keeping it current

• Every document has an owner. No owner, no entry.
• Review dates are real: the assistant flags anything past its date.
• When an agent says an answer was wrong, the fix is to correct the document, not the chat.
• Once a month, read the unanswered questions list and write the top three missing documents.

If your documents do not exist yet, do not wait. Writing the procedure for the top ten questions is the real work, and it helps your office whether or not you ever turn on AI. Tools that help draft SOPs (standard operating procedures, step-by-step written instructions) from notes are covered in the AI SOP generator guide.

How the tools work: your options and the limits

There are three routes.

• A ready-made knowledge base tool. You upload documents and get a chat window. Several such tools exist; the vendor pages I read were US-focused, and their claims about going live in a week, or about results, are vendor claims I could not check. Ask any vendor where the data is stored and whether your documents are used to train models. My comparison of AI knowledge base tools goes through them.
• Build it yourself in a general AI assistant. Claude and ChatGPT (two general-purpose AI assistants) both let you attach documents to a project and ask questions. Cheap, quick to test, but access control and logging are limited.
• A custom assistant connected to your documents and your WhatsApp or internal company website, with roles, logs and your own escalation rules. I build these for businesses; it makes sense once the simple version has proved that agents use it. For the mechanics, see how to build an AI knowledge base from your documents and letting staff get answers from SOP documents.

Try it yourself with Claude, an AI assistant. Take your commission policy document (a policy, not a deal) and paste it in with this instruction: "Answer only from the text I pasted. Quote the sentence you used. If the answer is not in the text, say 'not in the policy, ask the owner'. Question: if an agent leaves before a deal is paid, what happens to the commission?" Then ask something the policy does not cover and check that it refuses. If it invents an answer, you have just learnt what to guard against.

A first-month rollout for a 20-agent team

This is a suggested sequence, not a benchmark.

1. Week 1: count the repeat questions, pick the top 20 to 50, and note who answers each one today.
2. Week 2: write or tidy the documents for those questions. Add owner and review date to each.
3. Week 3: load them into a test tool. Try five agents, with five known-unknown questions to test refusals.
4. Week 4: open it to all 20 agents with a clear message about what it is for and what it is not. Log every question.

How to tell if it works

Use your own baseline. Count the repeat questions in the agents' WhatsApp group in the four weeks before, and in weeks five to eight after. Track how often the assistant said "I do not know" and how many of those turned into new documents. If agents keep asking in the group anyway, ask them why: usually the answer is too slow, too long, or wrong once. No independent Dubai figures exist for how much time such a tool saves, so rely on your own numbers.

How would I approach it?

I would start with the 20 most repeated questions, write the answers into documents with owners and dates, and test with Claude or a ready-made tool before spending anything. Only if agents use it and you need roles and logs would I consider a custom build. The real work is the documents, and that stays with you.