INSIGHTS / AUTOMATION
Afirm Labs
We helped a mid-size holiday homes company in Dubai put AI on their WhatsApp. Guests now get a reply in under two minutes, at any hour, with answers taken from live data in their PMS (property management system).
The AI works on its own only as far as the team is comfortable with. When a chat needs a person, it hands it over, and the team can see and change everything.
This case study covers the project from the first audit to launch: the decisions we made, the problems we ran into, and what any business owner can take from it, even outside holiday homes.
At a glance
The company: a mid-size holiday homes operator in Dubai with around 30 properties. Most guests find them on Property Finder and talk to them on WhatsApp.
The problem: every guest inquiry went to one phone.
What we built: one shared inbox for the team, and an AI that answers guests on WhatsApp using live data from the PMS. When it needs a person, it hands the chat over.
Today: live, guests get a reply in under two minutes, at any hour.
When a business outgrows one phone
In the UAE, WhatsApp is the main way customers reach a business, and for holiday homes it is the front desk. Guests rarely call, email or fill in a form on a website. They tap "Inquire on WhatsApp" on a listing and start typing. When the company was smaller, every inquiry went to the CEO's phone, and that worked well: he knew every property, price and house rule.
At around 30 properties, the same setup holds the business back. Every inquiry still lands on one phone, and most questions are simple: dates, prices, location, deposit. Each answer means checking the PMS and typing it back, and anything for the team gets forwarded as screenshots. One person can't be in a meeting, at a property and on WhatsApp at the same time. And the longer a lead waits for a reply, the more likely they book somewhere else.
The audit: everything on the table
We didn't start with AI. We started by looking at how the whole business runs, because the process that annoys you most is not always the one that pays off first.
Over a few sessions with the CEO and the team, we went through every area step by step: guest communication, maintenance, cleaning, vendors, finance operations and finance analytics. For each step we noted who does it, which tool they use, how long it takes and how often it happens. Then we ranked everything by what it costs the business and how much it matters.
Guest communication came out with the highest and fastest ROI. The reasons apply to almost any business choosing its first automation:
It's close to revenue. Every inquiry is a possible booking. Saving time on internal work cuts costs, but answering guests faster can bring in money.
It happens many times a day. A few minutes saved per conversation adds up fast.
The answers already exist. Prices, dates, deposits and locations are already in the PMS. The AI doesn't need to invent anything, only find the right fact and say it clearly.
It's safe to hand back. When the AI isn't sure, it can stop and pass the chat to a person instead of guessing.
The other areas didn't go away. Maintenance, cleaning, vendors and finance are still on the list, in order, for later.
Control first: one inbox for the whole team
Before any AI, we changed who can see what. All guest chats now arrive in one shared inbox, where the whole team sees every conversation and who replied. Guests still talk to one company number.
This order matters, because an AI replying in a chat nobody else can see is a risk. In a shared inbox, every message it sends is visible to the team from the first day.
The inbox also works on Android and iOS, so the team can reply from anywhere, not just from a desk. AI can be switched on or off for each chat, so a person can take over at any moment. The inbox is self-hosted rather than rented, which makes it faster and cheaper to run, and keeps guest data under the company's control.
WhatsApp: the part most companies give up on
To answer guests automatically, you need the official WhatsApp Business API from Meta. We often hear from companies that tried to set it up themselves and gave up. It involves business verification, approvals and strict rules on when you can message a customer.
The rule that surprises most people: if a guest's last message was more than 24 hours ago, you can't send them a normal message. You can only send a template that Meta has approved in advance. So first contact and follow-ups each need their own approved template.
We wrote templates for each situation the business needed and got them all approved. The AI knows which one to use and when. Property Finder inquiries now flow in automatically too, so the AI picks up each new inquiry the moment it arrives.
How the AI is built
Most AI assistants are one long instruction to a model, plus as much information as fits. That is fine for a demo. In real use it gets slow and expensive, it is hard to test, and fixing one answer often breaks another.
We split the work into separate steps. One finds which listing the guest means, one decides what information is needed and where to get it, one writes the reply, and a guardrail step checks it before it goes out. Each step uses the model that fits the job: a small, fast one for simple work, a stronger one where it matters.
Each step is tested on its own, against a standard that fits its risk. A price or a date must be exactly right every time, while the wording of a reply has more room. When something goes wrong, we know which step to fix, and the rest keeps working.
Wherever an answer can be exact, it doesn't come from the AI's memory. Prices, availability, deposits and listing details come straight from the PMS, and the AI's job is to understand the guest and explain the answer clearly. The guardrail step makes sure no sensitive or internal information reaches a guest, and it was tested thoroughly before launch.
We built and tested everything on the company's past guest conversations. Real guests write in bursts, leave out details and negotiate, and no made-up example covers that. It's also why the AI waits a moment before replying: it reads the whole burst, runs its checks once, and answers everything together.
A lesson learned: when no-code tools stop being enough
We started by building a prototype with automation tools like n8n. They are great for a first version: you can connect systems quickly and see if an idea works.
But past the prototype stage, we hit their limits. With many steps, several models and real guests waiting, we needed each step tested on its own, fast replies and behaviour we could predict. That is hard to do well in a no-code tool, so we rebuilt the core as a custom system.
The lesson isn’t that no-code is bad. Use it to test an idea cheaply. Once the business depends on it, build something you can test and trust, especially once it connects several systems and runs a lot of custom logic.
As autonomous as the team is comfortable with
Not everyone trusts AI with their customers, and they shouldn't have to. Here, the team decides how much the AI does on its own: no more, no less.
It works within strict rules set with the team, and suggests alternatives when dates are taken or a guest pushes on price. When a guest is ready to book or the AI isn't sure, it hands over: the CEO gets a WhatsApp alert with the reason and a link to the chat.
Anyone can take over a chat or switch the AI off for it, and every conversation is tracked in real time. The rollout works the same way: a few listings first, more as the team sees it working.
Where it stands now
The system is live, and here is what changed:
For guests: answers day or night, without waiting for one person to be free.
For the CEO: he is pulled in only when a guest is ready to book or something needs him.
For the team: one shared view of every chat, and nothing gets forwarded as screenshots.
For the business: data it owns, showing what guests ask most, which listings get the most questions and where conversations go quiet.
"Now the AI answers most questions on its own, and I only hear about the ones that need me. The best part is that it works with our existing systems."
— CEO
What any business can take from this
Five principles we follow on every automation project:
Map the work before choosing a tool. The best first project is usually close to revenue, repeated many times a day, and answered from data you already have.
Give people visibility before adding AI. When the team can see every conversation, mistakes are caught early and trust comes sooner.
Keep facts out of the AI's memory. Prices, dates and policies should come from your systems. Let the AI do the talking, not the remembering.
Decide how much the AI does on its own. Plan the handover first, start with rules you're comfortable with, and widen them as you see it working.
