Algeria AgencyArtificial intelligence
AI for an estate agency: sort the enquiries, do not write the listings
An agency does not lose sales for want of copy. It loses them in Tuesday morning’s enquiries, arriving through four doors and handled by nobody.
Tuesday, 9:40, an agency in Chéraga. Eleven messages on the agency phone, four calls missed during the 9 o’clock viewing, three forms from a property portal, and two comments under yesterday’s post. Seven of those twenty enquiries are about the same flat, two are from the same person, and one is a competitor checking a price.
That is where the work happens, and it is where the artificial intelligence tools sold to the sector do not look: an agency is offered copywriting for its listings, when the first photograph already does all the sorting. What is missing is not text, it is an order of play.
This article describes what a system can do with inbound enquiries, what it must never decide on its own, and why it should not write your listings. It does not cover matching your buyer list against a property coming in: that is the other direction of the work, and it is covered elsewhere.
Tuesday morning, and how many doors it comes through
Do the inventory once, over a week, and note only where each enquiry came from. An average Algerian agency receives through four to six channels: the landline, the agency mobile, the messaging app, portal forms, comments and private messages on social networks, and sometimes somebody walking in.
Those channels share neither a waiting time nor a trace. A missed call leaves only a number, a private message leaves text that gets buried under the next ones, a portal form lands in an inbox nobody opens on Saturday. The phone is still the counter, and the mobile share the ARPCE measures — 88.71 % of internet subscriptions in Algeria — says clearly enough that the written enquiries will arrive from a phone too, out of order.
The consequence is that your problem is not volume but dispersion. Twenty enquiries spread across five places cannot be compared with one another, so the most recent and the loudest gets answered, not the one closest to a transaction.
Before any tool, this inventory is worth doing by hand: it gives the number of doors, and the number of doors is what decides the difficulty of everything else. An agency with only two channels needs half of what this article describes.
The first move is not to qualify, it is to acknowledge
Property enquiries are comparative by nature: the same person writes to three agencies about the same neighbourhood within the hour. What holds them is not the quality of your answer, it is having had one while they were still looking.
An acknowledgement is not a commercial reply. It says the message arrived, it names the property so the person knows it was read, and it says when an agent will call back. It is the only message a system can send on its own without risk, because it commits to nothing.
The trap is turning it into an automated conversation. As soon as the system answers the second question it has to say things about a property — availability, room to negotiate, a viewing date — that are wrong the moment a colleague moves the file. An acknowledgement that promises availability is worse than silence, because it creates an expectation nobody will meet.
The rule that follows is short: the system writes what stays true whatever happens next, and nothing else. That fits in one sentence, and that sentence is the only thing you have to approve.
What a model extracts from a message, and what it must not guess
On a free-form message — "hello, is the two-bed in Bir Mourad Raïs still available? looking for September, budget around two, credit possible" — a model does one useful thing and only one: it puts what was said into named boxes. Property concerned, timing, budget mentioned, financing hinted at, channel, language of the message.
That work is reliable because it is restatement, not judgement. The model does not decide whether the budget is realistic, whether the person is serious, or whether the credit will come through. It turns text into columns, which lets you compare twenty enquiries with one another for the first time.
The distinction holds at one precise place: what was not said stays empty. A system that fills a box by inference — "probably a first-time buyer", "strong intent" — produces data nobody supplied, and that data will then decide a calling order. This is the moment the tool starts being wrong in a way you cannot see.
So ask for two columns rather than one: what was said, and what the system assumed. If the second column is empty in the demonstration you are shown, ask to see the same thing on your own messages, where it will not be.
The same buyer writes three times: deduplicate first
Before any qualification, the first saving is the duplicate. The same person calls, then writes on the messaging app because nobody picked up, then fills in a portal form in the evening. You have three lines, three agents calling back, and a buyer who thinks the agency is disorganised — which at that moment it is.
The matching runs on the phone number where there is one, and that is the easy case. The hard case is the portal form, which sometimes hides the number, and the social message, which carries none. What is left is the property, the timing, and the wording.
A model is useful here, provided it proposes and does not merge. Two close enquiries appear side by side with what brings them together, and a human confirms. An automatic merge makes a real enquiry disappear the day two different people want the same two-bed in the same neighbourhood, which is the normal case and not the exception.
It is also the easiest part to measure: count, over a week, how many of your enquiries are duplicates. Under 10 % the tool has almost nothing to offer you on this side; above a quarter, you have found your first hour saved every day.
One clarifying question, and which one
Many enquiries arrive with one missing piece that prevents classification. A system can ask a question, one only, and the choice of that question is a commercial decision rather than a technical one.
The temptation is to ask for the budget. It is the worst one: it looks like a filter, it puts the person in the position of being judged, and the answer is often inaccurate because it is given defensively. The question that really sorts is timing — when are you looking to move — because it is not threatening and it separates curiosity from a project immediately.
The second useful question, when the timing is near, is about financing: do you have the amount, or are you going through credit. It changes the calling order and the time you will invest, and it is asked without awkwardness if it comes after timing rather than before.
One question is enough because the next one belongs to the agent. An automated five-point questionnaire drives half the enquiries away and gets polite answers from the other half; what you obtain then is not information, it is a filled form.
Route to a person, not to a shared inbox
A sorted enquiry that lands in a shared inbox has not been handled, it has been moved. The routing rule is what turns sorting into work: every enquiry carries an agent’s name, a callback deadline, and a state with only three values — to call, called, closed.
Routing runs on the property, not on the workload. Whoever holds the mandate knows the file, the co-ownership, the viewing constraints; sending the enquiry to the first person free loses more time internally than it saves at reception.
One written exception is needed: absence. If the agent holding the mandate is on leave or out on viewings until evening, the enquiry moves after a set delay to a named person. That delay is a setting you must be able to change yourself, because it depends on the season and not on the tool.
None of this mechanism is intelligent in the technical sense, and that is precisely the useful observation: on this subject half the benefit comes from simple rules applied without exception, and only the other half from the extraction above.
The enquiries that are not enquiries
An agency’s inbound flow contains a constant share of approaches that lead nowhere, and recognising them beats handling them politely to the end. There is cold selling — insurance, surveys, search rankings — the price check by a fellow agent, the free-valuation request with no intention of a mandate, and the person looking to rent when you only do sales.
A model classifies those cases correctly most of the time, because their wording is stable. But the consequence of an error is not symmetrical: filing a real buyer under cold selling costs you a sale, filing a cold seller under buyers costs you three minutes.
The design that follows is never to delete, only to demote. A "probably off-topic" category that stays visible and is reread in two minutes at the end of the day gives the same time saving without the risk, and within a few weeks it teaches you what your flow actually contains.
It is also the only category where we advise letting the system decide alone, because the decision is reversible and the cost of its error is bounded. Everywhere else in this article, the proposal goes past a human.
What the system never answers on its own
Three things, and they are not negotiable. A property’s availability, because it changes within the day and often through a channel the tool cannot see — a sale agreed at the notary, an owner withdrawing a mandate. The price and the room to negotiate on it, because that is the agency’s commercial position and it belongs to the agent holding it.
The third is subtler: the promise of a viewing. A viewing commits the owner, the occupant if there is one, and an agent’s diary. A system offering slots from a synchronised calendar does less damage, but it will keep offering a viewing for a property whose mandate has just expired.
The general rule behind those three cases is that the system must never assert a state of the world it cannot verify at the moment it writes. It can say what is published, say that an agent will call back, pass an enquiry on. It cannot say what is true right now about a property.
Put that way, the limit is not a lawyer’s caution: it is what separates a tool that saves you time from one that produces cancelled appointments and buyers who travelled for nothing.
Why it does not write your listings
This is the use you will be sold first, and it is the one we advise against. Three reasons, in this order. The first is that the photograph decides: better-turned text does not make up for a listing whose images show neither the light nor the layout, and effort put into the text is almost always effort stolen from the photographs.
The second is legal and commercial at once. A listing commits the agency on floor area, storey, aspect, condition, charges. A model writing "bright" about a north-facing flat, or "close to all amenities" without checking, produces a statement the mandate obliges you to stand behind. Correcting it costs more than the writing saved.
The third is that the same model, on the same property, will write six near-identical texts for six portals, and multiple publication is already a known problem in the trade. Near-identical repeated text is what search engines handle worst, and what gives a buyer the feeling of having seen the property already.
What the model can do on a listing is narrow and honest: check that no mandatory mention is missing, flag an inconsistency between the stated area and the number of rooms, spot a word you have forbidden yourself. Checking, not producing.
An enquiry is personal data, and so is the trip abroad
The content of an enquiry — a name, a number, a budget, a guessable family situation, sometimes an address — is personal data, and automated processing of those messages falls within law 18-07 of 10 June 2018, amended and completed by law 25-11 of 24 July 2025.
Two practical consequences for an agency. The first is that the processing has to be described somewhere: what you collect, why, how long you keep it. That is what the register is for, and it is filled in on a real case in one morning.
The second is that if the extraction runs through a service hosted abroad, the message goes with it. The regime applying to the transfer and the architectures that avoid it are covered by a dedicated article in this same set, and the question arises before the tool is chosen, not after. If the sorting were ever to touch sensitive situations rather than search criteria, the impact assessment is what then decides whether the project starts.
This paragraph describes the state of a regulation at the date of publication and does not replace a lawyer’s advice. It is here because an agency installing automatic sorting for its messages has, on that day, begun automated processing of personal data, often without knowing it.
The week’s check: twenty enquiries, four columns
Take the last twenty enquiries received, across all channels, and fill in four columns by hand: the channel it arrived on, the property concerned, the time of the first reply, and the current state. One hour of work, a sheet of paper, no tooling.
Three numbers come out of it, and they are the only ones that decide. How many duplicates; how many enquiries never got a first reply; and the median delay of that first reply, compared across channels. An agency usually discovers its delay is minutes on the phone and hours everywhere else.
The reading is direct. If you have few duplicates, few misses and an even delay, there is nothing to buy: your organisation holds, and a tool would only add one more interface to watch. If the misses concentrate on a single channel, the answer is to close or redirect that channel, which costs nothing.
Only when the twenty lines show misses scattered across every channel, with duplicates, is automatic sorting justified — because the problem is then dispersion, which is exactly what the tool addresses.
What we do, and what we refuse
We bring the channels into one list, wire in extraction in two columns — said and assumed — write the routing rules and the acknowledgement message, and leave the duplicate-matching threshold adjustable by you. All of it connects to what you already use, without changing your trade software.
We refuse to write your listings, for the three reasons in section 9, and we refuse to let the system announce an availability, a price or a viewing. Those refusals are not start-up precautions to be lifted later: they are the conditions under which the rest works.
We promise no conversion figure. We have none measured for this trade in Algeria, and a percentage carried over from a foreign market would describe an enquiry flow with neither the same channels nor the same delays. What we know how to count is what the sheet in section 11 counts, and it belongs to you.
What you can do without us is the most profitable part: close a channel nobody watches, write the acknowledgement by hand, and name an owner per mandate. Many agencies find most of what a tool would sell them right there, and the rest depends on the photograph.
Frequently asked questions
Do we need trade software before doing this?
No, and the reverse order is often healthier. Sorting inbound enquiries runs on the tools you already have, and it teaches you what your flow contains. Choosing trade software after that step gives you a specification based on facts rather than on a demonstration.
Can the system answer in Arabic and in French?
Yes, and it should answer in the language of the message received, not the agency’s. That is a rule to write down explicitly, without which the reply follows the interface language. The produced text then has to be checked on screen, because Arabic display breaks in specific places.
How long before it is usable?
The rules part — channels gathered, acknowledgement, routing, states — goes in within days and delivers most of the gain. The extraction then needs tuning on your own messages, and it takes one to two weeks of real use before you know what it misses.
What about rental enquiries if we only do sales?
Recognise them and answer once, clearly, saying it is not your business. It is the simplest case to handle automatically and one of the most profitable: it takes up a visible share of the flow and requires no judgement.
Does this replace a negotiator?
No, and a tool claiming it describes the trade badly. What is replaced is the time spent working out who wrote what, on which channel, about which property. The first call, the viewing and the negotiation remain the work, and they are what decides.
Our competitors already use automatic replies. Are we behind?
Look at what those replies say before concluding. An automatic reply announcing availability or a viewing produces cancellations and wasted journeys, which costs more than silence. Being behind is measured by the first-reply delay, not by the presence of a bot.
Where we come in
On your last twenty lines, the total is not what decides: it is whether the misses concentrate on one channel or spread across all of them.
- We bring your channels into a single list, inside the tools you already have.
- The extraction shows separately what was said and what the machine assumed.
- Callback rules, the handover delay during an absence and the duplicate threshold stay editable by you.
No conversion rate will come out of here for this market: nothing of the kind has ever been measured with us, and a percentage from elsewhere would be talking about a flow that is not yours.
Read next
Estate agencies: the photograph decides, the listing only confirms
A buyer eliminates nine properties out of ten before reading a line. That sorting happens on photographs.The buyer register: calling the right person the day a property comes in
An agency holds two lists that never speak to each other: what it has, and who is looking for what. Matching them is the job.Training a model: the three cases where it is worth it
Training changes the shape of answers, almost never their content. What you must supply, and the three cases where it is the right tool.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.