Algeria AgencyReference page
Artificial intelligence in Algeria: what is legal, what it costs, what works
Almost everything that blocks an AI project in Algeria is decided before the first line of code: the law, the payment, the language.
This page describes the state of artificial intelligence applied to businesses in Algeria: what the law permits since law 25-11 of 24 July 2025, what actually makes up the cost of a project, the language question, and the cases where the right decision is to buy nothing. It covers the domain in breadth; each section points to an article that treats it in depth.
It is written for the person who decides, not the person who installs. Where an operation needs technical work, that is said. Where a question goes beyond our trade — a legal judgement, a tax characterisation — that is said too, and we name the profession it belongs to.
What it does not contain: no model comparisons, no product news, no promised quantified gain. All three go stale within months and none of them ever answered a question an Algerian business actually asks.
Three constraints specific to this market
The phone is the default screen. ARPCE puts the mobile share at 88.71% of internet subscriptions in Algeria, which decides the shape of an interface before any consideration of models: a dashboard designed on a wide screen and read on a phone is not the same product.
Paying in foreign currency is an administrative process, not a form. It constrains a project’s calendar as much as the development does, and a business discovering domiciliation after signing loses weeks on a project it believed was technical.
Skills are concentrated in a few cities. A project whose daily operation rests on a single person is a project that stops the day that person changes role, and that is true regardless of how well it was built.
What the law permits, and since when
Personal data processing in Algeria is governed by law 18-07 of 10 June 2018, amended and completed by law 25-11 of 24 July 2025, published in Journal officiel no. 48. The second does not replace the first: it turns principles into documents you have to produce.
Four obligations decide whether an AI project is feasible. A register of processing activities, kept current. A ledger tracing operations carried out on the files. The designation of a data protection officer. And notification to the supervisory authority within five days of becoming aware of a breach.
The fifth stops more projects than any other, and it is not new: transferring data outside the country is subject to prior authorisation from the authority. Law 25-11 adds an assessment of the level of protection offered by the destination country.
This paragraph describes the state of a regulation as at 21 August 2026 and does not replace a lawyer’s advice. The exact scope of each obligation is read in the text, article by article, and some points are read differently by different practitioners.
Sending data to a foreign model is a transfer
A call to an interface hosted outside Algeria moves whatever data it carries. If that data identifies a person — a name, a phone number, an address, the content of a customer message — the operation is a transfer outside the country in the sense of the law, and it falls under the prior authorisation regime.
That is almost never how a project is presented. A chatbot connected to the company mailbox is described as an automatic reply tool; nobody around the table describes it as exporting the customer file, and yet that is what the provider’s logs record.
Three architectures avoid the question rather than working around it: strip the identifying elements before the call, keep the identifiers with you and send only text, or run the model on a machine you control. They do not cost the same and they do not suit the same volumes.
Verification is one operation: export a week of conversations, highlight everything that designates a person, and count the lines that would still work with the highlight removed. The result decides the architecture before any quote.
The three documents a project has to produce
The register of processing activities describes, for each activity, its purpose, the categories of people and data involved, the recipients, the retention periods, the security measures and what leaves the country. A spreadsheet is enough for a long time; what matters is that it describes your real processing and not a copied template.
The ledger is the trace of operations carried out on the files: who consulted, modified or deleted what, when, and why. It is the part tools are nearly always able to produce and almost nobody switches on.
The breach procedure is the third, and the one written latest. The notification deadline runs in days from the moment you know, which rules out discovering the procedure on the day it is needed. One page is enough: who calls whom, in what order, with what wording.
These three documents are not an administrative preliminary to the project: they are what makes the project describable. A business that cannot say where its data is cannot decide what it is entitled to do with it, and no supplier can decide that on its behalf.
Paying a foreign supplier is the first real obstacle
CIB and Edahabia cards are domestic and denominated in dinars; they do not settle a subscription billed abroad. This is the wall every business hits when it simply wants to try an interface, and it arrives before any technical question.
The regular route for a company is bank domiciliation of a service import. Since ABEF note no. 471/DG/2025 of 9 July 2025, this requires prior approval from the Ministry of Foreign Trade: the company obtains the authorisation, presents it to its bank, and the bank then proceeds with domiciliation and the currency transfer.
Domiciliation carries a tax whose rate depends on how the operation is characterised. Article 123 of the 2025 finance law sets 5% of the transferable amount on royalties. Whether a subscription to a model is a service or a royalty is an accounting and tax question, not a technical one: it belongs to your accountant, and we refuse to settle it in their place.
No irregular arrangement is recommended here, and several circulate. A project funded through a payment method not in the company’s name is a project that cannot be invoiced, amortised or defended during an inspection — and it stops the day the person holding that payment method leaves.
What you pay for in an AI project, and to whom
A project’s cost breaks into lines that are not paid to the same people, and confusing them is what produces a wrong budget. Building is paid once, to whoever builds. Usage is paid monthly, to the model provider, in foreign currency. Hosting is paid monthly, to the host. Maintenance is paid by time spent. And the human work the system does not do is paid in salaries, by you.
The fifth line is the one quotes leave out. An automation handles what it recognises and sets the rest aside; somebody empties that pile, every day, and that person is either in the budget or the project fails silently.
This page gives no amount, deliberately. A price written into a reference page is a figure nobody re-reads when it changes, and the two variables that decide yours — your real volume and the day’s exchange rate — are not ours. Our offer pages carry our rates; the arithmetic is done with your numbers.
The only useful order of magnitude is taken at your end: count the messages actually received over a month, and the share whose answer fitted in three sentences. That ratio decides whether an assistant pays for itself, before any quote, and it takes an hour to establish.
Arabic, French, Darija: the question’s language is not the answer’s
An Algerian customer frequently writes their question in Darija, in Arabic or Latin characters, and expects an answer in standard Arabic or in French. A system that answers in the exact language of the question produces approximate Darija, and approximate Darija written by a machine is heard immediately.
The rule that holds in writing is asymmetric: accept Darija as input, never produce it as output. Spoken, the question is different, because transcription decides before the model does: what was not transcribed correctly cannot be understood, whatever the quality of what follows.
Choosing the output language is a decision per channel, not a global one. A website form, an instant messenger and a phone call have neither the same audience nor the same tolerance for error, and treating all three with one rule gives three results of which only one is right.
This constraint is also a market advantage. The content and tools available for this market are overwhelmingly French-language, while most questions arrive in Arabic or Darija — and the gap is closed by work, not by a subscription.
What models get wrong here
Algerian proper nouns, neighbourhood names and wilayas are the first breaking points. A model trained mostly on foreign corpora proposes a plausible and wrong spelling, and proposes it confidently: that is what makes it dangerous in a quote, an invoice or a delivery address.
Product references and internal codes fall into the same category. They appear in no public corpus, and a system that has not been given your catalogue invents them while respecting the shape — the right length, the right prefix, the wrong item. A search engine never exceeds the quality of the catalogue it is given, and that is the same constraint seen from internal search.
The correction does not happen inside the model. It happens in a matching layer placed after it: the output is checked against your real lists — customers, products, communes — and anything matching nothing is flagged rather than silently corrected.
Verification is immediate and needs no tool: give the tool you are considering twenty names and twenty addresses taken from your own file, and count the errors. Twenty lines are enough to decide whether a matching layer is needed, and the test costs half an hour.
By sector: what can be delegated, and what never can
The criterion is the cost of an error, not the difficulty of the task. Sorting incoming requests, summarising a long document, drafting a reply: an error shows immediately and is corrected in seconds. Calculating a price, approving a file, committing the company in writing: an error is discovered later and costs money.
The criterion crosses sectors without changing shape. In an estate agency, AI sorts enquiries and does not write the listing. In a clinic, it books the appointment and never touches the diagnosis. In a professional practice, it reads the documents received and does not produce the advice. In a school, it holds the enrolment window and does not decide an application.
Logistics is where intuition is wrong most often. The customer’s question is almost never “where is my parcel”: it is “will I have it before a given hour”, which needs an answer tracking does not contain.
None of these boundaries is technical. They describe what a business is willing to let a system decide when that system is right most of the time, which is a question of responsibility and is settled before the purchase.
The method: a hundred messages before any specification
The best specification available in a business is already written: it is the last hundred messages actually received, with the answers that were given. It describes the real questions, in the real wording, with the real proportion of special cases.
The second step is not a prototype, it is a document: what is handled alone, what goes to a person, and the hand-off rule. A project that starts with a demonstration defers that decision until the moment it costs the most to take.
The third is a single slice, put in your own team’s hands before any customer. That is where the gap appears between what the system handles and what the business believed it would handle, and the gap is measured in days rather than in meetings.
The fourth is a gradual opening with a log read every week. Without that log nobody can say whether the system works: there is no way to tell an assistant that helps from one that politely filters.
What can be measured, and what cannot
Three numbers are enough to supervise a running system: how many requests came in, how many went out handled, and how long the oldest of those remaining has been waiting. Taken once a week in a spreadsheet, they give a trend after three months, and a trend is what makes a decision possible.
The success rate a tool displays does not measure that. It counts what the program did, not what happened to the person at the other end: setting a request aside is a success from the program’s point of view and a wait from the customer’s.
What cannot be measured from a dashboard is the quality of what was said. That is listened to: a weekly session on a sample of real exchanges, at the start, reveals in an hour what no curve shows.
A missing figure is information. An exception queue that is always empty usually means the system is accepting cases it should have set aside, and the errors have gone to the customer instead of piling up in a tab.
When not to buy artificial intelligence
When the same question comes back twenty times a week, the problem is a missing page, not an absent assistant. Writing that page costs a day and settles the matter for good; an assistant answering in the page’s place costs every month and leaves the page missing.
When the process is written down nowhere, automating it freezes a mess instead of correcting it. A system faithfully reproduces what it is described, and what was never described is reproduced as the person describing it that day happened to remember it.
When nobody at your end can open a screen every morning, the subject is not the tool: it is the organisation. An exception queue with no owner is not discovered, it accumulates, and it accumulates silently while the report stays green.
And when the volume does not justify it. A switchboard taking fifteen calls a day does not need a voice agent: it needs the phone answered, and the decision is made by comparing a system’s monthly cost with that of the hours it frees.
What we do, and what we refuse
We build systems grounded in a business’s real content — pages, catalogue, procedures, message history — that cite their source and say they do not know when they do not know. We write the limit before go-live: what the system handles alone, what it hands over, and what it is not allowed to do.
We promise no quantified gain. It depends on your volume, the cleanliness of your data and the variety of your cases, three things we discover at the same time you do, and a percentage announced before looking would be an invented percentage.
We are not lawyers and we do not settle a point of law. We describe what the text asks, we map your flows, and we name the moment the question belongs to a lawyer or an accountant — the tax treatment of a foreign-currency subscription being one.
And we will often tell you the project is not worth starting. It is the least profitable answer available to us and it is what makes the others credible: a business that has been advised against a project is a business that can believe the next quote.
Frequently asked questions
Is an Algerian business allowed to use ChatGPT or another foreign model?
Using a tool is not forbidden; sending personal data to it is what is regulated. As soon as a message contains a name, a number or an address, sending it to a service hosted abroad falls under the transfer regime, which is subject to prior authorisation. The practical question is therefore not the tool but what is sent to it, and it is settled by a written usage rule rather than by a ban nobody will observe.
Should the model be hosted in Algeria?
It is one of the three possible architectures and rarely the first to try. It removes the transfer question and replaces it with a fixed cost — machine, electricity, administration — that only makes sense above a certain volume. Below that, stripping identifying elements before the call costs far less and answers the same constraint.
How much does an AI project cost?
This page gives no amount, deliberately: the two variables that decide yours are your real volume and the day’s exchange rate, and a figure written here would be wrong before it was read. What is stable is the list of lines — building, monthly usage, hosting, maintenance, and the human work the system does not do. Our offer pages carry our rates.
Can an assistant answer in Darija?
It can understand it adequately and it writes it badly. The rule we apply is to accept Darija as input and answer in standard Arabic or French: approximate Darija produced by a machine is heard immediately and costs more credibility than it earns. Spoken, the difficulty moves to transcription, which decides before the model does.
Where do we start if we have never done anything?
With two measurements that cost nothing and do not involve us. Count the messages received over a month and the share whose answer fitted in three sentences; list the software holding data about people and the software hosted outside Algeria. The first says whether a project makes economic sense, the second says in what form it is possible.
How to start
Tell us what you receive each week — messages, calls, documents — and where the data that produces is stored.
We look at a hundred real exchanges and tell you what is worth automating, what is not, and under which architecture the law permits it.
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.