![\"AI](\"https:\/\/storage.googleapis.com\/arrowlancer-prod.firebasestorage.app\/blog-ma\/maturite-ia-maroc-test-avant-devis-2026\/1.jpg\")
<\/p>\nAI projects that fail in Morocco rarely fail on the model. They fail at the moment the tool meets reality: the data cannot be pulled, the process was never written down, and nobody on the org chart is responsible for keeping the assistant alive once the provider has gone.<\/p>\n
Which is why the budget question usually comes too early. Before you request a quote, there is a more useful filter: is your company in a state to receive an AI tool? The answer moves the price, the timeline, and sometimes the decision to launch nothing at all this quarter.<\/p>\n
The test below runs internally, in two or three meetings, with no provider and no budget committed.<\/p>\n
A badly prepared AI project shows itself fast. In the first scoping meeting you ask a simple question: where is the data the tool needs, and who owns it? If three people give three different answers and a fourth mentions a spreadsheet on their own laptop, the project is already late before a line of code exists.<\/p>\n
The other signals are similar in kind. The need is phrased as a wish list (“an assistant that answers everything”) rather than a measurable task. Nobody knows how many times a week the task is performed today, or how long it takes. Reference documents exist in several versions, two of which still circulate as email attachments. None of this kills a project. But each item will be paid for in consulting days billed at development rates, when it could have been handled internally for far less.<\/p>\n
The useful question is not “do you have a lot of data” but “is your data reachable, current and unique”. Four checks settle it: where each required source lives, how you get data out of it (native export, API, database access, or only scanned PDFs), how often it is updated, and whether a single reference version exists.<\/p>\n
<\/p>\n
Take an internal assistant meant to answer questions about HR procedures. If three versions of the leave policy exist, the assistant will quote one of them confidently, and sometimes it will be the wrong one. A model does not arbitrate between two contradictory documents, it picks one. Two weeks of questionable answers are enough to lose user trust, and you do not win it back with a technical update.<\/p>\n
For document automation the filter is more physical. Invoices scanned crooked at low resolution, stamps sitting on top of the amounts, photos taken on a phone in a warehouse: extraction accuracy drops and you end up funding the human review you expected to remove. Take fifty real documents, not the cleanest ones, and have them tested before you sign. If you want the source-by-source version, the data quality audit before a BI or AI project<\/a> covers the checks in detail.<\/p>\n An honest test fits on one sheet. Ask the person who actually performs the task to describe it in ten lines: what comes in, the decision points, the exceptions, what goes out, and who approves. If the manager and the operator write two different versions, you do not have an AI problem, you have an implicit process. Automating it means freezing the confusion inside a tool and making it harder to correct.<\/p>\n Then look at the exception rate. Plenty of Moroccan back-office processes run on a large share of special cases: a client paying cash, a purchase order arriving over WhatsApp, a missing supporting document regularised later. If half the files fall outside the standard path, do not try to automate the hard half. Let the tool handle the standard route and send the rest to a person, with a clear handover rule.<\/p>\n One last point, often skipped: the before measurements. Monthly volume, time per unit, current error rate. Without those three numbers you cannot set an acceptance criterion and you cannot demonstrate a gain six months later, so the renewal conversation ends up being about feelings.<\/p>\n You do not need a data scientist. You need an owner: internal, named, with two to four hours a week genuinely freed up. The work is unglamorous. Read user feedback, spot wrong answers, update the knowledge base when a procedure changes, ask the provider for an adjustment when the same case keeps coming back.<\/p>\n Without that role the tool does not break, it expires. The knowledge base ages, answers become wrong at the edges, then plainly wrong, and the teams go back to the phone and the WhatsApp group. It is the most ordinary cause of death for internal assistants, and it appears in no quote.<\/p>\n You also need someone on the management side willing to refuse scope creep halfway through. And adoption deserves an early conversation: if staff assume the tool is preparing redundancies, they will not feed it, however good the engineering is. Saying what the tool is for, and what it is not for, is as much a prerequisite as clean files.<\/p>\n As soon as the project touches personal data, whether customers, employees or candidates, Law 09-08 applies exactly as it does to any other processing: defined purpose, legal basis, informing the people concerned, retention periods, formalities with the CNDP. Nothing AI-specific so far, which is precisely why many companies discover the topic at the end, once the tool is already built.<\/p>\n The AI-specific part is where the data goes. Most consumer models are hosted outside Morocco. Sending client files or payslips to a US API is a transfer out of the country, and transfers are regulated. Three routes exist depending on the case: pseudonymise or filter the fields before sending, restrict the categories of data processed, or host an open model on infrastructure located in Morocco. The exact formalities depend on your processing, so check them on the CNDP website or with your legal counsel rather than copying what a peer company did. The ChatGPT and Copilot at work case<\/a> shows how often informal usage arrives long before any official decision.<\/p>\n Before the quote, verify three things: whether a current processing register exists, who the contact person is on these matters, and whether your subcontracting contracts anticipate this kind of use.<\/p>\n Put these five questions to your management team. A question counts as a “yes” only if you can back it up on the spot, without going to check.<\/p>\n Four or five yeses: you are ready on this scope. A short scoping phase is enough, the quote will hold, and a first useful release typically lands in six to ten weeks. Two or three yeses: the most common situation in Moroccan SMEs. Plan a readiness diagnostic before the quote and accept that a real share of the budget, often 20 to 40 percent, will go to preparing data and process rather than building. Zero or one yes: do not launch this scope now. Either start with the data work, or pick a simpler target. The matching ranges are broken down in our AI project budget guide for Morocco<\/a>.<\/p>\n A provider who knows the territory refuses to give a fixed price before seeing real material. They ask for a representative sample, thirty to a hundred documents depending on the case, a database extract, and an hour with two business users. They also want to know who decides when people disagree about a rule, because that is where projects drift.<\/p>\n That work has a name: a short paid scoping phase, producing a written scope, measurable acceptance criteria, an architecture and an estimate that survives contact with the data. In Morocco this kind of engagement often sits between MAD 10,000 and MAD 30,000 depending on scope size, and is frequently credited against the project if the same provider builds it. The logic matches a digital transformation audit<\/a>: what you buy first is a decision, not code.<\/p>\n The warning signs are clear enough in the other direction. A quote delivered in 48 hours with no access to data. An accuracy figure promised before anyone has seen a single document. No mention of the CNDP anywhere in the proposal. No measurable acceptance criterion, which means no way to say whether delivery is compliant. And a first phase that covers “all company documents”.<\/p>\n The wrong conclusion here would be to launch a six-month data quality programme before touching AI at all. Few companies see it through, and the momentum fades. The cheaper answer is to shrink the scope to the place where your data is already good enough.<\/p>\n In practice, an internal assistant built on the forty genuinely current documents beats one wired to two thousand unsorted files. A pilot with ten willing users over six weeks, with a written acceptance criterion, teaches you more than another steering committee. And while the tool is being built, two workstreams can run in parallel at no development cost: naming the internal owner and freeing their hours, and handling the compliance side. Those are the two easiest delays to avoid, and the two we see most often.<\/p>\n Arrowlancer works in that order. We run this readiness diagnosis across data, process, skills and compliance before pricing an internal assistant, a chatbot or a document automation flow, because selling a tool to an organisation that cannot receive it serves nobody. It is the starting point of our AI solutions<\/a>, including the cases that require hosting inside Morocco. If you want to run the test on a specific scope, tell us about your project<\/a> before you ask for a quote.<\/p>\n","protected":false},"excerpt":{"rendered":" Most AI projects that fail in Morocco die on prerequisites nobody checked. A five-question test to see where your company stands before you ask for a quote.<\/p>\n","protected":false},"author":0,"featured_media":237,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[54],"tags":[903,906,63,69,468],"class_list":["post-236","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-digital_transformation_strategy-en","tag-ai-project-morocco-en","tag-ai-readiness-en","tag-cndp-en","tag-digital-transformation-en","tag-moroccan-smes-en"],"_links":{"self":[{"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/posts\/236","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/comments?post=236"}],"version-history":[{"count":1,"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/posts\/236\/revisions"}],"predecessor-version":[{"id":239,"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/posts\/236\/revisions\/239"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/media\/237"}],"wp:attachment":[{"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/media?parent=236"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/categories?post=236"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/arrowlancer.ma\/blog\/wp-json\/wp\/v2\/tags?post=236"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}Process: what you must be able to describe before automating it<\/h2>\n
Internal skills: who runs the tool once the provider leaves<\/h2>\n
CNDP compliance: the prerequisite discovered too late<\/h2>\n
The AI readiness Morocco test in five questions<\/h2>\n
\n
What a serious provider checks before pricing your project<\/h2>\n
Closing the gaps without losing six months<\/h2>\n
The right order: diagnosis, scope, then quote<\/h2>\n