Expect six to ten weeks after go-live for the first measurable gain, three months to know whether adoption is holding, and twelve to twenty-four months before saved hours alone repay a custom build. If your plan promises a return on an AI investment within six months based purely on time saved, the arithmetic will not hold.
That is not necessarily bad news. In a Moroccan SME, the AI projects that end up paying rarely do it through minutes shaved off a task. They do it through the errors no longer made, the quotes sent the same day instead of three days later, and the know-how that stays in the company when the purchasing manager leaves. Those effects arrive later and are measured differently.
To make this concrete, take a hypothetical but plausible case: a Casablanca import and distribution company, 60 staff, rolling out an internal AI assistant over its procedures, supplier contracts and product sheets. Every figure below is a working assumption, not a real client.
The counter on an SME rolling out an internal AI assistant
Set the parameters: MAD 220,000 of development, ten weeks of build, MAD 3,500 per month of running cost (hosting, model usage, corrective support), 18 target users across sales, order administration and purchasing. Here is what the curve looks like.
- Weeks 1 to 10: zero gain. Scoping, document clean-up, indexing, answer testing. The counter sits at zero and that is normal. A project that claims returns during the build phase is usually measuring something other than real gain.
- Week 11: go-live. 18 accounts created, 11 people log in during the first week, mostly out of curiosity.
- Week 14: 7 genuinely active users. Estimated gain of 15 minutes per day per active user. At MAD 100 per loaded hour (an assumption for an admin profile), that is roughly 35 hours a month, or MAD 3,500. Exactly the running cost. Net monthly: nothing.
- Month 4: narrowing. The tool stops being sold as everyone’s assistant and becomes the reflex for one specific move: answering supplier-terms questions during a customer call. 13 active users, 25 minutes a day. Around 108 hours a month, about MAD 10,800, minus MAD 3,500 of run: MAD 7,300 net.
- Month 6: the second-order effects show up. Two pricing-condition mistakes avoided per month, and a response time on price requests that drops from two days to a few hours.
Cumulative gain at six months: somewhere between MAD 25,000 and 35,000 against MAD 220,000 invested. Put that way, the project looks like a failure. It is not. At six months, what you judge is the slope, not the total. A net monthly gain that climbs from zero to MAD 7,300 between month 3 and month 4 and then holds tells you the mechanism works. A net gain that flattens at MAD 1,000 and stops moving tells you the opposite, however enthusiastic the steering meeting is.
Chatbot, OCR, internal assistant: three different clocks
Talking about AI return in the abstract helps nobody, because the three project families Moroccan SMEs buy most often do not run on the same clock.
A customer chatbot is the fastest clock, because its volume is known before signature. If you receive 1,800 WhatsApp messages a month and half are repeat questions (opening hours, stock on a reference, order tracking), the gain can be worked out on the back of an envelope. It is also the project where platform billing can quietly eat the benefit, a point covered in our piece on WhatsApp AI chatbot costs and pitfalls in Morocco.
Document extraction is the most predictable clock. You are counting units: supplier invoices, delivery notes, claim files. The gain is close to linear and the break-even point depends almost entirely on volume. At 400 invoices a month, a custom build defends itself. At 40 invoices a month, no honest calculation justifies it, and a serious provider should tell you that before quoting.
The internal assistant is the slowest and least linear of the three. Its return depends on a factor no contract can promise: whether people actually use it. It is also the one producing the value that is hardest to price and longest-lasting, since it turns spoken knowledge into something you can query. The timelines and cost lines for this kind of build are laid out in our guide to an internal AI assistant in Morocco.
The calculation nobody runs before signing
Most decision-makers compare the quote against zero. Zero does not exist: the status quo has a cost, it is simply invisible because you already pay it every month. Before judging an AI budget, put rough numbers on four things.
- Time spent hunting for information that already exists somewhere in the company. Ask three people to track it for a week, not in a meeting.
- The cost of recurring mistakes: a discount applied on expired terms, a warranty period quoted wrong, a reference ordered twice.
- How long a new hire takes to become productive, which depends directly on how readable your procedures are.
- Deals lost for lack of a fast answer. This is the vaguest line, so estimate it low rather than high.
Two cautions. Do not inflate this cost to justify the spend, or you build a case you will not be able to verify later. And remember that an hour freed is only a saving if it is reallocated to something, otherwise it is comfort rather than return. Starting budget ranges are in our guide to AI project cost in Morocco.
AI ROI Morocco: why the curve goes backwards
When the curve stalls, the cause is almost always one of these three, and rarely the model itself.
Soft adoption. The tool works, nobody opens it. Usually because it asks for a detour: one more login, one more tab, one more habit to form. An assistant embedded where people already work (internal chat, the ERP, the phone) gets usage levels a separate portal never reaches.
Dirty data. Three versions of the same supplier price list, two contradictory procedures, PDFs scanned crooked. The assistant answers with what it was given, and the first user who receives a wrong answer stops using it and tells the others. A data quality audit before a BI or AI project costs a fraction of the build and prevents that scenario.
The wrong use case. The most common one. A visible but cheap task gets automated instead of a boring, repetitive, expensive one. High-volume, low-judgement work automates well. Work that requires arbitration, negotiation or legal responsibility does not.
The checkpoints at 6 weeks, 3 months, 6 months
Put these three meetings in the plan before the first day of development, with the same people each time, including at least one end user.
Six weeks after go-live
You are not looking at return yet, you are looking at usage. Weekly active users against target, queries per active user, and above all the share of questions that get no useful answer. Above 25% of answers judged useless, fix the content before you keep training teams.
Three months
This is the real test. Three questions: is weekly usage rising, flat or falling? Is there at least one business move the team no longer does the old way? Does the measured monthly gain exceed the running cost? Two negative answers out of three justify a change of direction, not an internal communication campaign.
Six months
You compare the slope against the original projection, fold in the indirect effects (errors avoided, response times, dependence on one key person) and decide: extend to another team, freeze as is, or stop.
Measuring without fooling yourself
Three indicators are enough, as long as you define them up front and then leave them alone. Time saved must come from a measurement, not an impression: time the same task five times before go-live and five times after, with the same people. Errors avoided have to be counted on a specific category you were already tracking, otherwise you have no baseline. Real adoption is weekly active users against target, never accounts created.
Two classic traps when you set out to measure AI performance. Crediting the AI project with a gain that came from a reorganisation done at the same time. And redefining an indicator mid-course because the first number was disappointing. If you genuinely have to change it, write it in the minutes with the date.
The threshold that should trigger a stop
Three signals deserve a clean decision rather than a sixth enablement workshop. Fewer than 30% weekly active users at three months with no improvement after a serious relaunch. A running cost above the measured gain for three consecutive months with flat usage. And the inability to measure anything at all after four months, which is not a return problem but a requirements problem.
Stopping does not mean losing everything. The cleaned document base, the index, the ERP connectors and the access rules stay reusable assets for another use case, often narrower and more profitable. What costs real money is not stopping at six months. It is paying a run fee for three years on a tool twelve people have forgotten.
What to put in the quote to protect that return
An AI project is protected at quote stage, not at review stage. Ask in writing for: three success indicators with their measurement source and baseline value; the three checkpoints with dates and attendees; usage telemetry included in the delivery scope (active users, queries, unanswered questions) rather than offered as an extra; a formal decision point at three months with the option to shrink the scope; ownership of the index, the prompts and the data; and the running cost broken out line by line over twelve months. Our guide on how to read an IT quote before signing covers the clauses that matter most.
Frequently asked questions
Can an AI project pay for itself in under six months in Morocco?
Yes, but almost only on high volumes and highly repetitive tasks: invoice processing in the hundreds per month, first-level answers on a high-traffic channel. As soon as internal knowledge or judgement-heavy processes are involved, the realistic horizon is more like twelve to twenty-four months.
How do you measure the gain if nothing was ever measured before?
Take two weeks before go-live to record three numbers by hand: average time on the target task, monthly volume, and errors counted in one category. It is imperfect and perfectly adequate as a baseline. Without a starting point, no later calculation will be credible.
Should CNDP compliance be counted in the project cost?
Yes, as soon as the assistant touches personal data about employees, customers or candidates: declaration, documenting the processing, hosting and subprocessor clauses all cost time and sometimes professional fees. Check the obligations that apply to your case on the CNDP website.
Is it better to start small or cover a wide scope immediately?
Small, on one measurable business move, on a technical base designed to extend. A wide scope dilutes adoption, and an AI project without adoption has no return to discuss.
At Arrowlancer we scope every AI project (internal assistants, automation, document analysis, sovereign AI hosted in Morocco) with its success indicators and checkpoints written down before the first line of code, so you know when to judge the result instead of discovering it six months later. Our full range of services is on the site, and if you want to run this calculation on your own use case before signing anything, get in touch.