Resources·AI automation for a Moroccan SME: the 7 processes to start with
AI automation for a Moroccan SME: the 7 processes to start with
In a company of twenty or fifty people, nobody has a spare day a week for an AI project. The repetitive work already costs several hours per person per week, and it appears in no budget line.
So the useful question is not which AI to buy. It is which process to hand over first, and which one to leave alone. This guide answers both. It describes the seven processes we find in almost every Moroccan SME we audit, then gives the impact / effort / risk matrix that ranks them against your own situation.
What does automating a process with AI actually mean in an SME?
Automating a process with AI means handing a system the repetitive part of a task — reading, sorting, extracting, drafting, copying between tools — while decisions, client relationships and final approval stay with your team. The process does not change owner. It changes speed.
That definition rules out two things often sold under the same word. It rules out replacing a role, which is neither our trade nor what SMEs ask for. And it rules out the eighteen-month transformation programme, which mostly produces meetings. A useful automation is measured in weeks, on a task someone already performs today, using the tools the company already pays for.
Which seven processes should you automate first?
These seven recur because they share three properties: they happen often, they are written down somewhere, and a mistake shows up before it becomes expensive.
| # | Process | What the system does | What stays human |
|---|---|---|---|
| 1 | Inbound requests | Sort, qualify, draft a reply | Sending, negotiating |
| 2 | Quote and invoice chasing | Spot what is overdue, draft the follow-up | Tone and timing |
| 3 | Meeting notes | Transcribe, summarise, extract actions | Arbitration, reassignment |
| 4 | Quotes and proposals | Assemble a first draft from history | Price, commitment, signature |
| 5 | Supplier documents | Extract the fields, prepare the entry | Accounting control |
| 6 | Content and online presence | Produce first drafts, adapt them | Editorial approval |
| 7 | Weekly reporting | Gather the numbers, write the commentary | Interpretation, decisions |
1. Handling inbound requests
An SME receives enquiries by email, WhatsApp, web form and phone, and nobody can say how many arrive in a week. A system can bring them into one place, qualify them against your criteria, pull up the client's history and prepare a reply. Approval stays human. The visible gain is not drafting time: it is time to first response, which often decides the deal.
2. Chasing quotes and unpaid invoices
Nobody enjoys chasing, so nobody chases on time. The system watches quotes with no answer and invoices past due, drafts the follow-up with the right references, and presents it for sending. This is the process that pays for itself fastest, because it touches cash directly. Tone and timing stay a human call: a badly judged reminder costs a client.
3. Meeting notes and action tracking
Transcription is the one item on this list where the technology has been mature for years. What AI changes is the step after: pulling out decisions, actions and deadlines, then putting them where the team already tracks work. Skip that step and you produce notes nobody reads.
4. Preparing quotes and proposals
A quote is rarely written from scratch. It reuses lines, descriptions and terms you have used before. A system that reads your history produces a coherent first draft in minutes. Pricing, discounts and contractual commitment are never automated — those are decisions, not tasks.
5. Extracting data from supplier documents
Invoices, delivery notes, statements: entering them is slow, thankless, and it blocks the month-end close. An extraction system reads the documents and prepares the entries. The accounting control stays whole, and that control is what makes the automation acceptable. One prerequisite: if your documents arrive as crooked phone photos, fix the input first.
6. Content and online presence
Product descriptions, review replies, posts, articles: the cost is not the writing, it is the consistency. A system produces first drafts from your expertise and your real subjects; your team keeps the review and the editorial responsibility. Publishing text nobody has read is the fastest way to damage a brand.
7. Weekly reporting
Pulling the week's numbers out of three different tools often costs half a day from someone whose job it is not. That assembly work is fully automatable, down to the factual commentary. Interpretation, and the decisions that follow, remain Monday's meeting.
How to rank them: the impact / effort / risk matrix
The matrix below is our own scoping grid, used in AI Magic Makers audits. It is not an industry standard, and its scores are not measurements: they are the default values we most often observe, to be recalculated against your numbers in the first hour of work.
Three criteria, scored 1 to 3:
| Criterion | What you measure | Score 1 | Score 3 |
|---|---|---|---|
| Impact | Hours recovered per week, across the team | Under an hour | More than five hours |
| Effort | Data access, tools to connect, input quality | One tool, clean data | Three or more tools, manual entry |
| Risk | Consequence of an undetected error | Visible before sending | Commits the company or exposes personal data |
The reading rule fits in one sentence: start with high impact, low effort and low risk, and keep the high-risk processes for when the team already trusts a first system.
| Process | Impact | Effort | Risk | Suggested order |
|---|---|---|---|---|
| Quote and invoice chasing | 3 | 1 | 1 | 1st |
| Meeting notes | 2 | 1 | 1 | 2nd |
| Weekly reporting | 2 | 2 | 1 | 3rd |
| Inbound requests | 3 | 2 | 2 | 4th |
| Supplier documents | 3 | 3 | 2 | 5th |
| Quotes and proposals | 3 | 2 | 3 | 6th |
| Content and online presence | 2 | 2 | 2 | Commercial priority decides |
Two readings worth making. Quotes come late despite high impact: the risk is contractual, and you only touch it once the human checkpoint is proven. Document extraction comes late for a different reason: the effort is real, because it depends on the quality of what goes in.
Where do you actually start?
Scoping a first process in five steps
- 1Find the taskThe one nobody defends in a meeting.
- 2Time what existsStopwatch three real occurrences.
- 3Set the checkpointWho approves, and before what.
- 4Run it on real casesYour files, not a demo example.
- 5Measure at thirty daysSame occurrences, same stopwatch.
Once two or three processes are running, the question becomes oversight: who sees what is moving, and who approves it. That is settled on a shared board rather than in one more tool, and we cover it in Getting real work out of AI agents: start with your Kanban.
The step companies skip is the second one. With no starting figure, nobody can say whether the system saved time, and the project becomes impossible to defend internally as soon as priorities shift.
How do you measure what the automation returns?
Three measurements are enough, and they fit in one shared sheet.
| What | How | When |
|---|---|---|
| Time per occurrence | Stopwatch three cases, before and after | Week 0, week 4 |
| Rework rate | Share of outputs corrected before sending | Weekly, first month |
| End-to-end delay | From request to reply sent | Continuously |
Rework rate tells you the most. If it stays high after a month, the problem is not the model: it is the process, which was not written clearly enough to be handed over.
What Moroccan law requires before you connect AI to your data
Law 09-08 applies as soon as a process handles personal data, and your inbound requests, quotes and invoices all contain some. Morocco's data protection authority, the CNDP, sets out three points that shape your architecture.
First, every processing operation must be declared, unless it is excluded or falls under the authorisation regime. Second, some operations require prior authorisation rather than a simple declaration: sensitive data, genetic data, and cross-linking files between organisations with different purposes, among others. Third, transferring personal data abroad is only possible in the cases listed in articles 43 and 44 of the law (CNDP, Formalités, retrieved 20 September 2026).
That last point is the one that surprises people. An AI tool hosted outside Morocco processes your data outside Morocco. Depending on the data involved, that changes the formality you owe — and sometimes the technical answer: this is exactly where a local AI integration on your own machines becomes the reasonable option rather than a luxury precaution.
Where AI is the wrong answer
We regularly decline to automate a process. These are the cases.
The process is not written down. If two people do it differently and nobody knows which one is right, automation freezes the disagreement. Write it first; the gain often appears right there, with no AI involved.
The volume is low. A task that comes round twice a month justifies neither the scoping time nor the maintenance. Keep it manual and say so.
The inputs are dirty. Scattered data, unreadable scans, duplicate CRM records: the system will reproduce the mess faithfully, only faster.
The decision affects a person. Hiring, discipline, credit, disputes: AI can prepare the file, never decide it.
The process should disappear. The best outcome of an audit is often deleting a report nobody reads any more. Deleting costs less than automating.
What to do this week
Three actions, in order. List the seven processes above and note, from memory, who does each one and how long it takes. Then stopwatch three real occurrences of the first one on your list — half an hour of work that replaces a debate. Finally, decide who will approve the outputs: with no owner, an automation works for a month and then quietly stops.
If you want the ranking built on your numbers rather than our defaults, that is exactly what the AI audit produces: five days, your processes, the recoverable hours quantified and three prioritised projects. Once the first project is identified and the question becomes who builds and maintains it, it is handled either through consulting or through a team workshop, depending on whether you would rather we did it or taught your team to.
Frequently asked questions
Do we have to change tools to automate a process?
In most cases, no. The seven processes above connect to a CRM, an inbox, a spreadsheet or a management tool already in place. Changing software at the same time as automating is the surest way to fail at both.
Who should own automation inside an SME?
The person who performs the process today, not the IT department. They know where the exceptions are, and they are the one who will spot a wrong output. Plan on two hours a week of their time for the first month.
How do we know a process is ready to be automated?
Ask one question: can someone explain it end to end in ten minutes without saying "it depends" more than twice? If yes, it is ready. If not, the work to do is writing it down, and that goes well in a workshop with the team concerned.