AI for SMEs: what works, what disappoints, and where to start

An artificial intelligence project in an SME can fail for a reason that has nothing to do with AI: the data was not ready.

A sales forecast needs a reliable sales history. An assistant that answers customers needs answers written down somewhere. Till anomaly detection needs a till that actually records. When those foundations are missing, AI does not fill the gap: it exposes it, by producing confident, wrong results.

This guide separates what works today for a business in Cameroon, what often disappoints, what it costs, and the rules that apply to your data.

The order that matters: data, process, then AI

Before evaluating any tool, ask three questions, in this order.

Does the data exist? Not “could we produce it”, but: is it recorded today, in a usable form, with enough history? One year of properly entered sales beats five years of notebooks.

Is the process stable? Automating a process that changes every week means automating disorder. Stabilise first.

Is the decision actually being made? If nobody reads the current report, a smarter report changes nothing. AI amplifies a decision; it does not create one.

If you answer no to any of these, you will gain more by structuring your operations — the job of management software or a CRM — than by buying an AI tool.

Five uses that work today

1. Reading invoices and delivery notes

Supplier invoices, delivery notes, statements: automated document reading saves hours of data entry, and every extracted line can be checked against the original. It is often the best first project, precisely because the result can be verified.

2. Drafting WhatsApp replies

Take two messages a Douala shop receives one evening:

• “Bonsoir, le carton d’huile de 5 litres là, c’est à combien ? Vous livrez à Makepe ce soir ?”
• “Good evening, una get the 25 kg bag of rice? How much? Una fit deliver for Molyko today?”

An assistant connected to your catalogue and delivery rules drafts a reply: the catalogue price, whether the area is covered, the delivery fee you set, and a closing question (“Shall I put one aside for you?”). The employee reads it, corrects it if needed, and sends it. Response time drops, and the relationship stays human.

To connect WhatsApp to your catalogue and orders, see Nematek One WhatsApp Commerce.

3. Forecasting replenishment

For items that sell steadily, a simple statistical forecast gives the buyer a numerical baseline, which they adjust with what they know about the market. It requires a clean sales history and stock records that reflect reality.

4. Spotting till anomalies

Cash discrepancies, unusual discounts, repeated cancellations, abnormal returns: these signals can be detected automatically. The tool concludes nothing; it draws a human’s attention to what deserves a look.

5. Helping with writing

Meeting notes, standard replies, French–English translation, product descriptions. The time saved is real and the risk is low, provided a human reviews before anything goes out.

AI is also changing how your customers look for you. Our guide to Google AI Mode shows how to use it, and what lets a business appear in its answers.

Three uses that often disappoint

The autonomous front-line chatbot. Without a maintained knowledge base, it makes things up. A customer given a wrong price costs more than the time saved.

“Prediction” without history. Forecasting demand for a product launched two months ago is not an AI problem: it is an absence of data.

The intelligent dashboard on top of bad data. It gives doubtful numbers an appearance of rigour. It is the worst of both worlds.

What AI actually costs

The price of the tool is rarely the main line. The real budget includes:

• data preparation: cleaning, structuring, history — often the heaviest line;
• integration with existing systems, so the AI reads and writes in the right place;
• consumption: AI services are billed by usage, often in dollars or euros, and cost follows volume;
• human oversight, at least at the start;
• maintenance: processes evolve, models change, results drift.

Estimate real monthly consumption before committing, not after the first usage spike. A budget that only covers the subscription is a badly scoped project.

For the prices of the main models and a worked example in FCFA, see our LLM guide.

Your data: Cameroonian law, the GDPR and the AI Act

What data leaves the business? An AI service hosted abroad processes whatever you send it. Customer files, health data, salaries: each category deserves a written decision.

If you sell to people who are in the European Union, for example the diaspora in France or Belgium, the GDPR can apply even without an establishment in Europe (Article 3). Likewise, the AI Act, the EU regulation on artificial intelligence, also covers businesses located outside the Union when the output of their AI system is used in the Union (Article 2).

Two common-sense rules complete those texts. Who is responsible for the result? A recommendation produced by a model commits the business that acts on it. Can the result be checked? A system whose output nobody can explain should not decide a credit refusal or a dismissal on its own. The NIST AI Risk Management Framework offers a framework that scales to the size of the organisation.

Local realities that change the calculation

Bilingualism is an asset. Your customers write in French, English, Pidgin, and often a mix of all three. Writing and translation assistance therefore pays off quickly, provided you test it on your real messages.

Modest volumes are enough. An SME does not need “big data”: two years of clean sales support useful forecasts. The constraint is quality, not volume.

Connectivity imposes an architecture. A critical process that depends on a remote service stops when the connection drops. Always plan a degraded mode: what does the team do while the service is unavailable?

Starting in 90 days

Weeks 1 to 3 — pick one process. The one that consumes the most repetitive hours and whose output can be checked. Measure the time spent today: without that measurement, you will never know whether the project succeeded.

Weeks 4 to 8 — prepare the data and prototype. On your real data, never on a demo set. Compare the tool’s output with human work on the same sample.

Weeks 9 to 12 — deploy with oversight. A person keeps final approval. Measure again and compare.

If the gain is not measurable after 90 days, stop and change use case. That is rigour, not failure.

Scoping your first use case

The useful step is not choosing a technology: it is identifying the repetitive process whose cost can be measured and whose output can be checked.

We support that scoping through our AI and automation and AI consulting and development services. See also Automatisation, BI & Intégrations Nematek, or request a free assessment.