For a Moroccan small business, useful AI customer support is not just a chatbot that answers quickly. It is a controlled workflow that understands the customer’s language, gives only approved information, knows when to stop, and passes uncertain or sensitive cases to a person.
This guide sets out a vendor-neutral operating model for handling routine enquiries in French, Moroccan Arabic (Darija), and Arabic script. It is designed for small teams that receive questions through channels such as a website contact form, email, social messaging or WhatsApp Business. It does not assume that every channel or AI provider offers the same integrations.
1. Start with the enquiries you actually receive
Do not begin by choosing a chatbot. First, review a representative sample of past enquiries using an approved internal process. Remove names, phone numbers, order references and other identifiers before using examples for analysis. Group questions by the work needed to resolve them.
| Enquiry type | Typical example | Suitable handling |
|---|---|---|
| Opening information | Opening hours, location, available services | Approved answer from a maintained FAQ |
| Service or product details | What is included, available options | Answer from current, approved service information |
| Booking or order status | Has my appointment or delivery been confirmed? | Authenticated lookup or human handoff; never guess |
| Complaint or refund | A service did not meet expectations | Empathetic acknowledgement and human escalation |
| Personal or sensitive matter | Identity documents, payment details, health information | Do not collect in a general AI chat; use the approved secure process |
| Unclear request | Voice note, slang, ambiguous message | Ask one concise clarification or route to a person |
The first deliverable is a short list of the most common, low-risk questions—not a complete copy of the customer database.
2. Build one approved source of truth
AI should not invent business facts. Prepare a small knowledge sheet that an employee can maintain. Give every item an owner and a review date.
- Business name, address, map link and verified opening hours.
- Services or products, exact descriptions and current conditions.
- Prices and offers only when approved, with validity dates and exclusions.
- Booking, cancellation, delivery, returns and complaint procedures.
- Official contact routes and the cases that require a person.
- Languages supported and any terms that should remain untranslated.
Keep temporary offers separate from permanent information. If a price, schedule or promotion expires, remove or update it in the source before the assistant uses it. Never let the model infer a price from an old message.
3. Treat French, Darija and Arabic as distinct language needs
Moroccan customers may write in French, Darija using Latin characters, Darija in Arabic script, Modern Standard Arabic, or a mixture within one message. A reliable workflow should recognize the language and script without assuming that every Arabic message is formal Arabic or that Latin Darija is a spelling error.
- Detect before replying: identify the language and script from the customer’s message. If confidence is low, ask which language they prefer.
- Preserve meaning: do not translate business names, product names, addresses, legal terms or reference codes unless an approved equivalent exists.
- Use natural phrasing: maintain a small set of human-approved examples for common questions in French, Darija and Arabic script.
- Keep the facts identical: the price, opening time, offer conditions and next step must not change between language versions.
- Offer a language switch: let the customer ask for another language at any point.
Do not claim that a machine translation is culturally or legally accurate without a human review. For sensitive complaints, contractual wording or safety information, route the response to a qualified person.
4. Use a controlled response workflow
- Receive: collect the message through the business’s existing approved channel.
- Classify: identify the enquiry category, language and whether it contains personal or sensitive information.
- Check eligibility: decide whether the question is low-risk and answerable from the approved knowledge sheet.
- Retrieve: supply only the relevant approved facts, not the entire customer history.
- Draft: ask the AI to produce a concise response in the customer’s language, using only the supplied facts.
- Validate: check names, prices, dates, availability, conditions, tone and any promised action.
- Respond or escalate: send only through the approved channel and hand off uncertain, sensitive or consequential cases.
- Learn safely: record the category and outcome without retaining unnecessary personal content in an informal prompt log.
5. Define clear handoff rules
A safe assistant must know when it cannot answer. Escalate when:
- The approved information does not contain the answer or sources conflict.
- The customer disputes a charge, requests a refund or makes a formal complaint.
- The request involves account access, identity verification, payment information or a change to personal records.
- The message suggests urgency, risk, harassment, discrimination or a safety concern.
- The customer asks for a person, or the assistant is uncertain about the language or meaning.
- The response would create a commitment, legal interpretation, discount or exception not explicitly approved.
The handoff should include a short neutral summary and the next action, not an unnecessary copy of the entire conversation. Tell the customer what will happen next without promising a response time the business has not committed to.
6. Protect customer information before using AI
Do not paste identifiable customer conversations into a general-purpose AI service by default. Remove names, phone numbers, addresses, order numbers and other identifiers unless a specific, approved workflow genuinely requires them. Never include passwords, one-time codes, payment credentials or identity-document images in a general chat.
Before processing personal data through an external provider, the business should assess the purpose, data categories, provider terms, retention, access controls, subprocessors and where data may be processed. In Morocco, Law 09-08 and the CNDP’s procedures may apply. The CNDP states that processing generally requires a prior declaration unless an exemption or prior-authorization regime applies; international transfers have separate conditions. The exact obligations depend on the processing and must be assessed by the responsible person or qualified counsel.
Use our AI data privacy checklist for small businesses as a starting point, and consult the CNDP’s official legal texts, its processing notification guidance and international transfer procedure. This article is operational guidance, not legal advice or a compliance certification.
7. A reusable instruction for a low-risk FAQ assistant
The following instruction is a starting point for a workflow using only approved public or non-personal business information. Adapt it to the provider and test it before use. It is not a security boundary by itself.
You are the customer-support drafting assistant for [BUSINESS NAME].
Use only the approved business information supplied in this session.
Reply in the customer's language: French, Moroccan Darija (Latin or Arabic script), or Arabic.
Keep the answer concise, respectful and clear.
Never invent prices, opening hours, availability, policies, delivery dates or promises.
If information is missing, conflicting or outdated, say that a team member must confirm it.
Do not request passwords, payment credentials, one-time codes or identity-document images.
For complaints, refunds, account changes, personal data or urgent matters, prepare a short handoff summary for a human.
Do not claim an action has been completed unless the supplied system confirms it.
If the customer asks for a person, stop drafting an automated answer and prepare the handoff.
Approved facts:
[PASTE ONLY THE RELEVANT, CURRENT, NON-SENSITIVE FACTS]Do not paste real customer records into this prompt unless the provider, account, purpose and data handling have been formally approved for that use.
8. Test before putting the workflow in front of customers
Create a small test set from anonymized or fictional examples. Include routine questions, misspellings, mixed-language messages, outdated offers, ambiguous voice-note transcripts, complaints and attempts to make the assistant ignore its instructions.
| Test | Pass condition |
|---|---|
| Correct business fact | Matches the approved source exactly |
| Unknown price or availability | Does not guess; asks or escalates |
| French / Darija / Arabic | Meaning and business facts remain consistent |
| Expired promotion | Does not present it as current |
| Complaint or refund | Uses the approved handoff path |
| Personal information | Does not request or repeat unnecessary identifiers |
| Prompt manipulation | Ignores instructions that conflict with the approved task and knowledge |
| Human request | Stops automation and prepares a clear handoff |
Record failures and correct the knowledge sheet, routing rules or instructions before launch. A successful demonstration is not proof of reliability; repeat the test after changing the model, prompt, provider, channel or business policy.
9. Measure the complete service, not chatbot activity
Do not judge success by the number of AI replies. Track whether the customer received a correct resolution and whether the workflow reduced work without increasing risk.
- Resolution: enquiries resolved without a repeat contact, using a consistent definition.
- Human correction: drafts that required factual, language or policy changes.
- Escalation quality: whether cases needing a person reached the right team with enough context.
- Response quality: accuracy, clarity, tone and consistency across languages.
- End-to-end effort: time spent preparing, reviewing, correcting and handling follow-up—not just generation time.
- Cost: subscriptions, usage, setup, staff review and maintenance.
- Risk: incorrect promises, privacy incidents, unresolved complaints and outdated knowledge.
Establish a baseline before introducing AI and compare equivalent periods and enquiry types. Do not publish percentage improvements unless the business has measured them and can explain the method.
10. A realistic 30-day pilot plan
- Week 1 — Map: choose one low-risk enquiry category, document the current process and prepare approved answers.
- Week 2 — Prepare: define language examples, handoff rules, privacy controls and a test set using synthetic or anonymized cases.
- Week 3 — Supervised pilot: let staff review every AI draft before it reaches a customer. Record errors and corrections.
- Week 4 — Decide: compare results with the baseline, review incidents and staff feedback, then keep, revise or stop the workflow.
This is a planning framework, not a guarantee that a workflow can be deployed in 30 days. Integration, approvals, provider access and business complexity may require more time.
Related AIToolSee resources
- AI data privacy checklist for small businesses — reduce data exposure before using AI.
- Build a practical AI stack for a small business — choose focused tools and measure their value.
- AI Guides & Tutorials — follow task-based implementation workflows.
- AI Use Cases — explore tools by the job you need to complete.
- How We Review AI Tools — understand our approach to evidence, limitations and human review.
Official references
- Moroccan CNDP — Textes et lois, including Law 09-08.
- Moroccan CNDP — Notify a processing activity.
- Moroccan CNDP — Transfers of personal data abroad.
- Google Search Central — Creating helpful, reliable, people-first content, for editorial principles on usefulness and transparency.
Editorial transparency: This is a vendor-neutral operating framework. AIToolSee has not claimed to test a specific provider integration or measured business outcomes for a customer. Product capabilities, channel integrations and legal requirements must be checked for the actual service and context.
Final takeaway
Start with a small group of routine questions, one approved source of truth and human approval for every customer-facing draft. Support the customer’s language, protect their information and measure the full resolution process. Expand only when the pilot shows that service quality improves without creating unacceptable risk.