Short answer. An AI agent only works if it has something to read and something to do: a schedule to follow up, rules to answer from, a back-office to write into. The three uses that pay in Algeria, payment and order reminders, customer support in Arabic, French and darija around the clock, and qualifying WhatsApp requests, all share that prerequisite. Writing posts or summarising documents yields little; replacing a salesperson or deciding credit is off topic. Then one autonomy level at a time: the agent answers, then proposes, then acts within limits, then acts and reports. The first month contains no AI: it puts the data into a system. It is the month that decides whether the next five are worth anything.
This guide extends our roadmap for digitising an SME in 2027, where AI arrives in the fourth quarter, after collection, the flow and the back-office. It answers owners who write to us “I want to add AI” without knowing where, and founders who want an agent in their SaaS or their app.
The uses that pay, and those that are useless
Seven uses, ranked by gain for an Algerian SME, each with what must exist first. The right-hand column is the important one.
Reminders are the most profitable use, because they deal with money that already exists and is not being recovered: the unpaid instalment of a school, the unconfirmed order of an e-merchant, the late invoice of a wholesaler. The agent follows up at the right time, in the customer's language, with the exact amount and how to pay, and raises to a human those who do not respond. Prerequisite: the schedule or the orders live in a system.
Trilingual support answers at night and on Friday, from your rules and your back-office: where is my order, what are your hours, how much do I owe. It understands Arabic, French, darija and Arabic written in Latin letters, because that is how your customers write. Prerequisite: written rules and a status readable somewhere. Qualifying WhatsApp requests turns the message flood of a real-estate agency or a carrier into sorted requests, with complete information, dropped into a pipeline. Prerequisite: a pipeline.
The daily summary, document entry and copywriting are useful but secondary. The first requires one single base; the second, a volume of paper; the third requires nothing, which is why everyone already has it and nobody gains an edge from it. Replacing a salesperson or deciding credit is off topic: the agent prepares, the human decides. A company that lets an agent promise a price or grant a payment delay without limit regrets it within a week.
«An agent plugged into a notebook has nothing to read.»
The four autonomy levels, in order
The question is not “which AI”, it is “how much autonomy, and when”. Four levels, climbed one by one.
Level 1, it answers: read-only, status, hours, balance, rules. No risk, an immediate gain on repetitive questions. Level 2, it proposes: it drafts the reminder, suggests the appointment, prepares the reply, and a human validates before sending. This is the level where the team learns to trust it, and where you discover its mistakes before a customer sees them. Level 3, it acts within limits: the reminder goes out, the appointment is confirmed, a goodwill gesture is granted under a threshold you set; beyond, it escalates. This is the level that pays. Level 4, it acts and reports: every evening, what it did, what it could not do, the exceptions, the numbers. The human supervises instead of executing.
Jumping from level 1 to level 4 in a month is the surest way to lose the trust of the team, who will watch the agent fail in public, and of the customers, who will receive a wrong reminder. About one level per month, each measured.
How the agent fits into the company
An agent is not a chatbot dropped onto a website. It is five layers, of which the third, the tools, makes the difference between an agent that talks and an agent that works.
The channels are where your customers already write: WhatsApp Business first, your app or website next, e-mail for B2B requests. Same rules everywhere. The agent understands the three languages and their mixes, knows your rules, what it may say, do and promise, and escalates to a human as soon as it is unsure. The tools are reading and writing in your back-office: read an order, write an appointment, update a schedule, drop a request into the CRM. Without them the agent is a polite answering machine. The data lives in one single base, with hosting chosen with Law 18-07 in mind, and every action of the agent is logged: who said what, who did what. Control is human: thresholds, weekly review of conversations, numbers.
Where the agent runs, and how it is paid. AI models are rented per use, in foreign currency, from foreign providers; we detailed how to pay for cloud and AI from Algeria. For an SME the usage cost is small next to a salary; what costs is the integration with the tools, and that is what gets scoped in writing with a total price. For companies whose data must not leave the country, clinics, insurers, administrations, open models can be hosted on local infrastructure; it is heavier, and it is decided at scoping, not after.
«Without the tools layer, the agent talks; it does not work.»
Six months, one level at a time
The calendar of an AI agent in 2027 starts with a month without AI.
Month 1, prerequisites: the schedule, orders or requests enter a system; company rules are written, hours, return policy, delays, what you promise and what you do not; the back-office exposes what must be read and written. If this month reveals that nothing exists in a system, the right decision is to build the back-office first, as we explain in the SME roadmap. Month 2, level 1 on WhatsApp and in the app. Month 3, level 2, drafts validated by the team, mistakes fixed in the rules. Months 4 and 5, level 3, reminders sent, appointments confirmed, thresholds set and respected. Month 6, level 4, daily summary, measurement, review: response rate, unpaid bills recovered, requests qualified, team time freed.
An agent of this scope, three channels, three languages, connected to an existing back-office, sits in category 2 of our price grid, with the back-office to build on top if it does not exist.
The mistakes we see with AI in companies
- Plugging an agent into a notebook. It has nothing to read. Data in a system first.
- Starting with maximum autonomy. A wrong reminder sent to a hundred customers. One level per month.
- An agent that only understands French. Customers write in darija and Latin-script Arabic. The three languages and their mixes, from level 1.
- No escalation to a human. The agent makes things up when it does not know. It must be able to say “I'll pass this on” and do it.
- No thresholds. A goodwill gesture without limit, a delay granted to everyone. Written thresholds, beyond which it escalates.
- No log. A customer disputes, nobody knows what the agent said. Every conversation and every action is logged.
- Measuring nothing. Without response rate, recovered bills and qualified requests, the agent is an expense. One number per use, read every week.
Frequently asked questions
- What is the first use of an AI agent in an Algerian SME?
- Payment and order reminders, because they deal with money that already exists and is not being recovered, then customer support in Arabic, French and darija around the clock, then qualifying WhatsApp requests. All three require the information to live in a system.
- Can an AI agent answer on WhatsApp in darija?
- Yes, provided it is designed for the three languages and their mixes, including Arabic written in Latin letters, and escalates to a human as soon as it is unsure. WhatsApp Business is the first channel, because that is where customers already write.
- What must exist before adding an AI agent?
- Data in a system, schedule, orders or requests, written rules, and a back-office the agent can read and write. Without that, the agent talks but does not work. If nothing exists, the back-office is built first.
- How far should the agent act on its own?
- Up to level 3: it acts within limits you set, reminder sent, appointment confirmed, goodwill gesture under a threshold, and escalates beyond. Replacing a salesperson or deciding credit stays off topic: the agent prepares, the human decides.
- Where does the agent run and how is it paid from Algeria?
- Models are rented per use in foreign currency from foreign providers; the usage cost is small for an SME, the integration is what gets scoped. For data that must not leave the country, open models can be hosted on local infrastructure, which is decided at scoping.
- How long until the agent is useful?
- Six months: one month of prerequisites without AI, then about one autonomy level per month, it answers, it proposes, it acts within limits, it reports. Each level is measured before moving to the next.