AI agents · Morocco
AI agency in Morocco: custom AI agents and automation
We build AI agents that do one job in your business properly: answer customers, qualify a request, chase a quote, write the weekly report. Wired to your own data and tools, with the failure cases handled. Based in Marrakech, in French, Arabic and English.
Most AI projects stop at the demo.
An agent that shines in a meeting and fails on the third real customer is a liability. What decides whether it earns its place is the unglamorous part: what it is allowed to say, what it does when it does not know, when it hands the conversation to a person, and what it costs each month. That part is the job.
Already running. You can test the first one now.
Nothing below is a mock-up. The first item is the assistant of this site: open it and ask it something hard.
The assistant on this site
Answers in English, French and Arabic from a knowledge base generated from the site's own pages and price grid, never typed twice. A code gate removes any figure we have not published before it reaches you, and a qualified request reaches the founder as a written brief. Runs on open-weight models (gpt-oss) with a fallback chain across two providers.
Marrakech Private's back office
Booking-recovery jobs, weekly sponsor digests, post-stay follow-ups, programmatic pages regenerated from structured data, and a vision model that turns a screenshot of a lead into a CRM entry. In production on our own platform.
Our own monitoring
Every day a job pulls Search Console for our sites and raises alerts. A probe asks Gemini, grounded on Google Search, the questions our buyers ask, and records which sites it cites.
An MCP server
It gives Claude read access to Search Console, Analytics and that monitoring history, so an audit starts from measured data instead of a guess.
Five steps, in this order.
An agent is a small piece of software with a job description. We write the description first.
The job, written down
Which task, for whom, what a good outcome looks like, and what the agent must never do. One page, agreed before any code.
Your data and tools
The documents it answers from and the tools it acts on: CRM, calendar, spreadsheets, email, your website, WhatsApp through Meta's official API.
The guardrails
Answers come from your sources only; figures are checked by code, not by the model; a person takes over, with the context, when the agent should stop.
Production
Hosted in your accounts, logged, with a monthly running cost you can read line by line.
The monthly review
What it answered, where it handed over, what it got wrong. Fixed in the knowledge base or the rules, not with a longer prompt.
Published prices.
A one-off setup, then an optional monthly retainer for hosting and improvements. Model and API usage is itemized separately, never hidden inside a round number.
See the full grid- Quick wins: 3-5 automationsfrom 1 500–3 000 + 200–400/mo EUR15–25K + 2–3K/mo MAD
- Workflow: 8-12 automations + 1 agentfrom 4 000–8 000 + 400–700/mo EUR30–50K + 3–5K/mo MAD
- Custom: agents + RAG + data pipelineson request
Ranges from the pricing grid in force. A custom agent is quoted after a written scope; these ranges cover the automation work we publish.
Four things we tell you before you sign.
We integrate models, we do not train them. Claude, GPT and open-weight models such as gpt-oss handle the language. What we build is everything around them: your data, the tools, the rules, the hand-over to a person.
WhatsApp goes through Meta's official Cloud API, and we have not delivered one to a client on it yet. Unofficial libraries get numbers banned. If a WhatsApp assistant is your project, it will be our first client delivery on that API, and we scope and price it that way.
An agent is scoped before it is priced. The ranges below cover the automation work we publish. A custom agent gets a written scope first, never a catalogue price.
It costs something every month. Model usage, hosting and messaging fees are real. They are itemized and estimated with you before anything is built.
What people ask before they sign.
If your question is not here, ask it on WhatsApp: the founder answers.
Read further
Bringing AI into a Moroccan SME: where to startHow much does an AI agent cost in Morocco?
At ZAOURI, published automation work is 1 500–3 000 EUR of setup, then 200–400 EUR a month, or 15–25K MAD then 2–3K MAD a month for Moroccan clients. A custom agent is quoted after a written scope. Model and API usage is itemized separately.
AI agent or automation: what is the difference?
An automation follows fixed steps: when a quote goes out, send a reminder three days later. An agent reads a situation and chooses within limits you set: answer, ask a question, hand over. Most projects need both, and the automation usually comes first because it is easier to measure.
Can you build a WhatsApp assistant?
Yes, on Meta's official Cloud API, the only way to run one without risking the number. We have not yet delivered one to a client on that API; if it is your project, the scope and the quote say so.
Will you train a model on our data?
No. We connect existing models to your documents and tools. Your data stays in your accounts, and we do not use it to train any model.
Who builds it?
The founder, Thibaut Campana, who also built the assistant on this site. No sales team, no chain of subcontractors.
How do we start?
With a short written brief: the task, the tools involved, the volume. The assistant on this page can collect it for you, or send it on WhatsApp. We come back with a written scope.
Let's make you the answer.
Tell us what you're building. We'll tell you, straight, whether and how we can make it visible.
Direct line to the founder. No bots, no relay.
Prefer to write it down instead? Send the brief by form