AI development and automation for businesses in Morocco

Useful AI starts with a specific task, repeated often, whose cost you know. The use cases we build:

Before you commit a budget, we put numbers on it. That’s the deliverable of our “Choosing and integrating AI” advisory: costed use cases and a prototype.

1. 1

   List the repetitive tasks

   With your teams: who does what, how many times a week, how long it takes, what goes wrong.

2. 2

   Estimate the gain

   Time saved, response time, errors avoided: each use case gets an estimate, weighed against the cost to build and run it.

3. 3

   Prototype on your real data

   A prototype on a sample of your documents, with personal data masked, to check that the quality is there.

4. 4

   Build an evaluation set

   Real questions and cases with the expected answer. It measures quality before go-live, then after every change.

5. 5

   Pilot with human review

   A group of users, checked answers, collected feedback. We widen the rollout when the numbers hold.

6. 6

   Go live and monitor

   Monitoring, running costs, answer quality and a monthly report.

RAG (retrieval-augmented generation) means the assistant first searches your documents, then writes its answer from the passages it found, citing its sources. The model doesn’t need retraining: when a document changes, the answer changes.

Quality depends less on the model than on the preparation: splitting documents cleanly, keeping tables readable, respecting access rights (a salesperson has no business reading payslips) and being able to say “I don’t know” when the information doesn’t exist. That’s where most of the work goes.

A language model can be wrong with total confidence. We treat it that way:

• Evals before go-live on a set of real cases, then after every model or prompt change
• Guardrails: allowed topics, answers limited to sources, refusal when information is missing
• Human review for any decision that matters: in HireSense, a person reviews the report before any decision
• Personal data masked before processing whenever possible
• Zero data retention options with the model provider
• Document access filtered by each user’s permissions
• An NDA signed before you show us a single document

HireSense shows how we build AI that touches sensitive decisions. The assistant runs the interview in writing or by voice, with a text fallback; it scores answers against a rubric, backed by evidence, and produces a structured report. Personal data is masked and a person reviews before any decision. It works in French, English and Arabic, and integrates with an ATS (applicant tracking system).

An assistant on its own does little. We connect it to your CRM, ATS, help desk or business software, so it reads the right data and puts its work where your teams already are.

We work with models from Anthropic (Claude), OpenAI and Mistral, and choose per task: quality in French and Arabic, speed, cost per request, data constraints. The architecture lets you switch models; the evaluation set tells you whether the new one does better.

A focused assistant, built on a well-defined process and set of sources, typically takes 6 to 10 weeks, prototype and evals included. Projects are custom-quoted. Also plan for a monthly running cost (model calls, hosting, document search) that varies with volume: we estimate it before building, then track it in the monthly report.