Custom AI platforms in Tunisia: multi-tenant SaaS and AI agents

What is an AI platform, and how is it different from a web application?

A standard web application manages users and data. An AI platform adds features that read, write, analyse or act, and each call consumes a pay-per-use service. That detail changes the architecture: you need to know who triggered each call, for which client, with which data and at what cost.

When the platform is a SaaS, several companies use it at once. Each one expects its data to stay invisible to the others, its users to have different rights, and its subscription to match what it consumes. If your project has no AI features, our web application and SaaS page is a better fit. If you want to add AI to existing software rather than build a platform, see custom AI integration.

How do you keep each client's data isolated?

Data isolation is the first promise of any SaaS, and AI puts it to the test. An assistant searching "the documents" must search that client's documents and nobody else's. We apply a few constant rules:

• A workspace identifier on every record, with filtering enforced on the server, never only in the interface.
• Filtering before search: the documents passed to the model are selected from the client's workspace before the call, not sorted afterwards.
• Inherited permissions: an agent acts only with the rights of the user it assists.
• Separate logs, available to each workspace administrator.
• Automated tests that try to reach another client's data and must fail on every release.

For a Gulf group serving several subsidiaries, or a European franchise network, the principle is the same: each entity sees its own data, and head office sees everything if its rights allow.

How do you control AI costs: quotas, caps and billing?

Model providers charge by the volume of text processed. On a multi-tenant platform, one very active client can weigh heavily on the bill if nothing limits it. So we build consumption into the product itself: a counter per workspace and per feature, quotas included in each plan, alerts before limits are reached, and the option to buy extra credits or change plan.

Model choice is part of the equation. A simple task, such as classifying a message, does not need the most expensive model. We spread tasks across models according to the quality required and keep an abstraction layer so you can switch providers without rewriting the platform. Subscriptions, trials and plan changes are designed into the architecture, with a payment provider suited to your markets.

What does an agentic platform change?

An agentic platform does more than answer: its agents chain steps, such as reading a request, checking a record, preparing a quote and scheduling a follow-up. That is powerful, and it is also where mistakes cost most. So every tool given to an agent is a separate permission, binding actions wait for human approval, and every step is written to a log you can review.

Long tasks go through a queue, so the platform stays fast when several agents work in parallel for different clients. And we write down what an agent must never do before we write what it can do.

Which stack and which security for an AI platform?

We build with Next.js and TypeScript, PostgreSQL for data, n8n when an automation is better kept visual, and OpenAI- or Anthropic-class language models called through an API. Applications run in containers, on infrastructure you control or that we manage for you.

Security covers AI-specific risks on top of the usual ones: strong authentication, rate limiting, secrets never exposed in the browser, personal data reduced to the minimum before it reaches the model, and vigilance against malicious instructions hidden in a document or message. Depending on your markets, the framework is GDPR in Europe, Tunisia's Organic Law 2004-63 or Saudi Arabia's PDPL; legal sign-off remains with your counsel.

What do AzurLab and SkanLink show?

We apply these principles to our own products, which means living with the consequences:

• AzurLab, AD AZUR DIGITAL's CRM platform factory: an AI-augmented CRM with an assistant, analytics, email writing, automations, lead import, Meta ads and WhatsApp integrations, and client workspaces with isolated data. The application is private and requires sign-in; an agentic version is on the way.
• SkanLink, our live SaaS for digital business cards (QR and NFC): smart scanning of a paper card into a contact, a personal assistant, a chat assistant on every card, AI usage quotas and subscription plans, all multi-tenant.

Our other projects show the care we put into interfaces.

How does an AI platform project run with AD AZUR DIGITAL?

We start with the business model as much as the technology: who pays, for which plan, and what an active client costs in AI usage. A clickable prototype follows, shown to prospective users before any development, because a sound architecture does not prove that a market exists. Development moves in batches, from the foundation (accounts, workspaces, permissions, billing) to AI features, on a staging environment you test at every stage.

Our team is based in Sfax and works in French, English and Arabic, with a right-to-left Arabic interface designed from the start when your markets need it. The first step is a free audit of your project: we will tell you what calls for a platform, what can start more simply and which risks to address first. If what you mainly need is an assistant answering your customers, an AI agent as a service is quicker to set up.