How do you choose an AI agency in Tunisia: which questions should you ask?
Data, models, made-up answers, evaluation, ownership, usage costs: the ten AI-specific questions to ask before signing, the warning signs and the proof to demand.
Founder and managing director
In short
To choose an AI agency in Tunisia, do not judge the demo: judge how the provider handles what is specific to AI. Ask ten questions. Where does your data go, and is it used to train a model? Which model is used, and can it be swapped? How does the agent avoid inventing answers? How is quality evaluated before launch? What happens when it gets something wrong? What does usage cost, and who pays? Who owns the instructions, the knowledge base and the conversations? Does it handle French, Arabic and derja? Who monitors it after launch? Which products are in production? Be wary of promises of full autonomy, a "proprietary AI" that is never explained and the absence of any cost estimate. Demand a demo on your own data, a test set, a usage cost simulation and a reversibility clause.
Why is choosing an AI provider different from choosing a web agency?
A website is judged on screen: you see what you buy. An AI agent or feature gives different answers on every use, spends budget on every call and handles your data continuously.
General criteria still apply, and we cover them for choosing a digital agency and choosing an automation provider. This article focuses on what only AI adds: data, models, errors, evaluation, ownership and usage costs.
What are the ten questions to ask an AI agency?
For each one, here is what a good answer contains.
- Where does our data go, and is it used to train a model? A good answer names the vendors, the hosting countries and the contract terms that exclude training. In Tunisia the reference framework is Organic Law No. 2004-63 of 27 July 2004 on personal data protection, overseen by the INPDP.
- Which model do you use, and can we switch? The provider should be able to change models without rebuilding everything.
- How does the agent avoid making things up? Expect specifics: it relies on your documents, points to its internal sources and says "I don't know" when information is missing.
- How do you evaluate quality before launch? There should be a test set written with you, clear pass criteria and a measured result before each release.
- What happens when the agent is wrong? Human handover, searchable conversation history, tracked fixes.
- What does usage cost, and who pays? Model vendors bill by consumption, and channels such as WhatsApp have their own charges. Ask for a cost estimate per conversation or task, and a cap.
- Who owns the instructions, knowledge base, code and conversations? You should, with export written into the contract.
- Do you handle French, Arabic and derja? Test with real customer messages, including derja typed in Latin characters.
- Who monitors the agent after launch? Ask who reads conversations, how often, and what the report contains.
- Which AI products do you run in production? Not mock-ups: tools used daily by real users that you can try.
Unsure what profile you need? When to call an AI expert in Tunisia helps you decide.
Which warning signs should you watch for?
- Promises of full autonomy, with no human handover or scope limit.
- A "proprietary AI" never explained, often an off-the-shelf model behind the scenes.
- No usage cost estimate, or "it's negligible".
- No test set: quality is "checked by hand" on launch day.
- Accounts opened in the provider's name with model vendors or Meta, with no transfer planned.
- Refusing to work on your real data before signing, even anonymised.
Want to put your project through these questions? Send it to us: we answer in writing, point by point.
What proof should you demand before signing?
| Proof | What it shows | How to check it |
|---|---|---|
| Demo on your data | The agent works on your case, not an ideal one | Supply real documents and questions |
| Test set and results | Quality is measured, not assumed | Ask for the questions and pass rate |
| Usage cost simulation | The monthly budget is predictable | Vary the conversation volume |
| AI product in production | The team can operate, not just prototype | Use the tool yourself |
| Reversibility clause | You can leave with your data | Read the contract before signing |
| Reachable client | The provider keeps commitments over time | Call them without the provider |
On our side, we point to two platforms in production: AzurLab, an AI-augmented CRM, and SkanLink, a digital business card platform with smart scanning. Other projects are in our portfolio.
What if you also sell in Europe?
The EU AI Act requires, among other things, that people be told they are interacting with an AI system, with this transparency obligation applying from 2 August 2026 (source: Regulation (EU) 2024/1689, Article 50). Adjustments to the timeline are being discussed in Brussels, so check the version in force at launch. A Tunisian company serving European customers should ask its provider how these obligations are covered.
Where should you start?
Before talking to agencies, map your processes and data. Our automation audit exists for exactly that, and our AI and automation practice covers everything from AI agents as a service to custom platforms.
Related reading — Automating your processes with n8n: where to start · AI chatbots and customer service: what actually works
Read next
A project in this area?
Two minutes to tell us what you need, and we come back with a first read of your situation.
Reply within one business day