Building AI agents that speak Darija
A conversation on localisation, data scarcity and why language is the real moat for AI agents serving Moroccan customers and public services.
Conversational AI built for English or French breaks down in Moroccan Darija. We sat down with a founding team building agents for the local market to understand why language is a defensible moat — and a hard one.
On the problem
Most enterprise support, sales and citizen-service interactions in Morocco happen in Darija and Arabic. Off-the-shelf models handle them poorly, so the customer experience suffers exactly where it matters most.
On the moat
Anyone can wrap a model. Very few have the data, the dialect coverage and the evaluation discipline to make it actually work in Darija. That’s the moat.
On going to market
The team’s path mirrors what we see work repeatedly: start with a narrow, high-value use case (support deflection), prove it with one enterprise, then expand. Distribution — getting in front of banks, telcos and public services — is often harder than the technology.
Takeaway
Localisation isn’t a feature; for this market it’s the product. Founders who treat language as core, not cosmetic, are the ones building something durable. Browse more teams like this in the directory.