Arabic changes the product.
Building in English and adding Arabic at the end doesn't work, because Arabic changes the product itself.
A pattern we see in AI projects across the Gulf: build everything in English, get it working, then "add Arabic" as the last step. It doesn't work, because Arabic changes the product itself.
This applies to any second language your customers actually use, but Arabic is the one we meet most, and it makes the point sharply. Four things break when it arrives late.
The layout flips, and so does the logic
Right to left isn't a mirror. Numbers still read left to right. Icons that point forward now point the wrong way. A progress bar that filled from the left now has to fill from the right, and the code that drew it has assumptions baked in from screen one. Retrofitting this is weeks of small, tedious fixes. Designing for it from the first screen costs almost nothing.
Formal and casual Arabic are two different tones
An assistant that's friendly and light in English needs a deliberate decision about register in Arabic. Modern Standard Arabic reads as official. Gulf dialect reads as human. Neither is wrong, but a product that switches between them without meaning to feels broken. This is a product decision, and it needs to be made by someone who speaks the language, early.
Search that works in English falls apart
Retrieval, the part of an AI product that finds the right passage in your documents, is where Arabic hurts most. Plurals and roots behave differently. Diacritics appear in some documents and not others. Real documents switch between Arabic and English in the same sentence. Default chunking and embedding settings tuned on English text will return the wrong passages, and the model will answer confidently from them.
We tune retrieval on the client's actual documents in both languages before we trust a single answer.
You're only testing half your users
If the test set is in English, every release is checked against English users only. The Arabic users find the bugs in production. A test set has to cover both, from the first release, or the numbers on the dashboard describe half the product.
Arabic on the first screen, the first test, the first eval. Not step twelve.
Our rule
Every language the product will ship in is present on the first screen, in the first test set, and in the first round with real users. It is one of the seven lines we hold ourselves to, and it is the one that saves the most money, because it is the one that's most expensive to fix later.