AI models such as Claude, ChatGPT, and Gemini are now fairly fluent in Indonesian. For general conversation the results usually sound natural. The trouble starts when AI is put to work on everyday business tasks. A customer writes "kak, yg ukuran L ready ga? mau co skrg" (roughly, "is size L in stock? want to check out now"), staff say "SJ" when they mean a delivery note, and supplier invoices use a number format that differs from the American one most software assumes.
This article looks at the challenges of using AI with the Indonesian language in a business setting, from mixed language and local terms to tone and OCR of Indonesian documents. It also covers how to test with real data before customers or your team rely on it.
Mixed language and abbreviations
Indonesians rarely write in one tidy language. In business chats, Indonesian mixes with English ("invoice-nya sudah di-forward ya"), with regional languages such as Javanese ("nggih, matur nuwun"), and with abbreviations that only make sense in context. "PO" can mean purchase order to a distributor but pre-order to an online shop. "TF" (transfer), "DP" (down payment), "COD", and "ongkir" (shipping cost) show up in almost every sales conversation.
Modern AI models generally handle mixed language well, but they can still misread an abbreviation with several meanings. The fix is simple. Describe your business context in the system instructions, for example that you are an online fashion store and "PO" means pre-order. Include a few real customer messages as well, along with the correct way to reply.
Decide on the reply language too. A guesthouse in Yogyakarta, for instance, gets messages from local and international guests. A clear rule, such as replying in whatever language the guest used, stops the AI from answering an English question in Indonesian or the other way around.
Local terms and your own business vocabulary
Every industry has its own vocabulary. Contractors use SPK (work orders) and BAST (handover reports), clinics talk about poli (outpatient units) and rujukan (referrals), and notary offices deal with akta, salinan, and minuta. Your company probably also has internal names for products, branches, or processes that appear in no dictionary.
If the AI does not know these terms, its answers sound off or are simply wrong. Build a simple glossary with each term, its meaning, and an example of how it is used, then add it to the AI's context. In a RAG system that answers from company documents, a glossary also improves search. A staff question about the "BAST for the warehouse project" can then find a document titled "Berita Acara Serah Terima".
Formal or casual
Indonesian has a wide range of tone. Letters to government offices use standard formal language and "Bapak/Ibu". WhatsApp replies from online stores usually use "Kak" and a more relaxed style. Clinics and professional services often sit in between, polite without being stiff.
AI tends to fall back on one default style, often too formal or too eager. Write down the style you want. Specify the form of address, sentence length, whether emoji are allowed, and words to avoid. Better still, give the AI a few of the best replies your own staff have written as a reference. AI imitates examples more reliably than it follows general descriptions.
OCR for Indonesian documents
Many business documents in Indonesia still arrive on paper or as scans. Handwritten receipts from a building supply store, invoices with a company stamp covering the figures, blurry photocopies of a KTP (national ID card), or forms filled in by pen. OCR trained mostly on English documents can struggle with these.
Number and date formats need special attention. In Indonesia, "1.500.000" means one and a half million, and the comma marks decimals. A system that follows the American format can read these figures wrongly, and that kind of error is very costly if it reaches the books. A date written "05/10/2026" means 5 October to an Indonesian reader, while an American-format system reads it as 10 May. Make sure your system sets these formats explicitly, and flag low-confidence documents for human review. For ID documents such as the KTP, mask the NIK after reading it if the number is not needed in later steps.
Testing Indonesian-language AI with real data
A demo with tidy sentences almost always works. The real test is the messages and documents that reach your company every day. These steps help you test fairly:
- Collect real examples from customer chats, emails, and documents, with personal data masked.
- Make sure the examples include mixed language, abbreviations, typos, and the regional language your customers use.
- Include scanned documents that are blurry, skewed, or handwritten.
- Write down the correct answer or extraction result for each example.
- Ask the staff who normally handle the work to judge the tone and accuracy of the AI's answers.
- Repeat the tests whenever the prompt, glossary, or model changes.
These results also help you choose a model. A model that leads on general benchmarks is not necessarily the best fit for the way your customers write.
AI that understands how your business talks
Through our Company AI service, Averoa connects AI to your internal documents and data, with answers that cite their sources, access set per team, and NIK and NPWP masked before data reaches the model. We set it up to understand Indonesian and the business terms your team uses every day.
