Multimodal AI for Business: Voice, Images, and Their Limits

Most work in Indonesian businesses doesn't arrive as tidy text. Customers send WhatsApp voice notes, suppliers send photos of receipts, field staff photograph damaged goods, and important meetings exist only as a recording on someone's phone. Multimodal AI is AI that can understand inputs like these, not just text. This article covers what it can do, realistic use cases, and the limits to understand before you rely on it.

What multimodal AI is

A "mode" here means a type of input: text, images, or sound. A typical chatbot only takes text. Multimodal AI can take more than one type at once, such as a photo of a receipt plus the question "what's the total?". Several large AI models, such as Gemini, can already take text, images, and audio in a single request.

For a business, this means AI can sit at points that used to always need a person: listening, looking, and then copying into a system. AI doesn't replace decisions, but it can save the time spent retyping and summarising.

Use cases for voice

• Voice notes into written orders. A customer's voice note becomes text, and the key points, such as item, quantity, and address, are summarised for an admin to check.
• Meeting recordings into minutes. The recording becomes text, then a summary of decisions, tasks, and owners.
• Field reports by voice. Staff just talk into their phone, and the report lands in a consistent format.
• Assistants that answer by voice. Good for frequently asked information, such as opening hours or order status, with a quick route to a person.

Use cases for images

• Receipt or invoice photos into data. Date, shop name, items, and total are read from the photo and entered into a spreadsheet or bookkeeping system.
• Paper forms into digital data. Useful for businesses that still take sign-ups or surveys on paper.
• Product photos into draft catalogue copy. AI drafts the name, colour, and features, then your team edits before it goes live.
• Customer complaint photos. Photos of damaged or wrong items are sorted first, so your team can prioritise the urgent ones.

Limits you should know about

Multimodal AI is very helpful, but it isn't perfect. The most common limits:

• Noise and accents. Recordings made on the street or at a market, mixed regional languages, and your business's own jargon can be misheard.
• Blurry photos or handwriting. Folded receipts, faded ink, or hard-to-read handwriting lower accuracy.
• Numbers can be wrong. AI sometimes misreads numbers or invents content that isn't there. For totals, prices, and quantities, always build in a check.
• Business context. AI doesn't know your internal rules, such as item codes or discount limits, unless you give them to it.

That's why a good system doesn't save AI output straight away. The output is shown as a draft, uncertain parts are flagged, and a person gives final approval. Over time you can see how often drafts need correcting, then decide which parts are safe to automate further.

Privacy and personal data

Voice and photos often contain personal data: names, phone numbers, addresses, even faces. In Indonesia, processing this kind of data is governed by the Personal Data Protection Law (UU PDP). A few things to prepare from the start:

1. Tell customers that their voice notes or photos are processed with the help of AI.
2. Send the AI service only the data it actually needs. Avoid sending identity documents when you don't have to.
3. Decide how long recordings and photos are kept, then delete them on schedule.
4. Limit who on your team can open the original recordings and photos.

Cost: voice and images are heavier than text

Processing audio and images usually costs more than text, because the data is much larger. You can keep costs in check by shrinking photos before sending them, trimming silent parts of recordings, using a lighter model for simple jobs, and setting a daily spending cap. It's the same principle we apply to the Numekara AI tutor, which uses fallback across models along with daily quotas and spending caps.

How to start safely

1. Pick the one workflow that takes up the most time, such as transcribing order voice notes or retyping receipts.
2. Collect 20 to 50 real examples from your business, including poor-quality ones, with personal data masked.
3. Test AI on those examples and note how many are right, how many need correcting, and the kinds of mistakes.
4. Set up the workflow with a human check, then monitor the results for a few weeks.
5. Only once the results are stable, consider automating further.

A small pilot like this gives you an honest picture before you spend a lot. If accuracy isn't good enough yet, you find out early and can choose another route, such as tidier forms or a different split of work.

Multimodal AI is most useful when it prepares drafts from voice and images, and people check and decide.