OpenAI just rolled out vision and voice for ChatGPT. You can now show it a photo and ask questions about it, or talk to it out loud and hear it answer back. For anyone who has only ever typed at these models, this is a bigger shift than it looks, and the multimodal AI business uses are surprisingly practical.
The reason it matters: most real work in a warehouse, a workshop, or a shop floor is not text. It is a broken part, a messy shelf, a handwritten note, an error screen nobody understands. Until now, using AI meant first translating that physical reality into words yourself. Now you can just point a camera at it.
I want to walk through where this is genuinely useful for an Indonesian SME today, and, just as importantly, where it will confidently tell you something wrong. Because it will, and knowing that is the whole game.
Concrete multimodal AI business uses
These are not hypotheticals. They work with a phone and a ChatGPT app right now.
- Troubleshooting an error screen. Your point-of-sale throws a cryptic error and your one IT-literate staff member is out. Photograph it, ask what it means and what to try first. You get a plausible first step in seconds instead of waiting.
- Reading a broken part. A machine component fails. Photograph it and ask what it is and what a likely replacement is called. You walk into the parts shop knowing the name instead of pointing helplessly.
- Merchandising checks. Photograph a shelf and ask whether facings look full, whether anything is obviously out of place, or to describe the layout. Useful for a small chain where the owner cannot visit every outlet.
- Digitizing handwritten notes. Snap a photo of a handwritten order list, a stock count, or a delivery note and ask it to transcribe into clean text you can paste into a spreadsheet.
- Understanding a document fast. Photograph a dense contract clause or a form in a language your staff struggles with and ask for a plain summary before a human verifies it.
The voice side adds a different kind of value. A driver or a field worker with their hands full can talk to it. A staff member who reads slowly can hear an explanation instead. For frontline roles where typing is awkward, voice lowers the barrier to actually using the tool.
The accuracy problem: verify before you act
Here is the part the launch demos will not stress, and I will. Vision AI is impressive and unreliable at the same time. It reads a shelf correctly nine times and invents a product on the tenth. It transcribes handwriting well until one messy digit turns Rp 150.000 into Rp 750.000.
The rule is simple and non-negotiable: use it to draft and to orient, never to decide unchecked.
- For a transcribed order list, a human confirms the numbers before you fulfill or invoice.
- For a diagnosed error, treat the suggestion as a lead, not a verdict, especially before anyone touches wiring, money, or a machine.
- For a document summary, read the actual clause before you sign anything.
This is the same discipline I apply to every AI tool. The model produces confident output at speed, and confident wrong output is still wrong. You stay the reviewer. I wrote more about this reviewer mindset in the context of AI agent frameworks, and it applies doubly when a camera is involved, because a photo feels like objective truth and the model's reading of it is not.
Where it is safe, and where it is not
A quick way to decide whether to trust a multimodal answer: what does a mistake cost?
| Task | Cost of error | Verdict |
|---|---|---|
| Summarizing a shelf photo for a vibe check | Low | Use freely |
| Transcribing a handwritten stock count | Medium | Use, then human-verify numbers |
| Diagnosing a POS error message | Medium | Use as a lead, confirm before acting |
| Reading financial figures off a document | High | Use to draft, never to finalize |
| Anything involving safety or a legal signature | High | Human does the real reading |
Keep AI in the low and medium boxes and you get most of the speed with little of the risk. Push it into the high box unsupervised and it will eventually cost you more than it saved.
Practical setup for an SME
You do not need a project to start. You need one habit.
- Pick one repetitive visual task your team already does slowly, such as transcribing daily order notes or checking outlet shelves.
- Try it for a week with a real staff member and their phone.
- Track the error rate. How often does a human have to correct it? If corrections are rare and quick, keep it. If they are frequent, that task is not ready.
- Write the one rule for that task: what a human must always verify before acting.
That is the entire adoption plan. Small, measured, reversible. This is the same measured approach I recommend before any AI investment, which is why I usually point owners to an honest AI readiness check before they get swept up in a launch.
The practical takeaway
Vision and voice make ChatGPT genuinely more useful for the physical, non-text work that fills a real business day. The best multimodal AI business uses are the boring ones: reading a part, transcribing a note, orienting on an error, summarizing a document.
My recommendation: pick exactly one visual task this week, test it with your team, and set the "human verifies before acting" rule from day one. Treat the model as a fast, sharp intern with no accountability. It sees a lot and it is sometimes wrong, so you check its work. If you want help figuring out which of your workflows are worth pointing a camera at and which are too risky, that is a good conversation to have with a technology partner before you build any habit around it.