The BCG number stopped me mid-scroll.

92% Indonesia ranked first globally, not the US and not the UK, for worker GenAI usage rate. I was in the middle of a pretty bad week at work when I saw it, and I remember thinking: someone is going to use this number to make a really confident pitch deck, and it’s going to miss the point entirely.

Maybe that someone was past me.

A bit of context: I run Pionedge, a digital product studio in Indonesia. We build things. AI has been central to what we ship for a while now. So when I talk about deploying AI in this market, I’m not talking about strategy decks or hypotheticals. I’m talking about the specific sensation of watching a production model fail because a real user from Malang typed the way people in Malang actually type.

That experience will recalibrate your relationship with adoption statistics pretty quickly.

Here’s the thing about 92%.

It’s true. I’m not arguing with the data. But “using AI” in Indonesia covers an enormous range of behaviors from a solopreneur using ChatGPT to draft a caption for their Tokopedia store to an enterprise team running LLM-powered document processing at scale. Both count. They are not the same problem.

The gap between those two things is where products live or die.

When we first shipped one of our AI features, our internal testing looked great. The model handled queries well. Response quality was solid. We were quietly proud of it.

Then real users showed up.

The first week of production data looked nothing like our test data. Not a little different, completely different. Because in our testing, users knew they were being evaluated. So they typed carefully. Full sentences. Clear intent.

Real users typed like they were texting a friend, which in Indonesia means some combination of Bahasa Indonesia, Javanese, abbreviations from years of WhatsApp habits, and whatever English tech vocabulary they’d absorbed from YouTube. Not because they’re unsophisticated, because that’s how people communicate when they’re not performing for anyone.

Something like: “kak ini bisa gak dipake buat usaha aku, soalnya aku suka lupa2 sama pengeluaran”

Our model, trained on cleaner inputs, had no idea what to do with that. The intent is completely obvious to any human. To the model, it was noise.

We had to rebuild the data collection process. Longer than I want to say. And the lesson wasn’t technical; it was about assumptions. We’d built a system optimized for the version of our users we imagined, not the ones who actually showed up.

The second thing that surprised me was about trust, and it took me longer to notice because it was quieter.

There’s an assumption most AI products are built on: users will naturally understand that AI can be wrong. That they’ll verify things that matter. That the “AI disclaimer” in the footer does some actual work.

In some contexts, maybe.

In ours, particularly with users outside Tier 1 cities, we kept seeing something that made us uncomfortable. People treated AI outputs as final. Not because they were gullible. Because that’s a completely reasonable response to years of being told that AI is a sophisticated, powerful technology. If the computer says it, why would you second-guess it?

We ended up adding what we called internally a “humility layer.” When confidence was low, we didn’t just flag it with some small UI indicator nobody reads. We said it, in plain Bahasa Indonesia, in the middle of the response: “Saya kurang yakin dengan ini, sebaiknya dicek dulu.”

It made the product feel less confident, which was exactly the point. The product was less confident, and pretending otherwise wasn’t helping anyone.

The third thing is less interesting to talk about, but probably the most impactful in practice: infrastructure.

Outside major cities, mobile data is how people get online. Speeds vary. Data isn’t cheap. Users have spent years developing strong instincts for which apps are “worth it”, and a heavy, slow AI product doesn’t get a long runway to prove itself before it gets deleted.

We had to make architectural decisions I hadn’t really thought about before. What happens to the UX when the connection drops mid-session? How much can we cache without hurting output quality? What does response time look like on a slower connection in a second or third-tier city, not in our office?

None of those questions is in the AI product playbooks I’d been reading. Because those playbooks were mostly written for environments where broadband is assumed and latency is a minor consideration, not a core constraint.

Which meant we were figuring it out as we went. That’s fine. That’s a building. But it’s worth noting: if you’re building AI products for emerging markets, you’re often doing original work, not just implementation.

I don’t think Indonesia is a special case.

I think it’s a preview.

Linguistic diversity, trust environments that don’t match Western AI design assumptions, infrastructure that makes bandwidth a real constraint, this is most of the world. The AI conversation just hasn’t caught up yet because the conversation is still mostly shaped by people building in San Francisco and London.

That’s going to change. The question is whether the people building for these markets are taken seriously as contributors to how AI development is thought about or just treated as edge cases to be handled later.

Based on what I’ve seen, there’s genuinely important and underexplored work happening in these contexts. The problems are real, the constraints are real, and the solutions aren’t going to come from just porting over what worked somewhere else.

I don’t have clean answers to most of this. I’m still figuring out how you evaluate model performance across linguistic contexts that don’t have standardized benchmarks, how you design for trust calibration in markets with different AI literacy patterns, and how you build data pipelines that reflect real-world language without compromising privacy.

These are the questions I keep coming back to, probably for a while.

If you’re working on similar things — building, researching, thinking about this — I’d like to hear what you’re seeing.