Indonesia's AI gap: 54 percent use it, 14 percent work with it
HONG KONG, Nov 16, 2026 - (ACN Newswire) - An owner of a food business in Jakarta runs the whole company on WhatsApp. Orders arrive there. Suppliers send invoices there. Staff photograph receipts and send them to the same group chat. When someone offers to sell her "AI", the pitch is usually a

HONG KONG, Nov 16, 2026 - (ACN Newswire) - An owner of a food business in Jakarta runs the whole company on WhatsApp. Orders arrive there. Suppliers send invoices there. Staff photograph receipts and send them to the same group chat. When someone offers to sell her "AI", the pitch is usually a chatbot that answers customers faster. That pitch misses the job. The job is the work after the chat: matching invoices to orders, chasing payments, filing records.
This gap between talking to AI and working with AI shows up in the national numbers. Indonesia has one of the highest rates of casual AI use in Southeast Asia. It has one of the lowest rates of AI inside business workflows. The distance between those two numbers is where most companies are stuck, and the three ways out of that position cost very different amounts of money and time.
The two numbers measure different people
The e-Conomy SEA 2025 report, published by Google, Temasek and Bain & Company, puts standalone generative AI use among Indonesia's working population at 54 percent. The six-country regional average is 58 percent. Thailand sits at 63 percent, Vietnam at 65 percent, Malaysia at 68 percent and Singapore at 70 percent. Indonesia's internet penetration is 81 percent, so the base is there.
The consumer-side numbers run even hotter. The same report says 79 percent of Indonesian users have taught themselves to use AI tools. Veronica Utami, Google's country director in Indonesia, said at the report launch that 80 percent of users in the country interact with AI daily. She gave one detail that shows the scale: Indonesians generate 18 million images per day on the Nano Banana image feature alone. Revenue from AI-enabled apps in Indonesia grew 127 percent year on year, the fastest in the region. Consumer interest in AI runs about three times the global average, by her account.
The enterprise-side numbers tell the other half of the story. Microsoft's own adoption tracking puts AI use inside Indonesian organizations at 14.1 percent in one measurement period and 14.8 percent in the next. Compare that with the 54 percent individual figure and the ratio is about 3.65 to 1. People play with AI at home. Companies have not rebuilt their work around it.
The pattern holds across tools. Claude's usage tracking shows 23 percent penetration in Indonesian computer and math occupations and 25 percent in education. Lazada reports 31 percent of its Indonesian sellers using its AI sales assistant. These are real numbers, and they are all in the 20 to 30 percent band, far above the 14 percent organization figure and far below the 54 percent consumer figure.
The regional spread shows Indonesia is not an outlier in the casual-use number. It is an outlier in the size of the gap.
Market | Working-population GenAI use | Internet penetration |
Singapore | 70 percent | Above 90 percent |
Malaysia | 68 percent | Above 90 percent |
Vietnam | 65 percent | 87 percent |
Thailand | 63 percent | Above 90 percent |
Indonesia | 54 percent | 81 percent |
Philippines | 53 percent | 70 percent |
Source: Google, Temasek and Bain e-Conomy SEA 2025 🔗, as reported October 2026.
Where AI actually touched Indonesian commerce in 2025 was not the office. It was the selling floor. Video commerce, the format where sellers stream and sell inside chat-adjacent apps, grew its active seller count 75 percent year on year to 800,000 sellers. Transaction volume rose 90 percent to 2.6 billion transactions. The overall digital economy is closing on 100 billion dollars in gross merchandise value, with e-commerce at 71 billion dollars. Those transactions increasingly ride on AI-adjacent rails: recommendation feeds, AI-dubbed product videos, automated customer replies. The consumer adoption number and the commerce number describe the same people.
The surveys behind these numbers measure different populations, so the exact gap between 54 and 14 percent is an estimate rather than a measurement. Estimates vary on where casual use ends and operational use begins. The direction does not. Casual use runs far ahead of operational use, and the gap is wider in Indonesia than in most of the region because the consumer base grew so fast.
Why the obvious answer stalls
A company that wants to close this gap usually starts the same way. Someone buys subscriptions to a global AI tool, hands them out, and waits for productivity. In Indonesia this stalls for three reasons that stack on top of each other.
The first is language. Global models speak English first. Bahasa Indonesia works, but the languages of daily commerce do not. A shop owner in Surabaya who answers customers in Javanese or Sundanese cannot hand those conversations to a model that fumbles both. The e-Conomy data shows consumer AI activity concentrated in exactly these local-language interactions, which is why adoption exploded among individuals and crawled inside firms.
The second is money in the wrong currency. Global AI subscriptions bill in US dollars, typically 20 to 30 dollars per user per month. For a 50-person company that is 12,000 to 18,000 dollars a year before any integration work. That is not a large number by global standards. Against Indonesian SME margins it competes directly with payroll.
The third reason is legal, and it arrived on a schedule. Indonesia's Personal Data Protection Law, UU PDP 27/2022, took full effect in October 2024. Its implementing regulation GR 33/2026 requires controllers to keep a register of processing activities, run impact assessments before higher-risk processing, and honor the right to object to automated decisions. Cross-border transfers of personal data need a legal basis, and remote access from outside Indonesia counts. The Otoritas Jasa Keuangan has issued AI governance guidance for banks, and the regulator treats customer data inside AI prompts as personal data like any other.
Enforcement is no longer theoretical. In November 2026, the coffee chain Fore Kopi reported a leak of 5.7 million customer accounts to Komdigi, the communications ministry. The company filed the report itself and apologized publicly. Whatever the investigation finds, the case lands in a system that now has a working complaint channel and a 72-hour breach notification rule for controllers.
None of this makes global AI tools illegal. It makes them a compliance project. Sending customer chat histories through a model hosted in another country now needs a documented lawful basis, a transfer mechanism and an impact assessment. For a bank, the review alone can take six months. For a two-person trading company, it takes either a lawyer they cannot afford or a shortcut they should not take.
Path one: global subscriptions, kept in their lane
The cheapest path is also the most common. A company buys seats on a global tool and uses it where the compliance surface is thin: internal English-language work, drafting, summarizing, coding, analyzing its own non-personal data.
The cost is known and small. A team seat runs 25 to 30 dollars per user per month at current list prices. Deployment takes an afternoon. The tools are the best in the world at general reasoning and English output. For a company whose staff already read and write English at work, this path delivers value in week one.
The lane restrictions are the whole point. Customer data stays out. Personal data stays out. Anything the PDP Law would call processing of personal data stays out, because the models sit on servers in other countries and no company wants to build a cross-border transfer file for a chatbot subscription. One Jakarta bank described in a 2026 industry guide kept GPT-4o for internal English tooling while moving its customer-facing Bahasa chatbot to local infrastructure. That split is the pattern the market found: global tools for the back office, local systems for the front line.
The weakness of this path is that it barely touches the 14 percent number. Internal drafting makes individuals faster. It does not automate invoices, reconcile orders or answer Javanese customers. Companies that stop here stay stuck between the two adoption curves, paying a new subscription and reporting no change in output.
Path two: the sovereign hosted route
The second path is new in 2026. Sahabat-AI is the open-source model family built by GoTo and Indosat Ooredoo Hutchison, now at 70 billion parameters, running on Lintasarta's GPU Merdeka infrastructure inside Indonesia. The project had a public origin worth recalling: NVIDIA's chief executive introduced it at Indonesia AI Day in 2024, the model family reached 70 billion parameters by mid-2025, and the consumer platform arrived in early 2026. The public chat service is free, on the web and inside the GoPay app. The government put its weight behind the launch: Komdigi and Indosat presented the platform together in Jakarta on February 25, 2026.
Minister Meutya Hafid framed the goal plainly: technology must reach everyone, without language or cultural barriers. Patrick Walujo, GoTo's chief executive, called the 70B release a statement of support for local technology. Vikram Sinha, Indosat's president director, was more specific about the infrastructure. "Sahabat-AI is a national asset," he said. "GPU Merdeka stores and processes data directly at home."
The model speaks five Indonesian languages as first-class citizens: Bahasa Indonesia, Javanese, Sundanese, Balinese and Batak. Industry reviewers who tested the 70B model against global alternatives rate it the only option that handles Javanese well enough for production use. For a bank or a retailer whose customers write in those languages, that capability has no substitute at any price.
Hosting inside Indonesia solves the compliance problem by construction. Prompts stay on Indonesian soil, processed by an Indonesian operator, under Indonesian law. The managed inference service on GPU Merdeka prices at roughly IDR 7,500 per million tokens, which industry comparisons put at parity with the cheapest tier of global API pricing. For most deployment sizes, compliance comes free with the subscription.
The honest weakness is the ecosystem. The model family has accumulated about 35,000 downloads on Hugging Face, mostly of the smaller 8B and 9B variants. Independent review databases count fewer than 40 production deployments with public reviews. Tooling for fine-tuning, monitoring and evaluation is thin next to what surrounds the global models. A company choosing this path in 2026 buys compliance and language fit, and accepts that more of the integration work is self-service. It cannot yet buy the finished product it wants off a shelf, and whether a specific capability exists cannot be verified from a brochure; it has to be tested on the company's own tasks.
Path three: self-hosting the open weights
The third path takes the same model and moves it in-house. Sahabat-AI ships open weights under permissive licenses, so a company with its own GPUs can run the model entirely on its own hardware. Nothing leaves the building.
The economics flip at scale. Serving 50 million Bahasa tokens per day through a global API costs roughly 8,000 dollars per month. A self-hosted node with eight H100 GPUs costs roughly 12,000 dollars per month all-in, and then each additional token is free at the margin. Below a few hundred million tokens monthly, the API wins. Above it, ownership wins, and the gap widens with volume. A company processing millions of customer conversations can reach the crossover within its first year.
The costs that decide this path are not the GPUs. Self-hosting requires MLOps capability that most Indonesian companies do not have on staff: capacity planning, model updates, evaluation pipelines, incident response. Hiring or training that team is a multi-quarter project. The public-sector and bank deployments that succeeded started 12 to 18 months before their deadlines.
There is a middle version of this path that most SMEs should hear about. Managed inference on GPU Merdeka gives the data residency of self-hosting without the operations team. The token price is close to global API parity. For companies between the subscription tier and the hardware tier, this middle seat is usually the right answer.
The three paths, side by side
The table compresses the trade-offs. Prices are list or estimate levels as of late 2026 and change often.
Path | Monthly cost level | Time to first output | PDP Law position | Best fit |
Global subscriptions | 20-30 USD per user | Days | Transfers need a legal basis; keep personal data out | Internal work, English output, small teams |
Sahabat-AI managed (GPU Merdeka) | Near global API parity, around IDR 7,500 per million tokens | Weeks | Data stays in Indonesia | Customer-facing Bahasa and local-language service |
Self-hosted 70B node | About 12,000 USD for an 8x H100 node, marginal tokens free | 2-4 quarters including team | Fully on-premises | High-volume processors with MLOps staff |
Ref: costs compiled from vendor list prices and industry deployment guides cited in the Sources section.
The work after the chat
Whichever path a company picks, the subscription is the cheap part. The expensive part is the same for all three: connecting the model to the work. This is what the e27 column on Southeast Asian SMEs argued in November 2026. Most small businesses in the region already run on WhatsApp, the author noted, so the first AI purchase should not be a chatbot. It should be the system that takes what the chatbot hears and does the follow-up work: matching payments, updating the order sheet, filing the receipt.
That integration work is where the real money goes, and it is also where vendors' price lists go quiet. An integration project for a mid-sized retailer, connecting an AI service to its order flow and accounting, typically costs more in implementation labor in the first year than in model fees. The labor is billed in rupiah, by local firms, which is one more reason the domestic ecosystem around the models matters as much as the models themselves.
Indonesia's largest AI operator learned the same lesson at national scale. Indosat used AI-driven personalization on its own network and cut user churn by 50 percent over 90 days while lifting average revenue per user more than 6 percent, to 2.62 dollars. The capital expenditure saved through AI-optimized network planning reached about 86.5 million dollars. Asked what he would change, chief executive Vikram Sinha did not point at the technology. He said he would have pushed adoption across the organization harder and earlier. The machines were never the bottleneck. The habits were.
Who should buy which path
The choice reduces to three questions: where does the data live, what languages do the customers speak, and how many tokens per month.
- A company under 50 staff, doing internal work, with English-capable staff: buy the global subscriptions, keep personal data out, spend the savings on an integration project for the two workflows that actually cost time.
- A company with customer-facing service in Bahasa or local languages, handling personal data, with no MLOps team: use the managed sovereign route, and treat the language coverage of Javanese, Sundanese, Balinese or Batak as the feature that justifies the switch.
- A company processing millions of tokens daily, with or willing to build an operations team: self-host the open weights, accept the 12,000 dollar monthly floor, and stop paying per token at scale.
The two adoption numbers will not converge on their own. The 54 percent learned to use AI because it was free and fun. The 14 percent will climb only when the work after the chat gets built, and the three price tags above are what the building actually costs in Indonesia in 2026.
Data basis
Adoption figures come from the e-Conomy SEA 2025 report by Google, Temasek and Bain & Company, released in November 2025, and from Microsoft adoption tracking cited in the accompanying research. Enterprise figures for Claude and Lazada come from the tool operators' published usage data. Sahabat-AI capability, pricing and launch details come from the operators' announcements of February 25, 2026 and industry deployment guides published in 2026. Cost estimates for the three paths use list prices and the industry-standard deployment benchmarks cited in the text; all are estimates and change with pricing updates. The Personal Data Protection Law framework follows UU PDP 27/2022, GR 33/2026 and OJK guidance as described in published legal analyses. The Fore Kopi breach report is from November 2026. Data as of November 2026.
Limitations
The 54 percent and 14 percent figures measure different populations with different survey methods, so the 3.65 ratio is indicative rather than exact. Token and subscription prices shift quarterly, and the self-hosting crossover point depends on utilization that no public dataset discloses. Production deployment counts for Sahabat-AI cannot be verified beyond public reviews. Regulatory interpretation of GR 33/2026 is still settling in practice.
Sources
- BusinessWorld 🔗 — e-Conomy SEA 2025 regional GenAI adoption figures, October 2026
- VOI 🔗 — Google Indonesia launch remarks, 18M daily images, November 2025
- ITBeat 🔗 — Sahabat-AI platform launch with Komdigi, February 2026
- Technologue 🔗 — GoTo and Indosat executive remarks on Sahabat-AI 70B, 2026
- ITNews Asia 🔗 — Indosat AI Grid results and Sinha remarks, 2026
- Software Listing 🔗 — SEA sovereign LLM deployment costs and benchmarks, 2026
- e27 🔗 — SME chatbot workflows column, November 2026
- CNBC Indonesia 🔗 — Fore Kopi data breach report to Komdigi, November 2026
- Lexology 🔗 — Indonesia AI legal framework snapshot, 2026
About Global Infinity Labs
Global Infinity Labs is an AI innovation company specializing in MCP data services. This article was prepared by our expert team using insights and data provided through our MCP platforms.
Contact: support@globalinfinitylabs.com
Copyright 2026 ACN Newswire. All rights reserved. www.acnnewswire.com
Tokyo Overnight
Japan markets, decoded. Get the brief in your inbox.
No spam. Unsubscribe anytime.
