AI Applications for Business: 12 Categories, and Which Ones Actually Run

Indonesia's government puts national AI adoption at 92 percent (Source: Ministry of Communication and Digital Affairs, 2026). Four and a half months later, Indonesia's MSME minister said more than 90 percent of small businesses are still not using AI in any meaningful way (Source: ANTARA News, 2026). The two figures do not contradict each other. Both are true: almost everyone has touched AI, and almost nobody runs it inside their operations.

Map the spending data against the deployment data and the whole field of AI applications for business collapses into twelve categories. Few of them survive contact with a working company. Four absorb the largest share of the budget and produce savings you can measure, while the rest depend almost entirely on how clean your data is today. This article sorts the twelve into three tiers of maturity, so you can see which one deserves your next project.

Spending figures are blunt about it. Companies worldwide spent US$37 billion on generative AI during 2025, and US$19 billion of that went to the application layer (Source: Menlo Ventures, 2025). Within the US$7.3 billion of departmental spending, coding alone took US$4.0 billion, or 55 percent; IT operations took US$700 million, marketing US$660 million, and HR about US$365 million (Source: Menlo Ventures, 2025). That is a far more concentrated market than any best-apps listicle admits.

Adoption went up, impact stayed flat

Some 88 percent of organisations surveyed used AI in 2025, yet agent deployment sits in single digits across nearly every business function (Source: Stanford HAI, AI Index 2026). A survey of 3,235 business leaders in 24 countries found the same pattern on the results side: 66 percent reported higher productivity and 40 percent reported cost savings, but only 20 percent reported higher revenue (Source: Deloitte, 2026).

Indonesia shows the same gap in sharper form. Here, 97 percent of companies plan to deploy AI agents within 12 months, while only 23 percent are genuinely ready on data and infrastructure (Source: Cisco AI Readiness Index via Kompas Tekno, 2025). Twelve categories follow, ordered along that gap: what runs most often comes first, what stalls in pilot comes last.

Twelve categories of AI application, in three tiers

The first tier holds the categories that already run in production every day, at large corporates and at companies with much smaller teams. They share a profile: the task repeats, the format is predictable, and a person can check the output in seconds.

• Document and data extraction: purchase orders, invoices, delivery notes and bank statements that arrive as PDFs or photos become rows a system can process.
• Customer service and inbox triage: drafting replies and sorting what has to reach a human. In Indonesia the order is WhatsApp first, email later.
• Coding and internal tooling: building the applications themselves, with a measured 26 percent productivity gain in software development (Source: Stanford HAI, AI Index 2026).
• Marketing content production: product descriptions and ad variants at marketplace scale, where output can rise by as much as 50 percent (Source: Stanford HAI, AI Index 2026).

For online sellers and distributors, the first two usually pay off fastest. Seventy-one percent of adult internet users in Indonesia message a business on WhatsApp every day, the highest rate in the world (Source: Kantar for Meta via Kompas Tekno, 2026), so most of your orders, complaints and transfer receipts arrive as loose text and photographs. Olaif Purchase, a B2B ordering portal we built to process customer purchase orders, is shown as a named client with permission. On that build the hard part was never the model. It was agreeing which documents may pass automatically and which must be seen by a person.

Second-tier categories are worth doing, but the outcome is decided by the state of your data before the project even starts.

• Reconciliation and finance workflows: matching payments to invoices, flagging discrepancies, shortening the monthly close.
• Demand and stock forecasting: reorder points and slow-mover detection from the sales history you already hold.
• Internal knowledge base: answering questions about your own SOPs from your own documents. Dull, and it removes a surprising number of interruptions.
• Industry-specific vertical applications: built around one industry's workflow, and the fastest-growing slice of application spending at US$3.5 billion in 2025 (Source: Menlo Ventures, 2025).

Finance deserves more attention than it gets, and the reason is structural. QRIS had 65.77 million users and was accepted by 44.86 million merchants as of June 2026, with 96.68 percent of those merchants being small businesses (Source: Bank Indonesia via ANTARA News, 2026). Transaction data for a great many small companies here is already structured, long before anyone starts an AI project. Order exports from Shopee or Tokopedia do the same job for online sellers. Reconciliation rarely appears as its own budget line in any spending report, even though the work piles up at the end of every month.

Four categories that are still thin at your scale

The third tier is not bad work. Most of it simply does not pay off yet at mid-sized Indonesian scale. IT operations and the internal helpdesk are the second-largest departmental category by spending, at US$700 million or 10 percent (Source: Menlo Ventures, 2025), but that money comes from companies with their own IT teams and thousands of tickets a month. Security and fraud detection produce real results when somebody is there to act on what gets flagged. HR candidate screening is the smallest department named, around US$365 million or 5 percent (Source: Menlo Ventures, 2025), and the one most exposed to Law No. 27 of 2022 on Personal Data Protection, fully in force since 17 October 2024.

Last comes the category everybody talks about: autonomous agents working across functions. In the field, only 16 percent of enterprise deployments actually meet the definition of an agent; the rest are fixed-sequence workflows under a new name (Source: Menlo Ventures, 2025). And roughly one company in five has a mature governance model for autonomous agents (Source: Deloitte, 2026). That does not mean you will never get there. It means that making it your first project is the most expensive way to learn.

What this means for your business

If your company already has its own admin team and you are choosing one project for this quarter, the data points to a practical rule: take the category whose data you already hold and whose output you can check within a day. The cost of reading and typing has collapsed. The cost of deciding has not. A good application moves the reading to the machine and puts what is left on your team's desk as a list of exceptions to approve.

One point rarely makes any of these lists: the same category can be bought as a subscription or built around your own workflow. The difference only shows when your process is non-standard, and warehouse, billing and scheduling processes in Indonesia are rarely standard.

• Pick one high-volume process with a repeating format: incoming purchase orders, billing, or the first reply on WhatsApp.
• Measure the volume over two weeks: documents per day, minutes per document.
• Count how many come out wrong today, because without a starting number you cannot prove the improvement.
• Set an automation threshold and route exceptions to a person, from version one rather than after the first incident.
• Discuss agents after that, when the data and the ground rules already exist.

If you want to see what an AI application would look like for your own process, describe the workflow through our contact page. The Agentiq team will answer with concrete recommendations, not a template.