AI for Business: From Automation to Smarter Ways of Working

Two numbers from the latest global survey deserve to be read side by side. A full 88% of organizations now use AI in at least one business function, up from 78% a year earlier — yet only 7% report having scaled it fully across the organization, and the majority are still stuck at the pilot stage (Source: McKinsey State of AI survey 2025, as reported by Silicon Canals, 2025). Almost everyone is using AI. Almost nobody has actually changed how they work.

In our experience building systems for Indonesian businesses, that gap is rarely about technology. It is about task placement. AI automation is not a single product you buy once; it behaves more like a ladder with several rungs, running from simple rules you could write on one page up to AI judgment that still waits for a human sign-off. The businesses that pull ahead are not the ones with the flashiest technology; they are the ones that place each task on the right rung. That, in practice, is what a smarter way of working means.

The noise around the topic is real. Investors poured more than US$2.3 billion into over 680 AI startups across Southeast Asia in the past twelve months, and 79% of workers in the region say they are learning to use AI (Source: e-Conomy SEA 2025 report by Google, Temasek and Bain & Company). Tools are abundant. What is scarce is a way to decide which task needs which tool.

The four rungs of AI automation

• Rung 1, organized manual work: written SOPs, daily checklists, a tidy spreadsheet recap. Nothing is automated yet, but without a clear process none of the rungs above it can stand.
• Rung 2, rule-based automation: the right choice when inputs are stable and every rule fits on a single page. Think scheduled payment reminders sent through WhatsApp templates, or automatic recaps of marketplace settlement files.
• Rung 3, rules with AI assist: rules still carry most of the flow, and AI is called in only for the messy parts, such as reading a photographed receipt or tidying an order that arrived as free-form chat.
• Rung 4, AI judgment with human approval: AI analyzes, flags anomalies and proposes a decision, and a person approves it. Even on the top rung, you stay in control.

Picture every repetitive job in your business arranged on that ladder. The higher the rung, the messier the input the system can handle — and the more it costs to build and supervise. Payment reconciliation is the ladder in miniature. Settlement files from Shopee, Tokopedia or a QRIS acquirer arrive in a clean, structured format, so matching them against your sales records is second-rung work. A customer transfer with the wrong amount and no reference, on the other hand, is fourth-rung judgment. And this is no longer a big-corporate problem: QRIS processed 6.05 billion transactions worth Rp 579 trillion in the first half of 2025, across 39.3 million merchants, 93.16% of them small and mid-sized businesses (Source: Bank Indonesia, as reported by Kontan, 2025).

Three tests that decide where a task belongs

To place a task, we run three simple tests. First, how stable the input format is: a bank export file is stable, while WhatsApp chat full of abbreviations and photographed handwritten notes clearly is not. Second, whether every rule can be written down on a single page: if it can, stop at the second rung, because rules are cheaper, faster and easier to audit than any AI model. Third, what one wrong decision costs: the more expensive the mistake, the more firmly a human must approve before the system acts.

Typical Indonesian operating conditions make the first test the decisive one surprisingly often. Orders, purchase orders and approvals tend to arrive over WhatsApp as free-form text mixed with photos of handwritten notes. That is exactly where pure rule-based automation breaks down and the AI rungs start earning their keep. Sending a templated payment reminder, by contrast, will never need AI.

What does each rung cost?

Two anchor numbers keep the arithmetic honest. On one side, adding a single admin hire in Jakarta now starts from the 2026 provincial minimum wage of Rp 5,729,876 per month, up 6.17% on the year before, and that is before BPJS and holiday allowances (Source: ANTARA News, 2025). On the other, a budget-class AI model such as Gemini 2.5 Flash-Lite is priced at US$0.10 per million input tokens and US$0.40 per million output tokens (Source: Gemini API pricing documentation, Google, 2026). At those rates, our rough calculation: reading and classifying a one-page document, around 1,000 tokens in and 200 tokens out, costs in the region of Rp 3 per document.

In other words, the model call is no longer the expensive part. The real cost sits in building the system, the guardrails and the review process that make AI output trustworthy. That is why the one-page rule remains the cheapest rung of all: there is no model to supervise. There is a non-technical consideration too. Sending customers' personal data to a third-party AI API counts as data processing with real obligations under Indonesia's Law No. 27/2022 on Personal Data Protection, while rule-based automation running inside your own systems carries far lighter exposure. Keeping some tasks on the lower rungs can be a compliance decision, not a sign of falling behind.

What this means for your business

This ladder is not whiteboard theory. In Finance Autopilot, a finance back-office system we built, all four rungs live inside one product: rule-based parsing handles 9 bank statement formats and 7 card issuers, complete with arithmetic and VAT checks; categorization runs on vendor mapping; the LLM is invoked only where rules cannot reach, such as documents that exist solely as image scans; and every entry still passes human review before it is posted. Built by Agentiq for an anonymised client; all screens shown use fictional sample data.

The takeaway is simple: the biggest barrier is not technology, and it is not money either. APJII's 2025 survey puts AI adoption among Indonesian internet users at just 28.5%, with lack of knowledge as the number-one barrier, cited by 46.56% of respondents (Source: APJII, Indonesia Internet Profile Survey 2025, as reported by RedaksiBaru.id). An owner who holds a task-placement map like this one is already ahead of most of the market, before a single line of code gets written.

Practical steps this week

• Write down the ten repetitive tasks that eat the most of your team's time, then mark which ones have stable inputs and which are messy.
• Place each task on one of the four rungs, and start automating from rung two, not from the top.
• For every task that lands on an AI rung, decide who approves the output and what one wrong decision would cost, before any conversation about tools.

If you would like to see what this AI automation ladder looks like for your own processes, tell us about them through our contact page. The Agentiq team will reply with concrete recommendations on which rung is worth building first, not a template.