The hidden economics of autonomous AI agents

Many startup founders often make a costly mistake when first trying to understand the expenses associated with artificial intelligence (AI). They focus on the pricing details provided by the model providers, comparing input and output tokens and forecasting usage as if AI were a simple utility meter. However, this approach fails when companies transition to autonomous agents – software systems capable of planning, retrieving information, running commands, and troubleshooting without constant human intervention.

In this context, the model call is often not the most expensive part; the real cost lies elsewhere. Erik Perttu, Head of Engineering at Edu2Review, explains that while generating code with AI may be inexpensive, the trust-building and ensuring the reliability of the code represents the actual financial burden for engineering teams in Southeast Asia and India.

As AI coding assistants and agentic workflows become integral parts of daily development, these teams face hidden costs beyond merely asking models to write code. These costs arise from various activities necessary to make the generated code usable, safe, and reliable, including automated context retrieval, prompt construction, syntax parsing, multi-pass test validation, security scanning, and human review.

These tasks account for nearly 90% of the total financial and computational expenditure. Autonomous agents differ significantly from single-turn assistants. While the former may perform multiple complex tasks like inspecting local files, executing terminal commands, and re-prompting itself based on errors, the latter might simply suggest a function.

As a result, each loop in the autonomous agent workflow consumes significant resources. Without proper controls, these systems can quickly consume budgets in unexpected ways. Engineers may upload entire code repositories, database schemas, or raw application logs into high-context models when only a small portion of the data is required.

Agents might retry a failing unit test multiple times before human intervention, routing routine data aggregation to deterministic scripts, and reserving high-reasoning large language model calls for anomaly interpretation and executive synthesis. To address these issues, dataxet, an Indonesian media-intelligence firm, implemented a systematic approach to cost control.

The company narrowed the context fed into agents by using tightly scoped and pre-filtered data payloads instead of raw logs. It also routed routine data aggregation to predictable software and reserved high-reasoning large language model calls for anomaly interpretation and executive synthesis. Additionally, dataxet tracked token consumption by feature, pipeline, and engineering workflow, making AI usage more visible rather than abstract.

This visibility is crucial for managing AI costs effectively. The Agoda report highlights the difference in how senior and junior technologists view AI adoption. While senior technology leaders are more likely to view cost as the primary barrier (32%), junior developers are more cautious about potential risks. Thus, the lesson for Southeast Asian startups is clear: embracing autonomous agents without proper controls or imposing strict restrictions on their usage both prove unsustainable.