“AI agent” has become one of the most frequently mentioned terms in the pitch decks of Indonesian product teams throughout 2026. Most of the interesting content on X and LinkedIn comes in the form of sixty-second demo videos: agents reading emails, auto-replying, updating spreadsheets, sending notifications to Slack. Everything runs smoothly. Applause.
However, teams that have had agents in production for three to six months have quieter stories. Some admit that they eventually retired their agent after spamming customers twice with the wrong template. Some are still running, but their monthly tokens have tripled from the initial estimate. There are also those who succeeded, but with a footnote: “we use the agent for very specific tasks, not as a generic assistant.”
Before you spend two weeks assembling workflows in n8n and tell your team “we have an AI agent now,” the following questions need to be asked first. Not technical questions. Operational questions.
What Are You Actually Automating: Decisions or Execution?
Execution is easy. Send an email every nine AM. Pull data from the API every Monday. Update the sheet if there’s a new row. The workflow is straightforward, doesn’t need an agent; n8n without LLM is sufficient.
What’s costly is decision-making. Should this email reply use template A or template B? Should this message escalate to a supervisor or not? Is this order approved or does it need manual review? This is where the AI agent truly makes sense. But the decisions transferred to the agent are also decisions you must be ready to audit, monitor, and sometimes retract. If you’re not prepared to do those three things, just automate the execution for now.
What Is Your Tolerance for Incorrect Output?
Karpathy once said that an AI agent is like a smart junior developer but prone to making mistakes. A useful analogy, but it lacks one thing: a junior developer’s mistakes are evident from outputs that don’t compile. An AI agent’s mistakes can manifest as replies that are grammatically correct and sound reasonable, but send invoices to the wrong email.
For low-stakes cases like summarizing a personal inbox, one incorrect summary out of ten is still acceptable. For customer-facing cases like replying to customers via WhatsApp, one incorrect reply out of a thousand can become a viral screenshot. Calculate the cost of wrong output before deciding whether the agent is worth deploying without human review in the middle.
Who Fixes the Agent When It Blunders?
A poorly designed workflow in n8n can run silently for months if no one is monitoring it. No errors, no notifications, no red dashboards. Just: the results gradually become more nonsensical.
A simple question: if this workflow starts to behave strangely, who in the team gets called? Who has access to the logs? Who understands how to read the execution history? If the answer is only “the one who created it,” it means the agent is a human single point of failure.
“We treat agents like junior employees, they need supervision, feedback loops, and someone who actually reads their work.”
David Sacks, All-In Podcast, 2025
Identify who owns the workflow before deployment. Not after the problems arise.
What Is Your Monthly LLM Token Budget?
LLM token prices may seem small in the pricing table. But after the agent has been used for a few weeks, costs can rise sharply due to three factors: the context window expanding with each iteration, automatic retries when there are errors, and more edge cases than initially estimated.
Roughly calculate: using one million tokens per day, with a mid-range model costing about two to five US dollars. In a month? sixty to one hundred fifty dollars. That’s for a small, simple workflow. Once your agent starts calling other tools and engaging in multi-step reasoning, costs can increase three to ten times.
This is not a reason to avoid starting. But it is a reason to set a budget alert from day one. Many Indonesian teams only realize at the end of the month when the invoice has already been charged to the founder's credit card.
The First Day the Agent Runs Is Like the First Day Hiring a Junior
It still needs supervision. It still needs clear rules of engagement. It still needs to be evaluated weekly until you are confident it can be left alone.
What distinguishes teams whose AI agents are sustainable from those that retire quickly is not the sophistication of the prompts or the number of nodes in the workflow. What distinguishes them is realistic expectations at the outset and maintained operational discipline afterward.
References:
- OpenAI – Token usage and cost optimization guide → platform.openai.com/docs
- n8n Community – Observability patterns for AI workflows → community.n8n.io
- All-In Podcast – David Sacks on AI agent governance → allin.com/podcast
- Anthropic – Building reliable agentic workflows → docs.anthropic.com