The Hidden Kos of AI: Why Seni Bina Sistem Beats Model Size in Production✎ Edit

Most teams underestimate how messy PKS bank-statement automation gets once it is running in production.

Picture 100 PKS. Each one drops a 100-page bank statement on your ingestion queue. That is not 100 files. It is 10,000 pages of financial data-thousands of transactions, OCR noise, duplicate entries, internal transfers, bank charges, refunds, cash deposits, platform payouts, loan movements, and descriptions that only make sense in context.

If you let the AI agen reason over every page end-to-end, the token burn is brutal. A dense 100-page PKS file can easily push 200,000 to 500,000 tokens once you include extraction, classification, validation, correction, and report generation. Multiply that by 100 PKS and you are already at 20 million to 50 million tokens before anyone reviews the output.

Run that through a premium full-processing model like Claude and the bill adds up fast. Take a midpoint of 35 million tokens, split 80% input and 20% output: that is 28 million input tokens and 7 million output tokens. At Claude Sonnet intro pricing-$2 per million input and $10 per million output-you are looking at about $126. At standard pricing of $3 input and $15 output, it climbs to about $189.

The second approach is architected differently. The AI does not read everything over and over. A detached preprocessing layer parses the statements into structured transaction rows, cleans the data, detects duplicates, separates transfers, applies accounting rules, maps standard descriptions, validates the output, and only escalates unclear or risky transactions to the model.

Because the guardrails already own the workflow, you can route the exceptions to Qwen or another lower-cost model. The model is not asked to understand everything from scratch; it only resolves selected edge cases. If just 5% to 15% of transactions need AI review, the whole batch of 100 PKS might land between 3 million and 7 million tokens.

Midpoint that at 5 million tokens, again with an 80/20 input/output split: 4 million input tokens and 1 million output tokens. With a low-cost Qwen-style model handling only those routed exceptions, the AI inference cost can drop below $1 on some provider price cards. That excludes OCR, hosting, storage, engineering effort, and human review, but the inference gap is the point.

So the comparison is not really Claude versus Qwen. That is a surface-level benchmark. The real comparison is architecture. Claude touching every page directly costs you $126 to $189 in this scenario. A detached, routed system using Qwen only for exceptions can push the AI token cost below $1, depending on the provider.

This is why smart routing, segmentation, and guardrails are non-negotiable in production. The future of PKS financial-statement automation is not "send 10,000 pages to the biggest LLM". The viable model is: the deterministic system handles the structured work, the AI handles uncertainty, and humans close the loop on risk.

That is where the cost saving gets real.

#ArtificialIntelligence #AIAgents #DetachedSystems #SmartRouting #Guardrails #Perakaunan #PKS #FinancialPenyata #TokenKecekapan #Automasi #ESG