Kos Tersembunyi Kecerdasan Buatan: Mengapa Seni Bina Penghalaan Lebih Penting daripada Saiz Model✎ Edit

Most people see AI as the engine. From a logistics operations lens, the real cost driver is the routing architecture behind it.

Picture 100 PKS consignments. Each arrives with 100 pages of bank statement manifests. That is not “100 deliveries”. That is 10,000 pages of cargo passing through the hab - transactions, OCR noise, duplicates, internal transfers, bank charges, refunds, cash deposits, platform payouts, loan movements, and vague descriptions that refuse to fit a standard label.

If you run everything through one premium lane, the AI agen has to inspect, classify, and reconcile every single item from scratch. For one heavy 100-page PKS file, a full AI workflow can burn 200,000 to 500,000 tokens per PKS, covering extraction, classification, validation, correction, and report generation. Across 100 PKS, that becomes 20 million to 50 million tokens moving through the same expensive lane.

With a premium model handling the whole flow, the freight bill adds up fast. Take a midpoint of 35 million tokens, split 80% input and 20% output: 28 million input tokens and 7 million output tokens. At intro pricing of $2 input and $10 output per million tokens, that is about $126. At standard pricing of $3 input and $15 output, it is closer to $189.

The second approach is how we run a proper distribution centre. A detached system does the pre-sort first: it extracts the bank statement into structured transaction rows, cleans the data, detects duplicates, separates transfers, applies accounting rules, maps standard descriptions, and validates the output. Only the odd-shaped, damaged, or high-risk parcels get pushed to the AI exception lane.

Because the guardrails already control the workflow, a lower-cost model like Qwen can handle that exception lane. It is no longer asked to “understand 10,000 pages from zero”. It only processes the selected exceptions. If only 5% to 15% of transactions need AI review, jumlah token usage for all 100 PKS may drop to around 3 million to 7 million tokens.

Using a midpoint of 5 million tokens, again split 80% input and 20% output: 4 million input tokens and 1 million output tokens. With a low-cost Qwen-style routed model, the AI inference cost can fall below $1 under some provider pricing, excluding OCR, hosting, storage, engineering, and review costs.

So the real comparison is not Claude versus Qwen. That is like comparing a luxury courier to an economy courier while ignoring the sorting facility. The real comparison is architecture. Claude handling every page directly may cost around $126 to $189 in this example. A detached system using Qwen only for routed exceptions can cut the AI token cost to below $1, depending on provider pricing.

This is why smart routing, segmentation, and guardrails matter on the operations floor. The future of PKS financial statement automation is not “dump 10,000 pages into the biggest engine”. The smarter route is: the system clears the standard lanes, AI clears the exception lane, and humans inspect what is risky.

That is where the operational saving becomes serious.

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