Enterprise AI in Focus: Singapore’s AICC Says Its Routing Solution Slashes 47% of API Costs

Singapore-based AICC, an artificial intelligence platform that gives corporate clients a single connection to hundreds of AI models, claims its system has cut API costs by an average of 47%.

The start-up attributes the reduction to multi-model routing, matching specific workloads to models based on cost and performance instead of directing every task to a single premium system.

When an application connects to a model like OpenAI‘s GPT-4 or Google Gemini, it does not run the AI on its own servers. Instead, it sends data to the provider’s servers through an application programming interface (API), which processes the request and returns the output for a fee.

API costs have become a primary operational headache for companies moving generative AI out of the pilot phase and into daily business. As queries multiply across customer service, coding and document processing, even marginal price differences for individual computing requests stack up to high operating costs.

AICC operates as an aggregator, providing access to more than 300 models such as OpenAI’s GPT models, Google’s Gemini, Anthropic’s Claude, Alibaba’s Qwen, ByteDance’s Doubao, DeepSeek and xAI’s Grok through a unified network connection. The platform’s routing software automates model selection on a task-by-task basis. Routine classifications or basic jobs like information extraction are offloaded to less expensive, lightweight models, while tasks requiring complex reasoning are channelled to frontier systems.

However, the reported 47% saving remains an unverified internal claim. AICC has not named the corporate clients involved in the calculation, nor has it disclosed the sample size, tracking period or baseline spending data used to measure the reduction.

The broader underlying problem of compounding computational expenses is a growing concern for businesses. An EY US survey published in July this year shows that 82% of senior executives investing in AI are concerned about escalating token usage costs.

Among organisations using token-based systems, 98% said spiralling costs had prompted them to reconsider elements of their AI deployment strategy. Only 64% said they were actively monitoring token consumption with clear budgetary guardrails.

Actual enterprise spending has also trailed initial market hype. EY found that 35% of senior leaders surveyed a year ago expected their organisations to be spending at least $10 Mn on AI by this point, compared with just 23% who reported hitting that milestone in the latest data.

Again, 98% reported a positive return on their AI investments, indicating that the core operational challenge has shifted from finding value to managing the structural economics of adoption.

For businesses across Southeast Asia, these calculations are acute. The 2025 e-Conomy SEA report, compiled by Google, Temasek and Bain & Company, identified the high upfront cost of AI technology as a major return-on-investment hurdle in a region where relatively low labour costs can make the economic case for replacing or augmenting existing processes harder to justify.

Fragmented regional regulations further complicate deployment and inflate compliance budgets.

On the other hand, regional AI investments are accelerating. More than $2.3 Bn was invested in Southeast Asia’s 680+ AI start-ups in the past 12 months, with AI accounting for more than 30% of private funding value in the first half of 2025. For enterprises across Singapore, Indonesia, Malaysia, Thailand, Vietnam and the Philippines, this forces a pragmatic financial trade-off. Does every operational task demand a premium model? Or does the incremental capability justify the margin?

Automated multi-model routing aims to solve this dilemma, though it introduces technical trade-offs.

On August 11, US chip designer Nvidia introduced NeMo Switchyard, an open-source model-routing library built to distribute AI workloads based on capability, cost and infrastructure requirements. Nvidia noted that dynamic routing avoids the latency and expense of defaulting to a large model, while keeping premium computing power available for tasks that require it.

Yet a benchmark conducted by AI framework developer LangChain using Nvidia’s Switchyard highlighted the performance risk. Routing requests between Nvidia’s Nemotron 3.5 Lightning and Anthropic’s Claude Opus 4.8 delivered a 74% cost reduction compared to using a frontier model exclusively, with only 7% of calls sent to the frontier model. However, overall accuracy fell by about six percentage points.

This drop underscores the core friction of model routing. The cheapest individual API call can prove costly if the output is flawed. If a cheaper model yields an inadequate result, requiring a human employee to intervene or forcing the system to rerun the query through a premium model can erode initial cost efficiencies.

For AICC, the commercial strategy relies on abstracting this complexity. Rather than requiring engineering teams to maintain distinct integrations with individual model providers, businesses link their applications to a single junction point. This structural flexibility can also reduce dependence on a single vendor. If a provider changes its pricing tier or suffers a performance degradation, corporate traffic can be dynamically rerouted without a total overhaul of the client’s software infrastructure.

While AICC’s unverified 47% benchmark cannot serve as an industry standard, the platform’s positioning reflects a broader pivot in enterprise AI. As deployment volumes grow, corporate buyers are stepping back from the race for absolute intelligence. For enterprises, the next wave of efficiency gains will come not from deploying more powerful AI, but from learning how to buy exactly what they need.