Hidden operational costs from AI can pose massive challenge for enterprises

Enterprises across Singapore and the broader Asia-Pacific region are discovering that the hardest part of adopting artificial intelligence isn't choosing the right model — it's managing the costs that quietly accumulate once those models go into production. From token consumption and retrieval inefficiencies to governance overhead and human oversight, the operational expense of running AI at scale can quickly outstrip the cost of the models themselves.

Ed Keisling, Chief AI Officer at Progress Software, told iTnews Asia that many organisations celebrate successful proofs of concept without anticipating the operational demands that emerge when AI agents become more autonomous. At enterprise scale, retrieval inefficiencies, reasoning loops, and infrastructure monitoring compound into significant expenditure — costs that are rarely visible during the pilot phase when datasets are small and governance requirements are minimal.

The challenge is particularly acute for Singapore businesses that are aggressively deploying agentic architectures. As companies across the financial services, logistics, and government sectors here push AI agents into production, they are encountering what Keisling calls the "rework tax" — money spent fixing errors, tuning retrieval pipelines, and retrofitting governance onto systems that were never designed for autonomous decision-making at scale. Agentic Retrieval-Augmented Generation (RAG), which makes retrieval goal-driven and continuously validated rather than static, offers a path forward by grounding AI outputs in approved enterprise knowledge.

The solution, according to Keisling, isn't just better models but better infrastructure. He urges CIOs and CFOs to invest in standardised enterprise knowledge layers that multiple AI assistants and automation workflows can share from a common governance foundation. Without that shared layer, every new AI deployment duplicates engineering effort while adding governance complexity — multiplying operational costs rather than distributing them.

Why it matters for Singapore: Singapore's Smart Nation push and National AI Strategy 2.0 have encouraged rapid enterprise AI adoption, but the operational cost trap threatens to slow that momentum. For financial institutions governed by MAS guidelines and government agencies bound by strict data governance rules, the hidden costs of scaling AI are not just a budget concern — they are a compliance risk. The enterprises that treat retrieval architecture, governance, and observability as strategic investments rather than technical afterthoughts will be the ones that turn Singapore's AI ambitions into sustained business value.