Some have pointed out that state-of-the-art artificial intelligence models can be strong at solving complex problems but may deliver inefficient results in basic tasks at companies.

The South China Morning Post reported on April 19 that Databricks senior vice president David Meyer (데이비드 메이어) said in a recent interview that large language models are not always suitable for everyday office work.

Meyer cited invoice error identification as a representative case. He said that when a state-of-the-art model is asked to find incorrect numbers, it often “fixes the numbers outright, rather than extracting errors so follow-up corrections can be made”. In corporate systems, it is more important to precisely mark which items are problematic and pass them to the next processing stage than to immediately produce the correct answer, he said. AI can easily skip this step at its discretion, he added.

Such limitations also appear in other technology areas. Meyer said that advanced models such as Anthropic’s Claude are strong at coding but can fall behind models trained and fed with data specialised for data engineering. Data engineering includes transforming large datasets and cleaning up missing values or zero values. “No matter how state-of-the-art an AI model is, it cannot do everything equally well,” Meyer said.

As a result, how companies run enterprise AI is also changing. Meyer said smaller open-source models refined with reinforcement learning can handle specific tasks at much lower training costs. This approach can cut costs by several orders of magnitude compared with state-of-the-art large models, he said.

The strategy is also reflected in Databricks’ own products. Genie, an AI assistant that turns natural language into data queries, uses a structure that combines multiple agents and AI models. An analysis of customer usage patterns showed that companies in practice preferred smaller models over state-of-the-art large models. Small models with fewer parameters are seen as advantageous because they cost less and have shorter latency.

“Small models, by nature, generate the first token and respond much faster,” Meyer said. As services scale to very high queries per second, companies need low-cost models that can handle that volume, he said. That means companies are choosing models based not only on performance itself, but also on processing speed and unit costs in real operating environments.

Still, companies cannot immediately adopt the low-cost, high-performance models they want. Meyer said interest is high in Alibaba Cloud’s Qwen series, and that Chinese open-source models have reached surprising levels in performance, latency and cost. He said it is also a reality that use is limited in current corporate environments due to regulatory and compliance concerns.

Despite these constraints, the pace of corporate AI adoption is accelerating. “Many companies are pushing AI as fast as possible for fear of falling behind,” Meyer said. He added that listed companies weigh more carefully how AI investment spending affects financial statements, while unlisted companies tend to be relatively more aggressive in spending.

Ultimately, competition in the enterprise AI market appears to be shifting from who can release the biggest model first to who can process real work faster and cheaper. The direction suggested by Databricks is also placing more weight on using general-purpose ultra-large models together with task-specialised small models.