Snowflake adds automatic AI model routing to cut enterprise costs

Snowflake has introduced features that let customers pick the most suitable AI model for each job automatically, weighing quality, speed and user preferences while giving tighter control over spending, according to TechRadar. The company says the tools arrive as budgets tighten and businesses look for a better return on their AI outlay.

Grace Adamson, AI Product Marketing Manager at Snowflake, told TechRadar that the first wave of generative AI centred on raw model power, with firms experimenting to reach the strongest outputs or crack the hardest problems. She said organizations have since moved toward producing concrete business results, which means asking how much value each AI interaction generates.

According to Adamson, plenty of work does not need the priciest or most capable system. Certain jobs call for deep reasoning, she said, while routine sorting, condensing or data extraction can be done well by leaner options, including newer open source models, for far less money.

Snowflake describes this shift as intelligence efficiency: how well a company converts models, data and computing power into results it can measure. The company points to dynamic model routing, built-in governance and AI placed next to an organization's own trusted data so that requests carry business context. Together, Snowflake says, these let workloads run on the most fitting model while costs stay in check.

Adamson identified an abundance of models, rather than a shortage, as a major headache for large firms. Fresh options arrive every week with differing strengths, prices and performance, and assigning the right one to each task becomes an operational strain.

Dynamic routing inside Cortex AI Gateway tackles this by sending each workload to a model judged on a mix of output quality, responsiveness and expense for that specific use case. A straightforward job might go to an efficient open source system, while a demanding request needing deeper reasoning could be handed to a frontier model.

Snowflake says governance stays in place throughout. The same policies and controls, masking included, follow any model handling a request, whether open or proprietary, so a cheaper choice does not weaken security. Customers also decide which models and providers their users can reach, which Snowflake says matters for global firms facing regional availability rules or regulated sectors with compliance demands.

For end users, Snowflake argues, the choice should be invisible, while builders and administrators need straightforward setup and management. Adamson said model selection will always count at a governance level but should not fall to every employee.