CNBC Finds Enterprises Favor Older AI Models for Routine Work

At Salesforce’s Dreamforce conference, enterprise leaders told CNBC that older, lower-cost AI models are often enough for everyday sales and customer-service work. Their model choices suggest that, for many companies, the immediate challenge is less about accessing the newest capability than controlling the cost of putting AI into routine workflows.

That is a notably different priority from the race to build ever more capable frontier systems. Salesforce customers and partners at the event said they were still deciding their AI budgets and weighing models from Anthropic and OpenAI against cheaper open-source alternatives. G2 Chief Innovation Officer Tim Sanders told CNBC that most results from AI agents come from the prior generation of models, not frontier capability.

A split between routine requests and harder judgment

The emerging approach is selective, rather than a blanket rejection of the newest models. Docusign uses both frontier and open-weight models, according to its CEO Allan Thygesen. It reserves larger frontier systems for judgment-intensive work, including complex clause analysis, reasoning across multiple documents and summarization, where the higher cost can be justified.

The mechanism behind that split is model routing: directing each request to the AI system that best balances the work required with its cost. Thygesen said Docusign uses routing to send requests to the most cost-effective system. Nice described a similar preference, telling CNBC it does not rely on high-end models for most workloads and finds current models, including ones a generation old, effective for its customers’ needs.

Costs reshape the software calculation

The preference is not only about whether a model can complete a task. AI use introduces variable token costs into software services that traditionally had little marginal cost to deliver. Sanders said a move from conventional subscription software to agentic software could pull gross margins from more than 85% to about 45%, though that estimate reflects his assessment rather than a disclosed industry average.

What enterprise teams are optimizing for

• Adequate performance for common sales and customer-service tasks, rather than the highest available capability.
• Lower spending by assigning requests to the least costly model that fits the task.
• Time to assess new releases before connecting them to production systems; Nagarro waits about three months before integrating the latest models.

Salesforce’s own Agentforce example fits this practical posture. CNBC cited a Salesforce support page indicating its bots do not rely on Anthropic’s Claude Fable 5.1 or OpenAI’s GPT-6 Astra. The unresolved question is how often enterprise work will remain in the routine category as agents take on more consequential tasks—and whether the savings from routing will outweigh the appeal of using the strongest model more broadly.