Model orchestration keeps growing
Enterprises are treating models as a portfolio rather than a single vendor bet, with reporting noting companies mix different models for different tasks — for example using one provider for synthesis and another for fact‑checking (uctoday.com). The practical consequence is more work on model selection and routing inside products rather than wholesale vendor consolidation (uctoday.com).
Companies are increasingly building products that choose among several artificial intelligence models instead of sending every task to one system. (mckinsey.com) A router in this context works like traffic control: it sends a simple request to a cheaper or faster model and a harder request to a stronger one. A 2025 survey paper on large language model routing said the goal is to improve response quality while minimizing cost. (arxiv.org) That design is showing up as enterprise spending moves from pilots into operating budgets. Andreessen Horowitz said in its June 10, 2025 survey of 100 chief information officers across 15 industries that organizations were “much more sophisticated at mixing and matching multiple models” to balance performance and cost. (a16z.com) McKinsey said on November 5, 2025 that 88 percent of respondents reported regular artificial intelligence use in at least one business function, up from 78 percent a year earlier. The same survey found nearly two-thirds had not yet begun scaling artificial intelligence across the enterprise. (mckinsey.com) That combination has pushed more work into the layer between the user and the model. Teams now test which model handles drafting, coding, search, extraction, or review best, then build software that routes each request accordingly. (idc.com) IDC wrote in late 2025 that by 2028, 70 percent of top artificial-intelligence-driven enterprises will use architectures that dynamically manage routing across diverse models. IDC also said model routing can combine models in sequence, not just pick one winner for each prompt. (idc.com) Researchers are treating that routing step as its own technical problem because model quality, latency, and price all change over time. A Findings of the Association for Computational Linguistics paper from 2025 described routing under budget constraints as a way to avoid running every query through the most expensive model. (aclanthology.org) OpenAI said on December 8, 2025 that weekly messages in ChatGPT Enterprise had increased about eightfold over the prior year, while use of Projects and Custom GPTs had risen 19-fold year to date. Those figures point to more structured workflows, where routing and model choice can be embedded inside repeatable business processes rather than left to individual employees. (openai.com) The result is not a clean consolidation around one vendor. The result is a portfolio approach, with procurement, evaluation, and routing logic becoming part of the product itself. (a16z.com)
