At Deluxe, a payments and data company based in Minneapolis, no line of code reaches an AI model directly. Instead, a gateway routes each request to one of several models—GPT-5.6, Claude Opus, Claude Sonnet, or others still under evaluation—depending on which of the platform's 50-plus AI agents is handling the work. The developer doesn't choose which model runs the task, and neither does IT in the traditional sense. Most chief information officers have already committed to supporting more than one AI model, but the question of who decides which model handles what task, and how those decisions adapt as models and economics shift, remains unsettled.
Organizations that manage this well assign distinct responsibilities for two decisions that often get conflated. A central platform team owns the routing layer, setting which models are approved, how requests move between them, and who pays for it, according to Sumeet Mahajan, a partner of AI and data at Grant Thornton. The business unit running the workflow makes the task-level call, using criteria the platform team establishes. At Deluxe, an AI governance council determines whether a use case is allowed based on security, legal, compliance, data, and responsible AI rules. Once the use case passes that threshold, the leader responsible for the business outcome selects the model and owns the results. At Merchants Fleet, a New Hampshire-based fleet management and leasing company, the Artificial Intelligence Readiness Council sets guardrails for AI use, then business leaders make task-level calls within them. Only one in five organizations has a tested incident playbook for a model failure, according to Grant Thornton's most recent AI Impact Survey of 950 senior IT leaders.
When companies choose which model should handle a task, cost is always part of the discussion but rarely the deciding factor. Five elements govern model choice at Deluxe: quality, risk, latency, economics, and operability, with the weighting changing by workload. "A cheaper model that generates more retries, exceptions or human intervention is often the more expensive model," said Yogaraj Jayaprakasam, chief technology and digital officer at Deluxe. More sophisticated organizations treat model selection as fit for purpose rather than best in show, according to Michael Adler, director of AI governance and data protection at the law firm Akerman. Before it's considered for a task, a model should clear baseline requirements in confidentiality, data retention, and contract terms—fail one, and the model should be ruled out regardless of price.
The biggest risk in multimodel workflows isn't any single model but the gaps between them. Each AI vendor logs its own calls in its own format—one might capture the full prompt and response, another keeps only metadata, and a third system handles tool calls entirely on its own clock, resulting in an abundance of logs with no single record of what happened. This becomes critical during incident response, yet most plans were built to handle a single model failure rather than multiple models failing together across a handoff. The solution is an independent record that sits above vendor logs and follows a request across every move it makes. Quality, drift, and economics are watched continuously at Deluxe, and a new model release triggers evaluation. A scheduled review happens more often for higher-stakes deployments, and an event-driven review occurs when a model version changes, an incident happens, or a credible new challenger clears a predefined bar. "The model can change," Jayaprakasam said, "but accountability for the workflow and its outcome cannot." That kind of ongoing evaluation represents new work—when a company runs one model, the vendor handles integration, but when several models operate in production, the enterprise takes ownership of model evaluation, agent design, workflow orchestration, and cost management. The shift from single-vendor dependency to internal orchestration creates a permanent governance challenge that most enterprises are still learning to staff and fund.