Where Frontier Models Earn Their Cost

A frontier model is the most capable, most expensive tier of commercial AI available at any given moment, and for most enterprise processes it is a negative return on investment. MIT's Project NANDA found that despite $30 to 40 billion in enterprise generative AI spending, 95% of pilots produced no measurable P&L impact, and the researchers were explicit that model quality was rarely the cause. The failures came from integration, from tools that never adapted to the organization's processes, and from budgets concentrated in the wrong functions. Gartner projects that over 40% of agentic AI projects -- deployments where models plan and execute multi-step tasks on their own -- will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

The instinct to reach for a better model is almost always wrong. Understanding why requires holding two dynamics together: the capability of the affordable models, and the depth of the tasks those models are asked to perform.

The Affordable Models Already Cover Your Problem

The tasks enterprises assign to AI sit comfortably within the range of the affordable models. Think about the problems your AI initiative faced four months ago and ask yourself honestly: was the bottleneck model quality? A model with twice the reasoning depth cannot resolve data that was never production-ready, processes that resist automation, governance that was never designed, and unit economics that never penciled out. Swapping out the current model for the frontier model makes the same stalled pipeline stall at three times the cost.

The system you built was designed around the model you built it with. The content preparation, process pipelines, prompt structures, validation gates, and the human-in-the-loop review all encode assumptions about what that specific model could and could not do. Swapping in a more capable model without reevaluating the pipeline means paying frontier prices for a system that was engineered for a different engine. Think of the model as a car's engine and the fixed code that prepares its inputs and enacts its outputs as the car's body. If the solution is a Honda Civic, putting a twin-turbo V12 in it produces novelty rather than performance.

For most organizations the conclusion is uncomfortable and freeing: forgo frontier model adoption for production processes. Route routine, high-frequency tasks to smaller and domain-specific models with deeper context, which Gartner notes outperform generic frontier deployments at a fraction of the cost when aligned to a specific process. This matches what MIT found about where returns concentrate: back-office automation and narrow processes with measurable baselines.

The Test for the Exception

Frontier spend is justified only in a narrow class of processes. Before targeting a process for a frontier model, it must pass four conditions:

• The value per output makes model cost a rounding error. A few dollars of inference against a decision worth thousands, or a migration worth a labor-year, clears the bar. A few dollars of inference against a routed document or a tagged record never will.

• The process is capability-bounded. The current limit must be what the best model can reason through, with the surrounding data, integration, and governance already solved. If any of those remain unsolved, the process is process-bounded: more model buys nothing, because the constraint sits upstream of it.

• The output is verifiable. Migrated code runs against a test suite. A flagged compliance gap is checked against the regulation. Verification converts model output into bankable value; without it, a more capable model just produces more confident output you still cannot trust.

• The next model release improves the process without a rebuild. The frontier model must sit at a swappable tier in the architecture, so that when a better engine ships in two months, the gain lands immediately. If capturing the gain requires reengineering the pipeline, the process fails the test.

The fourth condition carries a counterintuitive corollary. If the best available model would fully solve the problem, the problem was not deep enough to justify the frontier model spend, and the obvious question follows: what do you do when the next model ships? The right targets are processes where the current state is unsatisfactory, where the frontier model still cannot handle the full depth, and where each percentage point of improvement in quality or efficiency registers as direct value. In those processes, every model release compounds the return instead of stranding the investment.

The Processes that Pass

Legacy system modernization. Banking, insurance, and manufacturing run on decades of COBOL, PL/1, and undocumented mainframe logic, with the engineers who wrote it retiring. Morgan Stanley's internal DevGen.AI tool interpreted nine million lines of legacy code and saved its development team an estimated 280,000 hours. This use case passes every condition: the value per converted module dwarfs inference cost, the constraint is genuinely the model's ability to extract business logic from dense legacy code, generated code verifies against test suites, and each model generation translates more of the codebase correctly without changing the surrounding scaffolding. The economics also differ in kind from transaction processing: a migration is a one-time transformation whose cost amortizes across the asset's remaining life, while per-document inference must clear its unit cost on every record forever.

The supervisor tier in multi-agent systems. The production pattern that survives cost review puts cheap, fast models on the routine steps and reserves the frontier model for the decisions those models cannot make: decomposing an ambiguous request into a plan, reconciling conflicting outputs from worker agents, and handling the escalations. Gartner's guidance is direct on this point: frontier inference should be gated and reserved for high-margin, complex reasoning, while routing everything else downward. The frontier model becomes the engineering manager of a large staff of inexpensive specialists. The design holds because the expensive tier touches a small fraction of total volume, and because swapping in a stronger supervisor improves every downstream decision without touching the workers.

High-stakes reasoning across heterogeneous sources. The escalated underwriting referral that crosses policy language, loss history, and reinsurance treaties. The regulatory change whose blast radius spans thousands of customer communication templates. The contract portfolio reviewed against a new compliance regime. These are the cases where smaller models demonstrably break down: open-ended, multi-step reasoning across documents that were never designed to be read together, where a missed dependency carries regulatory or financial consequence. Here the deterministic pipeline still does the orchestration, the chunking, and the audit trail. The frontier model is invoked at the single point where depth of reasoning is the product, and the verdict it renders is reviewed against a defined standard.

What to Ask Before Moving Forward

The question for decision makers is whether any process in the portfolio passes the four conditions, and the answer at most enterprises is that none currently do, because the data and process foundations have not been built. That answer is the diagnosis, and it points to the sequence that produces returns. The organizations reporting AI-driven P&L gains are the ones that stopped chasing engine upgrades and built the body first: the pipelines, the verification gates, the governance, and the unit-cost instrumentation that make any model, cheap or frontier, accountable for the value it produces. Once that foundation exists, the frontier model takes its proper place: an expensive specialist, deployed at the few points in the enterprise where its depth is the deliverable, swapped upward every quarter, and paying for itself each time.

Sources: MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025"; Gartner, "Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (June 2025) and inference cost commoditization research (March 2026); Epoch AI, LLM inference price trends; Xiao et al., "Densing Law of LLMs" (arXiv:2412.04315); Morgan Stanley DevGen.AI reporting.