The hidden cost of getting AI decisions wrong
FICO's Mike Trkay says the AI cost conversation focuses on tokens and models, but the real question isn't what a model costs to run -- it's what a wrong decision costs.
The enterprise AI race has rewarded speed over strategy, and the costs are becoming clear. Competitive pressure drives organizations to focus on the newest, largest and most powerful models.
However, the deeper problem is that many organizations are deploying AI without the accountability infrastructure to govern it responsibly. As the market matures, disciplined architecture and governance drive a sustainable AI advantage, not raw model scale.
The current cost conversation misses the mark
Right now, enterprise AI cost conversations center around tokens. Token spending is the most visible line item on the invoice. It is commoditized, easy to track and a natural path of least resistance for finance and procurement conversations. However, this emphasis is fundamentally misguided.
Organizations fixating on token spend and model selection are leaving the real ROI opportunity -- governance infrastructure and decision accountability largely untouched.
Research from FICO's 2025 State of Responsible AI in Financial Services report underscores the scale of that misallocation: 75% of respondents believe that a unified platform approach, combined with better cross-functional collaboration, could increase AI ROI by more than 50%.
The real cost question is not about what a model costs to run, but what a wrong decision costs in practice. In high-stakes, regulated environments, an incorrect AI output can result in a regulatory breach, financial loss or irreparable reputational damage. In practice, that looks like compliance teams manually reworking wrongly denied applications, audit teams scrambling to reconstruct model documentation when regulators come calling, and penalties that can run into billions annually. Behind those numbers are real outcomes: a borrower denied credit with no clear reason why, such as a claim wrongly flagged as fraud or a patient misdiagnosed -- each one requiring an audit trail to explain the decision.
Accountability for that outcome rests with the enterprise deploying the system, not the vendor providing the foundation model. When an algorithm produces a biased lending decision, regulators look at the institution that authorized the workflow. That reality must change how organizations budget AI. Responsible AI relies on continuous funding for safety and validation infrastructure, sustained well past the initial rollout.
Mike Trkay, CIO & chief customer officer at FICO
Bigger does not equal better
There is enormous institutional pressure to consolidate around a small number of large foundation models. The logic is understandable: simplify the vendor landscape, reduce integration overhead and benefit from the breadth of a model trained on vast data. The problem is that model size is a proxy for breadth rather than capability or trustworthiness. In high-stakes environments, breadth is often a liability.
Large general-purpose language models are designed to handle ambiguous, wide-ranging tasks, synthesize unstructured data at scale, and generate net-new content where the context window needs to be wide and flexible. That architecture is well-suited to those tasks, but not to high-volume, low-latency and highly regulated decisioning.
For high-stakes workflows, domain-specific models trained on curated industry data remove operational and regulatory risks inherent in general-purpose LLMs. In these regulated environments, they consistently outperform large general-purpose models on precision, explainability, and auditability. They also protect what matters most. Securing proprietary data requires ringfencing it. Domain-specific models enable enterprises to do exactly that: train AI to understand specific business logic rather than general patterns extracted from public data.
Choosing between a large foundation model and a smaller, purpose-built one comes down to whether the task demands broad reasoning or high-volume, consistent execution. For processes that involve millions of decisions a day under strict regulatory oversight, a fine-tuned, domain-specific model will consistently outperform a massive foundation model on both accuracy and unit economics. This is where tiered routing, matching the right model to the right task at the workflow level, becomes the real driver of operational efficiency.
The enterprises that will gain lasting competitive advantage are those that build governance, explainability and auditability into their AI foundations -- not those that chase model capabilities while treating accountability as a problem to solve later.
Mike Trkay is the chief customer officer and CIO at FICO, where he leads the global professional services, customer success and customer support organizations, alongside corporate IT and operations for FICO's software-as-a-service delivery. With over 25 years of leadership experience, Mike is a strategic executive focused on driving time-to-value and long-term value realization for B2B software customers. He is recognized for operationalizing large-scale, mission-critical AI platforms and enabling the back-office data and systems necessary for high-velocity business operations.
