INTERVIEW: AI Model Fatigue is Now a Trust Problem. Can AI Governance Keep Pace?

In this interview with Impact Newswire, Mary Ann Miller, Fraud & Cybercrime Executive Advisor and VP of Client Experience at New York-based digital identity verification company Prove, examines how rapidly evolving AI models are creating a new enterprise trust problem. Miller discusses the risks of AI model fatigue, the difficulty of continuously validating models and the growing dependence on vendors and complex AI ecosystems. She also explores how the rise of AI agents is changing fraud prevention, why identity and intent are becoming critical to trusted AI interactions, and why organisations need stronger monitoring, traceability and contingency plans as AI-driven attacks become faster and more sophisticated.

The AI industry has moved from a relatively simple question: which model is best? – to a much harder one: which model can an organisation continue to trust when models, vendors, costs, capabilities and risks are changing almost continuously?

That is the emerging problem behind what can be described as AI model fatigue. Enterprises are increasingly testing and deploying multiple models from different providers, while model upgrades can change performance, pricing, latency, safety behaviour and even the way an application responds to the same prompt. For organisations using AI in fraud detection, identity verification, finance, cybersecurity or customer decisions, the challenge is no longer simply validating a model before deployment. It is establishing whether the model and the behaviour around it remain trustworthy after deployment.

A June 2026 study by IBM’s Institute for Business Value, based on 1,000 senior executives, found that 71% said switching their primary AI vendor or model would be difficult. Even more strikingly, 91% said they did not fully understand their AI dependencies across vendors, models and infrastructure. IBM also found that 68% struggled with data-residency and sovereignty requirements across geographies.

That creates a paradox: companies increasingly want flexibility to move between models, but their AI infrastructure may be making them more dependent on particular vendors and model ecosystems.

Traditional software validation assumes a relatively stable product. AI systems are different. Models are frequently updated, fine-tuned, deprecated or replaced. The model underneath an application can therefore change without the surrounding application changing.

This becomes particularly important in fraud and identity systems, where an AI model may make decisions based on behaviour rather than a fixed set of rules.

KPMG’s Q2 2026 AI Pulse survey found that 53% of organisations were using AI agents, while the proportion orchestrating multiple AI agents across workflows had doubled from 9% to 18% in a single quarter. At the same time, only 26% of organisations reported full, real-time visibility into the cost of running AI, despite 66% having monitoring dashboards and 61% having approval processes. 

That gap matters because organisations can know that AI is being deployed without necessarily knowing precisely which model is being used, how it is behaving, what it costs or whether its output is still producing the expected business outcome.

KPMG’s global Q2 survey, covering 2,145 C-suite and business leaders across 20 countries and jurisdictions, similarly found that organisations are moving from AI experimentation toward broader deployment but are increasingly focused on accountability, governance, economics and measurable value. 

The financial stakes are substantial. U.S. leaders surveyed by KPMG planned to invest an average of $202 million in AI over the following 12 months. Yet only 26% had real-time visibility into AI operating costs. 

The problem becomes even more acute as organisations move from chatbots and copilots toward AI agents capable of taking actions.

An AI model generating an answer presents one kind of risk. An AI agent deciding to initiate a transaction, approve an identity, contact a customer, access a database or trigger another software system presents another.

KPMG found that organisations are increasingly coordinating multiple agents across workflows. Forty-eight percent said they were using agents to automate cross-functional workflows, while 49% were using them to support joint decision-making.

That makes intent increasingly important. A conventional fraud system might ask whether a transaction is anomalous. An AI-enabled system increasingly has to ask more complicated questions: Who initiated the action? Was it a human or an agent? What was the agent instructed to do? Is the behaviour consistent with that instruction? Has the system been manipulated? And can the organisation prove why the action occurred?

The distinction is important because malicious behaviour can increasingly be generated through legitimate AI infrastructure. An action may technically originate from an authenticated account while the underlying behaviour is inconsistent with the account holder’s actual intent.

The growing number of models also creates a governance problem. Enterprises are no longer necessarily choosing one AI provider. They may use different models for coding, customer service, document analysis, fraud detection, search, reasoning and agentic workflows.

AI governance has historically focused heavily on model performance: accuracy, precision, recall, hallucination rates and benchmark scores. But enterprise AI is increasingly creating a different governance requirement: accountability for actions.

KPMG found that employee resistance to AI increased to 20% from 5% in the previous quarter, with 53% of that resistance attributed to trust and ethical considerations. At the same time, only 36% of organisations had implemented direct token or usage controls, suggesting that many businesses are still developing basic mechanisms to understand and control how AI is being deployed and consumed.

That disconnect points to a broader challenge: AI adoption is accelerating faster than the systems designed to make it trustworthy. As organisations deploy increasingly autonomous AI agents and rely on multiple models and vendors, questions about who or what is behind an AI-driven action, whether it is acting within its intended purpose and how that behaviour can be verified are becoming increasingly important to fraud and cybersecurity teams.

Mary Ann Miller, Fraud & Cybercrime Executive Advisor and VP of Client Experience at Prove, spoke to Faustine Ngila about the emerging AI trust problem, why “intent” could become a critical concept in fraud prevention, and why continuous validation may replace the traditional “set it and forget it” approach to AI model oversight.

1) IBM’s 2026 study of 1,000 senior executives found that 71% say switching their primary AI vendor or model would be difficult, while 91% say they do not fully understand their dependencies across AI vendors, models and infrastructure. Is “AI model fatigue” now becoming an enterprise risk in its own right, and what does that dependency mean for trust? 

I am not surprised by IBM’s survey feedback. Financial Services has had OCC regulatory requirements for model risk governance since 2011, with significant updates in 2026 to better address modern AI, machine learning and automated vendor technologies. Historically, these requirements were often treated like a “hot potato,” but today agile, continuous and consistent model management is non-negotiable. Model fatigue is one of many risks that requires constant vigilance. Data inputs can become polluted over time, and when they do, organizations need to detect and correct those issues quickly to maintain trust. 

2) KPMG found that 66% of organizations have AI monitoring dashboards and 61% have approval processes, yet only 26% have full, real-time visibility into the cost of running AI at scale. Does this gap suggest companies are adopting AI faster than they can actually monitor and govern it? 

Absolutely. Many organizations are running fast without adequate monitoring. It’s like hearing the starting bell for a marathon but having no way to measure how each runner performs along the course. When everyone crosses the finish line, you can’t accurately evaluate the results of, which is the whole point. AI shouldn’t go live without end-to-end risk management, cybersecurity and advanced monitoring already in place. That also includes agile contingency plans for those situations where AI needs to be paused safely and responsibly. 

3) As companies move from individual models to networks of agents, how much harder does it become to determine which system is responsible for an action? 

It quickly becomes a puzzle. Auditability and traceability need to be built into these solutions as native capabilities, whether agents are operating independently or as part of a larger network. Without that foundation, determining which system made a decision, or why it made it, becomes increasingly cumbersome and time-consuming. 

4) Reco’s 2026 State of Agent Security report found that 80% of AI tools operating in enterprise environments have no IT oversight. What does that tell us about the growing gap between the AI systems companies know they have deployed and the systems actually operating inside their organizations? 

At the end of every AI interaction, people are looking for trust, accountability and safety. That requires agile, consistent oversight from IT, cybersecurity, data science and other technology teams, supported by the right talent and governance. No one understands the importance of that responsibility more than the teams tasked with securing and operating these technologies every day. 

5) As AI agents gain access to both enterprise data and external networks, should companies be validating not just what an AI model produces, but what it is actually authorized to do? 

AI safety isn’t about slowing innovation, but rather making sure we’re building systems that protect human well-being, security and trust. Organizations need to validate not only what an AI system produces, but also what it’s authorized to access and do. In my opinion, this is one area where we simply have to get it right.

6) Google Cloud says the emergence of an “agentic web,” where AI systems can reason, plan and execute transactions, introduces new fraud vectors and requires businesses to distinguish between humans, bots and legitimate AI agents. What do you make of this? 

This is one of my favorite topics at the moment. Not only will intent be critical, but it’s creating an entirely new vocabulary inside organizations. Terms like “intent-related services” and “intent data” are already becoming common across data science, fraud and cybersecurity teams. Equally important is establishing and maintaining identity throughout every step of an agentic workflow. Verifiable identity and intent will be mission-critical for trusted AI agents across virtually every use case you can imagine across the spectrum. 

7) Google Threat Intelligence Group reported on Sept. 8 that in the second quarter of 2026 it observed threat actors compromise a cloud resource and then plan, build and execute an AI-enabled mass credential-harvesting campaign in under six hours. Does the speed of AI-driven attacks mean traditional monitoring and model-validation cycles are simply too slow? 

Yes, AI has given velocity attacks an upgrade of epic proportions. While real time AI scoring has long been a familiar part of fraud prevention, the number of signals that now require “alarm bells” for immediate action has expanded dramatically. Organizations should be conducting proactive fraud and cybersecurity red-team exercises that specifically simulate AI-driven attacks to proactively identify and fill gaps before adversaries can exploit them. 

8) What do you think of the fact that companies are becoming dangerously dependent on AI models they cannot easily switch away from, and what should they be doing now to build resilience? 

We’ve already seen AI use cases that have had to start, stop and restart as organizations adapt to changing conditions. Companies should be updating their cyber incident response plans now to include AI-specific scenarios. Every organization should ask, “What is our fallback mode if we have to pause AI?” They also need to answer a second question: “Would an AI failure create a material business impact?” If the answer is yes, that risk needs to be formally assessed, rated and incorporated into resilience planning moving forward. 

Stay ahead of the Stories shaping our world. Subscribe to Impact Newswire and join our 
WhatsApp Channel for updates on global tech, business, and innovation—all in one place.

Dive deeper into the future with the Cause Effect 4.0 Podcast, where we explore the ideas, trends, and technologies driving the global AI conversation.

Got a story to share? Contact Us to reach a global audience with Impact Newswire.