Harsh Verma is the Principal Software Engineer – AI/ML at Palo Alto Networks, an IEEE Senior Member.
Since the first model was deployed, the AI industry has been obsessed with capability. Bigger, faster models, larger context windows, more tools and impressive demonstrations have made headlines repeatedly. This logic was that if it could do more, it was worth it. That assumption is now being challenged.
A different conversation is gaining ground within large organizations. Companies now realize that when there are possible financial, legal, operational or reputational consequences, an AI system that is correct 90% of the time may still be unacceptable.
The question has shifted. It is no longer, “Can AI do impressive things?” It is, “Can we actually rely on it?”
These enterprises are tired of AI making things up (hallucinations). While AI is great to test, businesses have learned that its unpredictable mistakes cost way too much money when they actually try to use it for real, day-to-day work.
My perspective on this is that the next premium tier of AI will not be the most creative one, but the one that is most reliable and predictable.
Why Hallucinations Are A Business Problem, Not Just A Technical Problem
Hallucinations, which are when the system says something wrong with confidence, have been viewed as a technical problem that can be fixed later. But today, companies are tired of it, and it is now a business risk.
According to Deloitte Global’s 2025 Predictions Report: Generative AI, organizations looking to expand generative AI beyond early pilots will need to address challenges around trust, capabilities and responsible deployment.
They are now prioritizing risk management and output quality over any experimental capabilities. These are more difficult for industries with higher stakes, like healthcare, financial services, insurance, public sector operations and critical infrastructure.
In these areas, one wrong answer from an AI system can cause actual harm to patients or clients or cause compliance issues. This has caused a shift in what organizations actually want from AI now. They now want consistency, repeatability, auditability, explainability and governance. They want a system that behaves the same way every time, one that they can predict.
The Rise Of Deterministic AI Architectures
Large organizations and users are tired of hallucinations, but the response to hallucination fatigue is not full abandonment. It is redesigning how we build these systems in the first place. Instead of allowing AI systems to operate with few constraints, organizations are increasingly adding layers of structure around how models access information, execute tasks and make decisions.
These can include retrieval-augmented generation (RAG), orchestrated workflows, access controls, evaluation mechanisms and human approval checkpoints. The goal is to make AI systems more predictable, auditable and grounded in relevant information.
RAG is one example of this approach. IBM and Google Cloud describe RAG as a way to connect AI models to external knowledge sources, allowing them to retrieve relevant information rather than relying solely on what the model learned during training. This can improve factual grounding and reduce hallucinations, although the quality of the result still depends on the relevance and quality of the retrieved information.
My perspective on this is that the market is moving from model-first thinking to system-first thinking. Reliability is becoming an architectural challenge rather than a model challenge.
Controlled AI In Healthcare
One of the clearest examples of predictable AI architecture comes from the Mayo Clinic’s AI-ECG research. Rather than using a general-purpose model, researchers developed a convolutional neural network trained to analyze ECG data for specific clinical applications. The approach demonstrates how AI can be designed around a defined dataset, task and clinical use case
Organizations are increasingly choosing reliable and traceable systems over unrestricted results. In controlled environments, predictability is more valuable than creativity.