Scott Zoldi, Chief Analytics Officer, FICO.
When it comes to extracting value from technology investments, large enterprises have a playbook perfected over decades: Evaluate solutions from multiple providers, choose the most performant and reliable candidate and then invest in and deploy it, with reasonable assurance that the new technology can be operationalized effectively and with trustworthiness.
Generative artificial intelligence (GenAI) large language models (LLMs) such as ChatGPT, Gemini, Claude and many others ripped that playbook to shreds. They fueled the biggest shadow IT deployment in tech history, which has morphed into the largest enterprise tech investment ever; according to Gartner, global AI spending will reach $2.5 trillion this year, up from $1.75 trillion in 2025.
Despite these staggering investment numbers, trust in AI lags. A recent global study by Stanford University’s Human-Centered Artificial Intelligence (HAI) found that 73% of surveyed experts expect AI to have a positive impact on how people do their jobs, compared with just 23% of the public. Additionally, only 31% of U.S. respondents reported trusting their own government to regulate AI, the lowest level among the countries surveyed; the EU is trusted more than the U.S. or China. Finally, although GenAI has been adopted by 53% of consumers worldwide, individual utility differs markedly from that of enterprises that may rely on generative AI to make critical, high-impact and regulated decisions.
All of these facts run counter to the operationalization of AI and GenAI at enterprise scale. Here, companies’ decisions directly affect customers’ lives—and serious errors can devastate the bottom line. For GenAI decisioning systems to deliver sustained return on investment (ROI), they must be effective, trustworthy and fully auditable.
How GenAI Must Level-Up To Meet Enterprise Requirements
The global trust gap in AI presents a pivotal opportunity for enterprises to responsibly address GenAI decisioning in ways that deliver continuous, governed and significant value. This is being accomplished today through bespoke, built-from-scratch versions of domain-specific small language models (DSLMs), which Gartner predicts will comprise half of the GenAI models enterprises will use by 2028.
Bespoke DSLMs can be designed to efficiently perform specific enterprise decision-making tasks and, in alignment with Responsible GenAI principles, can be purpose-built to address a very narrow domain, such as ensuring compliance with banking customer service regulations. These types of models are focused on narrow decision-making tasks and are trained on two distinct, carefully curated datasets: enterprise domain data and specialized task data that is highly specific and carefully audited.
Rigorous data curation significantly improves accuracy and reduces hallucination, criteria that are paramount when operationalizing GenAI technology in enterprise decisioning systems. Furthermore, bespoke DSLMs’ full auditability and domain- and task-specificity are distinct from those of commercially available language models; frontier LLMs offer no insight into or control over their large training corpora, nor do they necessarily allow any auditing of that data.
The New Playbook For Effective, Trustworthy AI
GenAI decisions that are hallucination-free and fully auditable are the key to AI that is trustworthy and can pass regulatory scrutiny. Bespoke DSLMs provide a new playbook for successfully deploying enterprise AI that:
• Mitigates bias: Data that introduces bias is explicitly omitted from training sets, and the auditability of these models enables monitoring to quickly detect model drift.
• Ignites innovation: The trustworthy decisioning that bespoke DSLMs delivers is a framework for innovation, allowing innumerable new, focused GenAI decisioning applications to be developed and deployed quickly, with confidence.
• Delivers sustained business value: Because these specialized models are carefully developed within a Responsible AI framework that allows for transparency, audit and monitoring, they can reliably deliver ROI on organizations’ GenAI investments. More models can enter production faster while greatly reducing wasted spend on abandoned pilot projects.
The Building Blocks Of Agentic AI
Bespoke DSLMs are being deployed in enterprise environments today to address critical decisioning use cases that impact broad swaths of customers—a dramatically different scenario than employees’ individual use of GenAI for personal benefit.
Data control and auditability are the keys to successfully operationalizing trustworthy GenAI at enterprise scale and driving sustainable ROI. Looking forward, the exclusive use of domain data and task training also establishes the validity of bespoke DSLMs for complex, audited agentic work. With these models, enterprises can unlock ROI now and multiply it with myriad more bespoke DSMLs as agentic AI becomes real.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?