It's looking like 2024 is the year of ROI for generative AI

### Overview On an episode of TechTarget's Targeting AI podcast, hosts Esther and Sean speak with Kashyap Kompella, CEO and founder of RPA2AI Research, analyst at Real Story Group, and visiting professor at BITS School of Management. The discussion centers on the shift from 2023's rapid generative AI hype to 2024's focus on enterprise return on investment (ROI), operational governance, risk management, and international regulation.

### Key Topics - **Shift to ROI and Business Value:** Following the initial wave of technological enthusiasm in 2023, 2024 demands proven business value. Currently, infrastructure providers like Nvidia and select consulting firms capture most direct revenues, while adopting enterprises grapple with demonstrating measurable financial returns. - **Production Bottlenecks and Shadow AI:** Historical Gartner findings cited indicate less than 50% of enterprise AI projects reach production, largely driven by data silos, talent deficits, and slow enterprise adoption cycles. Meanwhile, Salesforce survey data cited shows widespread "bring your own AI" practices: approximately 28% of employees use generative AI at work, with roughly half doing so without corporate authorization, risking data leakage and IP exposure. - **AI Governance, Hallucinations, and Safety:** Enterprise deployments are constrained by AI hallucinations and unpredictable failure modes. Approaches to mitigation include Anthropic's Constitutional AI (shifting from human feedback reinforcement learning to AI-assisted feedback for scalability and rule alignment) and rule-based system guardrails. - **Explainability as an Adoption Requirement:** High-stakes domains—such as healthcare, finance, and human resources—require transparent logical models. Kompella notes that lack of explainability often stalls complex deep learning models, prompting organizations to fall back on simpler decision trees or linear models. Regulations like New York City's mandatory third-party AI hiring audits reinforce the necessity of explainability. - **Global AI Regulation and Copyright Battles:** Divergent regulatory approaches are emerging globally. The European Union has enacted the comprehensive EU AI Act; the United States relies on a fragmented hybrid model combining existing agency authorities (FTC, EEOC) and state-level laws; and countries like the UK, India, and Israel lean toward light-touch, existing-law frameworks. Concurrently, major copyright litigation—such as lawsuits involving the New York Times, authors, Getty Images, and past precedents like the legal tech startup Ross—is defining intellectual property rules for training inputs and generated outputs.

### Technical Details - **Architectural History:** Transformers and early generative architectures originated around 2017, progressing through OpenAI models (GPT-1 in 2018, GPT-3 in 2020) before public access via ChatGPT catalyzed widespread adoption. - **Guardrail Architectures:** Discussion contrasts traditional Reinforcement Learning from Human Feedback (RLHF)—which requires tens of thousands of human-graded QA pairs and struggles to scale—with Reinforcement Learning from AI Feedback (RLAIF) and Constitutional AI to enforce corporate values and regulatory restrictions. - **Model Diversification:** The technology landscape is diversifying from monolithic frontier models into multimodal architectures, open-source variants (such as Meta's models runnable locally), small language models optimized for mobile environments, and domain-specific systems.

### Takeaways - Enterprises must transition from unmonitored experimentation to formal corporate usage policies, employee AI literacy programs, and explicit data governance to protect corporate assets. - Software distribution networks give established incumbents (e.g., Microsoft, Google) a distinct competitive edge in enterprise software integration over standalone point solutions. - Commercialization and enterprise production in 2024 will hinge on solving technical explainability, establishing fair data licensing compensation models, and complying with shifting jurisdictional AI regulations.