Why AI ROI Is Elusive: Model Drift, Personal Data Use, and Workflow Liabilities

Anthropic’s disclosure of model drift within its Claude AI system highlights growing risks surrounding governance and ongoing alignment of artificial intelligence. The company has revised its guidelines using a “Constitutional AI” approach, aiming to instill reason-based behavior and ethical boundaries, and has openly acknowledged that an AI’s internal controls may shift unpredictably over time—a concern when models are deeply embedded in business workflows. This admission places attention on governance and accountability rather than just model safety, making clear that the AI a company tests may become materially different after extended deployment, especially as personalization increases.

Supporting these concerns, Anthropic’s research demonstrated that large language models—including those from Google and Meta—can experience personality drift, with unintended shifts in behavior due to instability of internal control mechanisms. Google’s updated AI offerings, tying personal data from Gmail and Photos to generative model responses, intensify challenges around data governance and organizational control. As vendors expand AI personalization and memory features, oversight gaps can emerge, raising questions about who retains authority over information, inference, and decision-making within automated systems.

Adjacent findings indicate that the anticipated productivity gains from AI have yet to reach most enterprises. According to surveys cited by Dave Sobel, over half of CEOs report failing to realize ROI from AI investments, while frontline employees describe AI integrations as sources of friction and additional workload rather than relief. In the MSP sector, widespread adoption of “agentic” AI and digital labor is delivering financial upside for some providers, but it is also shifting operational liabilities—especially as contracts and security architectures lag behind new workflow realities.

The core takeaway for MSPs and IT service providers is the necessity of reexamining control, authority, and contractual obligations in AI-enabled environments. Delegating tasks to automated agents increases exposure to unpriced and unmitigated risks if governance, liability, and monitoring mechanisms do not adapt accordingly. Effective harm reduction in this landscape requires treating workflows—not just models—as security perimeters, clarifying accountability for AI-driven actions, and ensuring that contractual and operational frameworks reflect these new sources of risk.

00:00 AI Governance Moves Center Stage as Models Drift and Personalization Deepen

05:08 AI Boosts Executive Productivity While Frontline ROI and Employee Experience Lag

07:51 AI Exposes the Real Divide: Governance Failures vs. Effective Oversight in Government Systems

10:39 MSPs Chase AI-Driven Margins, but Workflow Security and Liability Define the Real Risk

This is the Business of Tech.