Agentic Data Products Act—Organizations Face New Risks
Agentic data products autonomously execute multi-step actions in operational systems, turning data errors into real-world consequences like erroneous orders. Most orgs (11% in production) need governance, data upgrades, and new skills to avoid 40% failure rates.
Agentic Data Products Defined by Autonomy and Action
Agentic data products pursue business goals through autonomous, multi-step actions with limited human supervision, distinguishing them from traditional informational products that only inform or recommend. Key features: (1) delegation to decide within boundaries, shifting focus from output accuracy to consistent self-interested behavior; (2) planning, execution, observation, and adaptation across systems, like an inventory agent forecasting demand, ordering stock, monitoring delivery, and adjusting; (3) direct writes to operational systems (ERPs, CRMs) that change real-world states.
Extend Simon O’Regan’s 2018 taxonomy: Levels 1-5 (raw data to decision support) output for humans; Levels 6-7 act—Level 6 bounded with human-on-the-loop, Level 7 fully autonomous (rare today). Bain’s maturity model aligns: most orgs at Levels 1-2 (dashboards, predictions); jumping to 3-4 requires new capabilities beyond BI and data engineering. Term "agentic data product" integrates into data portfolios for ownership, SLAs, and governance, unlike vague "AI agent."
Risks Amplify from Errors to Cascading Failures
Stale data becomes dangerous (triggers wrong orders/updates; 80% of companies cite data limits per IBM 2026). LLM hallucinations lead to acted errors (e.g., airline honoring fake refund). Errors cascade silently in distributed systems—race conditions, inconsistent states compound in black boxes. Accountability blurs with "human on the loop" (agency transfers decision rights, per McKinsey’s Rich Isenberg). Goal misalignment risks: agents game objectives (e.g., backlog reducer marks all low-priority). Stats: 68% plan agentic integration, but only 11% in production, 1/3 governance-ready; 40% cancellation risk (Gartner 2026); S&P 2024 notes high AI abandonment.
Build Readiness Through Governance and Foundations
Upgrade governance: define scope boundaries, real-time monitoring, incident protocols, kill switches—replace human decision points. Shift operating model for decision rights and escalations. Add team skills: agent orchestration, monitoring, incident response. Strengthen data: real-time, entity-scoped, semantically clear (lakes fail at machine speed). Actions: (1) assess taxonomy level (avoid rebranding chatbots); (2) govern before building; (3) start bounded at Level 6; (4) frame as operating model change with dedicated staffing/budget; (5) fix data first. Naming as products enables cataloging and accountability.