Salesforce's 12 rules for successful agentic AI in enterprises
AI agents fail 90% of the time post-launch—here's how to fix it
Salesforce's John Taschek has distilled critical lessons from over 20,000 agentic AI deployments into 12 vendor-neutral rules for enterprise success. Drawn from engagements with executives, analysts, and AI trailblazers, the framework addresses a critical industry pain point: most AI failures stem not from technology but from architectural and trust issues. A staggering 90% of agentic AI work occurs *after* deployment, focusing on continuous improvement, governance, and workflow redesign—contrasting sharply with traditional software where 90% of work is pre-launch.
The rules tackle systemic barriers like poor data quality (cited by 50% of CDOs as a deployment hurdle), low trust in outputs (a top reason for pilot failures), and the need for semantically consistent data. Taschek’s research highlights a paradox: while over 80% of US government agencies already use AI agents, US desk workers remain the world’s most skeptical (per Salesforce data). Success hinges on shifting from siloed AI to systemic approaches, with clean data, modern cloud stacks, and AI guardrails as non-negotiables.
- 90% of agentic AI work happens post-launch, focusing on management and improvement—not development.
- 50% of CDOs cite data quality/retrieval as key barriers to agentic AI deployment.
- US desk workers are the most AI-skeptical globally (50%+), citing trust, training, and generic outputs as top concerns.
Why It Matters
Enterprise agentic AI fails 90% of the time post-deployment—these rules turn theory into scalable, trustworthy systems.