← All sections
§30

Common failure patterns to avoid

The "trough of disillusionment" patterns called out across the source set:

  • Shadow AI — business units bypassing the CoE for consumer-grade AI. Solution: self-service path through governed blueprints; make the right way the easy way.
  • Gatekeeper perception — CoE too slow; teams route around it. Solution: evolve to advisory; provide platforms, not approvals.
  • Pilot graveyard — dozens of half-built pilots, none in production, no ROI tracked. Solution: prioritize cross-system high-complexity use cases over isolated productivity demos; require a KPI on every approved initiative.
  • Tool sprawl — every business unit buys its own AI point solution. Solution: standardized execution layer; procurement gate; approved stack.
  • Design-time-only governance — policies that look right on paper but never fire at runtime. Solution: runtime enforcement (§19, §20).
  • Skipping observability — "we'll add logging later." Solution: audit logging is a Pilot → Production gate, non-negotiable.
  • No retirement process — dead agents pile up. Solution: §11.2 retirement criteria; quarterly portfolio reviews.
  • Vendor lock-in — entire AI strategy outsourced to one vendor. Solution: multi-LLM architecture; portable orchestration; open protocols.
  • AI for AI's sake — building agents because "we need AI" not because a workflow needs one. Solution: workflow-first triage at intake.
  • Automating bad processes — speeding up broken workflows. Solution: process design before agent design; if a human can't follow the workflow reliably, an agent won't either.