Does agentic AI succeed through process redesign or workflow layering?

Deloitte's Tech Trends 2026 chapter argues that "many agentic AI implementations are failing" because organizations are "layering agents onto old workflows" instead of redesigning operations around them. "True value comes from redesigning operations, not just layering agents onto old workflows," which means "building agent-compatible architectures, implementing robust orchestration frameworks, and developing new management approaches for digital workers." The chapter names three "fundamental infrastructure obstacles": data architectures built around ETL and warehouses that are not positioned for agents that "need to understand business context and make decisions"; governance and control frameworks, since "traditional IT governance models don't account for AI systems that make independent decisions and take actions"; and processes still "designed around human staff" rather than entities that "don't need breaks or weekends" and can "complete a high volume of tasks continually." Deloitte's own 2025 survey found "nearly half of organizations cited searchability of data (48%) and reusability of data (47%) as challenges to their AI automation strategy."

The chapter's reasoning is that inserting an agent into an existing process captures only incremental value, because the process itself was built for human execution. The companies it profiles illustrate the alternative: HPE selected "an end-to-end process where we could truly transform rather than just solve for a single pain point," now running "a dozen agentic proofs of concept" across sales, services, supply chain and engineering, gated by a rule that "we require material ROI signed off by the finance partner and the head of that business unit" — discipline that "kept experiments as experiments." Toyota is extending an agentic tool from visibility into vehicle estimated-time-of-arrival toward resolving supply issues and drafting emails, replacing a process that "used to involve 50 to 100 mainframe screens."

Deloitte's companion survey work found that Why are companies deploying agents faster than governance matures?; this chapter supplies a mechanism behind that gap, pointing to specific data-architecture and governance obstacles rather than treating "immature governance" as a single undifferentiated problem. It also sits in tension with Why do production AI agents stay deliberately simple?, where practitioners keep deployed agents narrow and off-the-shelf rather than redesigning processes end to end — Deloitte's case examples suggest a more ambitious trajectory is available, but only once ROI discipline and architecture investment are in place first.

The excerpt does not measure how many organizations actually attempt end-to-end redesign versus layering, nor does it quantify any success-rate difference between the two approaches; HPE and Toyota are illustrative cases rather than a representative sample, and the 48%/47% figures are self-reported survey challenges, not independently verified architecture audits. The claim that process redesign is the dividing line between agentic AI success and failure remains Deloitte's own diagnosis rather than a tested causal finding.