Most Enterprise Agentic Projects Are Doomed, Here's Why — Jess Grogan-Avignon & Jack Wang, Accenture

Jess Grogan-Avignon and Jack Wang from Accenture present a practitioner's account of why most enterprise agentic AI projects fail — not because of technical shortcomings, but because organizational scaffolding built for human-pace operations cannot accommodate machine-speed delivery. Drawing on years of deploying AI at large enterprises in telecoms, utilities, government, and healthcare, they frame their findings around five structural tensions that predict agentic project success before a line of code is written.

The core thesis is illustrated by a real client engagement: the team built a working agentic application in two weeks, but required an additional 12 months to reach production because infrastructure, security, AI gateway, data governance, and application teams all needed to align. As a benchmark for the urgency: GitHub reported one billion commits in 2025 and is on pace for 14 billion by year-end 2026, at a rate of 275 million commits per week. The approval and deployment infrastructure, however, was never redesigned for that throughput.

The five tensions are: First, speed — every human approval chain must become executable, auditable code rather than a signoff meeting or a change-freeze queue. AI can accelerate the construction of this governance automation, but the mindset shift from meetings to machine-readable processes is the harder lift. Second, value framing — enterprise finance is wired for committed-benefit business cases that assume knowable scope, solution, and ROI upfront. For agentic AI, the solution and the business case are often discovered through doing the work. The authors cite Accenture's own research showing AI achievers generate 50% higher revenue growth than peers — not from cost-cutting but from entirely new capabilities. They argue CFOs must adopt a VC portfolio mentality: back a range of bets knowing most will not compound, rather than demanding guaranteed payback from each individual project. Third, delivery model — agentic systems are non-deterministic and emergent, meaning they cannot be scoped or milestoned like traditional software. The correct delivery language is hypothesis-driven loops: build, evaluate, iterate, accumulate statistical confidence. Teams need to hire and reward people who can translate probabilistic outcomes into stakeholder confidence, not just ship features. Fourth, trust — the authors introduce a progressive autonomy ladder for production deployment. Stage one is shadow mode, where an agent runs alongside human workflows but cannot affect outcomes; divergences between human decisions and agent recommendations are used as iteration signals. Stage two is advisory mode, where the agent runs live and recommends actions that humans approve or reject. Stage three is controlled autonomy with narrow, low-risk actions and explicit kill switches. Broader autonomy is earned only by achieving evidenced outcome confidence — not by completing project plan milestones or passing binary tests. Fifth, competitive moat — transactional memory (CRM, ERP, SOPs) gives every incumbent a floor, not a fortress, since every competitor has an equivalent. The real moat is living memory: the edge cases, corrections, emotional intent, and behavioral signals generated uniquely at each enterprise's specific scale and context when customers interact with the product. Every feature shipped should either generate a feedback signal or act on one; features that do neither are immediately copyable.

The session closes with a three-part prescription: bet like a VC across a portfolio of AI initiatives; upgrade organizational operating speed to machine pace by treating governance automation as the CTO's primary technical debt; and engineer for trust from day one through progressive autonomy and feedback loops, not just for feature completion.