Research insight · AI
· 1 MIN
Agentic operations: the new structure of infrastructure work.
Agents change who prepares the work, not who owns the decision. The structure around them decides whether they help.
The hype has a failure rate.
Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
The failure is rarely the model; it is the operating structure around it.
How agent teams fail.
A Berkeley-led study of multi-agent systems (Cemri et al., 2025) identified 14 failure modes in three groups: specification and system design, misalignment between agents, and weak task verification.
Each maps to an organisational question: who does what, what do they share, and who checks the result.
What works in industry.
Research on industrial maintenance (O'Donncha et al., IBM and Imperial College London, ACL 2026) shows the pattern that holds up: agents reason only over structured, traceable evidence from work orders, meters and failure knowledge, a deterministic verification loop suppresses unsupported conclusions, and the system acts as a governed decision-support layer.
The new structure.
Four roles change. Specialised agents take intake, checking, modelling and drafting. A shared, governed state replaces copies of the truth. Verification becomes a role, not an afterthought. People keep authority within delegations, and the audit trail records who decided, when and why. Agents prepare. People decide.
- 01Intake
- 02Check
- 03Model
- 04Draft
- 05People decide
- 06Verified outcome
The gate is a person, not a step.
In short
The organisations that benefit will redesign the work around evidence and authority first, and add agents second.
Sources
- Gartner press release, 25 June 2025: over 40% of agentic AI projects will be cancelled by end of 2027. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
- Cemri et al., “Why Do Multi-Agent LLM Systems Fail?”, arXiv:2503.13657. https://arxiv.org/abs/2503.13657
- O'Donncha et al., “Evidence-Driven Reasoning for Industrial Maintenance Using Heterogeneous Data”, ACL 2026 Industry Track. https://aclanthology.org/2026.acl-industry.50.pdf
Related insights
Research
The research behind the insight.
turns research, industry expertise and verified outcomes into methods the engine can run.
Explore Research