Agents Don't Fail in the Sandbox: What Actually Breaks When You Ship AI Agents to 1000's of Users
About this session
Everyone's building AI agents. Few are shipping them into enterprise production with governance, cost control, and user trust intact.
As the Lead Consultant, Atlassian AI, at Eficode, I've spent the last 18 months helping organisations move from sandbox experiments to governed, production-running agents, from 50 user pilots to 1,000+ seat rollouts. I've watched the same failure patterns repeat across every engagement, and none of them are the ones people expect.
This talk shares the honest, field-tested lessons from deploying AI agents in regulated enterprise environments: what breaks when you hand agents real workflows, why the "blank agent problem" kills more initiatives than technical complexity ever does, and what actually happens when every team starts building their own agents without a shared governance model.
I'll cover:
Why 95% of enterprise AI agent investments show no measurable return and the three patterns that separate the ones that do
The governance gap: what happens when every team builds their own agent with their own knowledge sources, and you realise nobody owns the question "what can this agent see?"
A real customer case where we turned scattered compliance inputs across 70+ attachments into structured, actionable planning (including what failed first)
The blank agent problem: why giving people a powerful agent with no structure is worse than giving them nothing
Adoption patterns that survive contact with real users vs. the ones that collapse after the workshop high
No product demos. No theory. Just the patterns, anti-patterns, and hard-won lessons from the field.
Speaker
Key takeaways
- A framework for identifying where agents will deliver ROI vs. where they'll become expensive toys
- Governance-first architecture patterns for multi-agent environments
- The cultural and organisational shifts that make the difference between a pilot and a practice
Related sessions
- When AI Joins the Team: Building an AI-Native Engineering Culture
- Observability for AI Agents: Tracking Performance and Iteration with MLflow
- Your Inbox Is the New Interface: Building AI Agents on Chat applications
- AI-GENERATED VOICE SYSTEMS: INTEGRATING COPYRIGHT LICENSING, BIOMETRIC DATA PROTECTION AND AUTOMATED ENFORCEMENT MECHANISMS