Agentic AI
Enterprise AI Needs More Than Chatbots: Why Operational Intelligence Matters
Agentic systems work best in enterprises when they're embedded into real operations and teams, not bolted on as isolated AI layers, argues technologist Arun Joseph.
InfoQ
Most enterprise AI projects fail not because the technology is weak, but because teams treat agentic systems as greenfield problems—ignoring the messiness of existing infrastructure, teams, and organizational boundaries. That's the core challenge Arun Joseph raises in reflecting on building one of Europe's earliest enterprise-scale agentic platforms at Deutsche Telekom and now founding Masaic, a multi-agent systems company focused on operational intelligence.
Two Classes of Agentic Systems#
The agentic AI conversation often conflates very different use cases. Joseph draws a distinction between Level 1 agents—chatbots and basic task orchestration that wire together existing APIs and functions—and Level 2 systems: operational intelligence platforms that synthesize data, reason across multiple domains, and drive business outcomes.
A practical example illustrates the difference. In heavy machinery operations, a Level 2 system might ingest IoT telemetry, service records, incidents, and frontline worker insights to answer: "We're seeing a spike in error codes from machinery in the Midwest—what should we do?" The system would drill down through data, identify that 73% of cases stem from air filter management failures, automatically generate a contextualized standard operating procedure with human approval gates, and recommend stock-up actions across warehouses.
This requires ephemeral agents—temporary task-specific agents spawned and run in parallel—that don't exist as pre-built entities but are dynamically created and synthesized into actionable outcomes.
The Fault Line Problem#
Why do 95% of enterprise AI projects stall? Joseph identifies the first major culprit: crossing organizational fault lines.
Enterprises aren't startups. Every API, database, and service lives behind teams with different incentives, knowledge, and approval workflows. When a new AI program spins up in isolation—bright engineers with research papers and the latest frameworks—it hits reality: that messy product-ordering API nobody fully understands, teams that guard their services, schemas that shift, and process change that requires consensus.
The solution isn't a new, detached technical team. It's embedding agentic work inside existing fault lines, staffing projects with both domain experts who understand the problem and the people closest to customer pain (frontline workers, not just engineers). Without that grounding, teams will stall waiting for API clarifications, organizational approval, and people who actually own the outcomes.
