The Trust Phase: What Does it actually Take to Trust Agentic NetOps - ILOG SA
Financial Performance:
- The transcript does not contain specific financial metrics (revenue, earnings, margins, cash flow, or debt) as the session is a technical webinar on product capabilities rather than a corporate earnings call.
- No industry-relevant financial data or market capitalization figures were disclosed.
Guidance and Future Outlook:
- Management expects the primary customer focus to shift in 2026 from feature capability ("can it do this?") to trust, security, and evaluation ("will it do this reliably tomorrow?").
- Future product development aims to increase determinism within probabilistic AI models through ontological grounding via knowledge graphs and continuous real-time evaluation.
- Long-term goals include transitioning human operators from repetitive manual prompting to strategic verification and policy governance.
- The market trend indicates a move toward "agentic" NetOps where AI detects, predicts, and recommends actions, with customer willingness to fully autonomous operations expected to grow slowly over time.
Business Segments and Product Lines:
- Primary focus is on IBM Network Intelligence (INI), an operational agentic layer designed for autonomous network investigation and remediation.
- INI functions via a "Sense, Reason, Govern, Act" loop, integrating with existing observability tools (e.g., Prometheus, Splunk) rather than replacing them.
- Key capabilities include:
- Time Series Foundation Models: For correlating telemetry into symptoms.
- Agent Orchestration: Using master/sub-agent structures to handle network-wide incidents.
- Governed Actions: Human-in-the-loop approvals for configuration changes.
- Digital Twin Verification: Pre-validation of changes using Forward Networks Predict before live deployment.
- The product integrates with IBM Bob for AI-driven playbook authoring and Red Hat Ansible Automation Platform for execution.
Market and Competitive Landscape:
- Adoption Rates: 55% of IT organizations now require AI capabilities from network management vendors; 55% are actively using AI tools; 29% plan deployment within 12 months.
- Trust Levels: Only 31% of surveyed enterprises completely trust AI network tools, while 64% are "somewhat trusting" (preferring "trust but verify").
- Autonomy Preferences: Less than 6% desire fully autonomous operations; the majority prefer "agentic AI for assisted operations" where AI recommends actions for human approval.
- Barriers to Entry: Security and compliance risks are the #1 barrier to deployment, followed by budget and integration complexity.
- Competitive Pressure: Vendors are increasingly pressured to move from "black box" AI claims to transparent architectures and explainable reasoning.
Risks and Challenges:
- Data Quality Issues: 36% of organizations struggle with data collection errors and storage; 35% face inconsistent formatting across vendors, preventing effective AI utilization.
- Reliability Decay: Compounding errors in multi-step agent workflows can reduce end-to-end reliability to as low as 36% if individual steps have 95% reliability.
- Hallucination and Non-Determinism: AI models can provide false insights or incorrect tool calls; 46% of users report seeing false/mistaken insights, though this does not always erode trust if errors are detected.
- Security Risks: Agents lack human judgment and strict adherence to role-based access control (RBAC) can lead to unauthorized actions in production environments.
- Cost Economics: Token usage, inference costs, and tool call failures impact the cost-per-resolution metric.
Management Commentary and Tone:
- Tone: Pragmatic, transparent, and technically confident; emphasizes reliability over marketing hype.
- Key Message: Customers no longer care about model capabilities but rather about auditability, reproducibility, and security.
- Philosophy: Treats AI agents as "junior engineers" that require supervision and "polishing" via human feedback.
- Stance on Data: Management asserts that data quality is the root cause of AI failure in NetOps, emphasizing the need for clean, context-rich inputs over raw telemetry.
Analyst Questions and Answers:
- Q: What is the single most common reason agent deployments stall?
- A: Security and compliance risks are the primary blocker, followed by integration complexity and budget. Security teams require strict attribute-based access control (ABAC) rather than simple RBAC to ensure agents act with appropriate judgment.
- Q: How can buyers validate vendor trust claims?
- A: Buyers should demand transparency on model training, access to POC environments for continuous monitoring, and evidence of "explainable AI" where feedback loops correct the agent's behavior.
- Q: How should buyers balance tool quantity (MCP tools) against hallucination risks?
- A: Adding tools increases context but also error surface; tools should only be added if they reduce agent uncertainty. The focus should be on outcome quality, not the number of tools.
- Q: How does IBM handle non-determinism in production?
- A: IBM treats evaluation as continuous, using golden datasets, code-based graders, and real-time instrumentation (OpenTelemetry) to track drift and measure "Resolution Honesty" and "Faithfulness."
- Q: What is the single most common reason agent deployments stall?
Other Key Points:
- Demo Evidence: A live demonstration showed INI autonomously diagnosing an MTU mismatch causing OSPF adjacency failure in ~7 minutes, validating the fix via a digital twin, and generating a remediation playbook.
- Audit Metrics: In internal testing (1,900 investigations over 6 weeks), INI recorded a 6% tool call error rate, resulting in a cost of ~$0.027 per audit, demonstrating the feasibility of continuous auditing.
- Human-AI Collaboration: Successful implementation relies on "tacit knowledge" integration, where engineers can reorient the agent's reasoning trajectory, improving future performance.
- Evaluation Framework: IBM proposes four rules for agentic ops: 1) Evals as specifications, 2) Measure business outcomes (MTTR, cost), 3) Monitor for drift, and 4) GA is where testing begins.
- Company
- ILOG SA
- Keywords
- ILOG SA