Share with your CIO
Broadcom’s automation division argues that enterprises are miscategorizing their AI deployment problem, framing it as an AI question when it’s really an orchestration question. The piece cites MIT and Gartner data showing that 95% of generative AI pilots produced no measurable P&L impact and 40% of agentic AI projects may be cancelled by 2027, then proposes a Volume-times-Variability framework for deciding when to use deterministic automation versus AI agents at runtime. The practical prescription is a centralized “intelligent control plane” governing both traditional scheduled jobs and agentic tasks within the same workflow.
What this means for your business
The argument lands hardest for organizations that have already deployed AI agents and are now staring at escalating token costs, degraded SLA performance, and audit trails that amount to a plausible story rather than a verifiable record. If your automation estate is fragmented, meaning different teams running agents, scripts, and orchestration tools with no shared governance layer, you’re not in a pilot problem, you’re in a sprawl problem. The Volume-Variability matrix is a useful triage tool regardless of which vendor you use to operationalize it.
The math here is worth sitting with. A ten-step agent chain where each individual step is 95% reliable completes correctly only about 60% of the time. That’s not a model quality problem, it’s a compounding probability problem that deterministic orchestration doesn’t have because the path is fixed at design time. The practical implication is that any high-volume workflow with predictable logic, think batch processing, compliance reporting, or scheduled data pipelines, should never have been agentic to begin with. Routing those back to deterministic automation isn’t a retreat from AI strategy, it’s a cost and reliability fix that frees agent capacity for genuinely unstructured decisions.
Broadcom is a vendor selling orchestration software, so the frame here tilts toward making the control-plane problem look large and urgent enough to justify enterprise platform spend. That specific incentive does produce one real distortion worth noting: the piece understates how much operational discipline is required to actually identify which agent decisions have “hardened” into repeatable patterns and promote them to deterministic steps. That feedback loop is harder to build than the matrix implies. Still, the core diagnostic, that hiding deterministic logic inside an agent creates observability debt, holds up on its own merits. If your team can’t answer “what did the agent decide at step six and why” from a log, you’ve already got the problem they’re describing.
Concept deep-dive: Deterministic automation
Deterministic automation means a workflow whose steps, branching logic, and outputs are fully defined at design time, the same input always produces the same path. Think of it as a recipe versus a chef: the recipe runs identically every time, the chef improvises. Enterprise schedulers and ETL pipelines have operated this way for decades. The business relevance is cost and auditability: deterministic execution carries near-zero per-run compute cost and produces a complete audit trail, two properties agentic AI systems structurally cannot match.
Based on reporting from The orchestration imperative: Balancing agentic AI with determinism, originally published 2026-11-02 11:28:00.
