The mandate for orchestration has never been more urgent. According to 2024 Enterprise Management Associates (EMA) research, 91% of technology leaders view AI-enhanced orchestration as critical within the next two years, and 80% are shifting toward enterprise-wide orchestration to run connected digital services.
Much of that urgency comes from AI itself. Enterprises have invested heavily but returns lag. MIT’s 2025 State of AI in Business study found ~95% of generative AI pilots had no measurable P&L impact, and Gartner expects over 40% of agentic AI projects to be cancelled by 2027. MIT also noted that while half of generative AI budgets go to sales and marketing, the biggest returns actually come from unglamorous back-office automation.
Deploying autonomous AI agents without a centralized strategy threatens to multiply existing enterprise fragmentation. To successfully scale, IT leaders must bridge the gap between traditional automation and agentic AI.
The Architectural Debate: Deterministic vs. Agentic
When building or renovating workflows, architects face a fundamental choice: lean on deterministic automation or build agentic AI applications.
Deterministic automation is the traditional operational backbone. Engineers sequence tasks with strict dependencies, and every run follows the exact same path.
Agentic AI works out steps at runtime based on a goal. Agents use tools, APIs, memory, and large language models (LLMs) to reason through the next best action. The practical difference is when reasoning happens: once at design time, or repeatedly at runtime—paid for each time.
Agentic advocates counter that models are improving, and Markdown instruction files can ground agents to run routine steps locally. However, an instruction file that runs identically every time is just a script by another name.
The real difference is where that script lives. Buried inside an agent, deterministic logic sits outside the orchestrator’s view, lacking a shared schedule and consistent audit trail. Hidden scripts widen observability gaps. Deterministic steps belong in the orchestration layer.
The Physical Realities of Agentic Workflows
Agentic flexibility carries a runtime price:
- Cost: LLM tokens are a recurring tax. Every agent decision burns tokens, whereas deterministic execution costs nearly zero per run.
- Latency: Model hops aren’t instantaneous. In workflows with strict Service Level Agreements (SLAs), per-step model calls degrade execution times.
- Accuracy: Errors compound. A ten-step agent chain where each step is 95% reliable completes correctly only ~60% of the time.
- Auditability: When a script fails, you get a stack trace. When an agent fails, you usually get a plausible explanation and little else.
The Decision Rule: Volume × Variability
To cut through the noise, enterprise architects need a practical framework. The guiding principle: push non-determinism onto the smallest surface area where it truly adds value.
IT leaders can categorize workloads by Volume (how throughput multiplies per-run costs and errors) and Variability (how unstructured inputs break traditional decision trees).
- High Volume, Low Variability: Needs AI only at design time. Agents are a poor fit because token, latency, and error costs repeat every run. Use fully deterministic automation.
- Low Volume, High Variability: Runtime AI makes sense here, as inputs are too unstructured to map into rigid rules.
- High Volume, High Variability: The answer is a hybrid. A deterministic skeleton handles scheduling, dependencies, retries, and state. Agents are injected only at steps where inputs require judgment.
This dividing line is dynamic: when an agent consistently makes the same decision, that path should be promoted to deterministic automation.
Establishing an Intelligent Control Plane
The future of the enterprise is hybrid. Standard deterministic jobs and intelligent agentic tasks will coexist in single, end-to-end orchestrated workflows.
Managing this safely requires an intelligent control plane—a centralized governance framework connecting stakeholders, AI models, and enterprise systems. This plane must deliver financial accountability, allowing platform owners to transparently attribute token consumption and infrastructure costs across business units.
Agentic AI is an architectural shift, not a standalone tool. Implementing governance and observability across a fragmented automation estate is the crucial first step. IT leaders must define their enterprise orchestration strategy today to regain control, fortify against security threats, and scale AI initiatives efficiently.
To read more about what is needed from an enterprise control plane, read the research: From Outcomes to Authority: Defining the Enterprise Control Plane.
Bio: Rajeev Kumar brings over two decades of enterprise software development and product management expertise to Broadcom’s global customers. He is passionate about discovering customer problems and finding innovative ways to solve them. Rajeev leads the product management and strategy teams for Broadcom’s AOD Automation product lines.
