I worked with 20+ large and global enterprises involving AI first, enterprise business solutions in the supply chain ecosystem. The following observations and practices reveal similar patterns regardless of industry sector.
1. ‘Silver Bullet’ Prompts vs. The 3 Layer Approach
- Current Belief: Deploying prompt generated agents with agentic workflows and expecting generalized LLM models to automate and consistently deliver repeatable actions and with relevant explainability.
- Implementation Reality: Relying on generalized LLM model intelligence to compute and express industry and enterprise specific business actions requires small language model (SLM) knowledge to automate industry and enterprise specific needs and requirements. SLMs enable contextual actions using proprietary business information.
2. Prompt Engineering Business Logic vs. Deterministic Functions
- Current Belief: Relying on prompts to solve big data, business math problems and deliver actions that are mathematically accurate and consistently repeatable.
- Implementation Reality: Core business logic should reside in different software layers. Position SLM and LLM models and prompts as a reasoning and extraction layer – separate from deterministic mathematical function layer.
3. Agentic Business Workflows vs. Human-in-the-Loop
- Current Belief: Develop, automate, and productionize all business agentic workflows.
- Implementation Reality: Deploying enterprise class agentic workflows requires human in the loop (HITL) workflow processing. Autonomous business workflows and agents introduce unmanageable organizational and compliance risks at key business decision points. “Trust but verify” is an underlying enterprise class first principle.
Takeaway
For B2B enterprises (manufacturing to business customers), following a roadmap that integrates agentic workflows with vendors and customers, embeds proprietary business knowledge in specialized small language models (SLMs), and keeps humans in the loop materially increases the likelihood of successful AI application deployment and demonstrable ROI.