Prompt engineering, RAG and fine-tuning are stages, not alternatives

Adapted from the update iNDustry Labs sends the AI Executive Group, a group of manufacturing and logistics leaders across the South Bend–Elkhart region.

Agent development approaches

AI agents can sound like a major technological leap. In practice most companies are already on the journey, even if all they are doing today is asking ChatGPT or Gemini a question. Agent development typically moves through three complementary approaches: prompt engineering, retrieval-augmented generation and fine-tuning. Each has distinct strengths, and the best outcomes come from combining and weighting them rather than treating them as alternatives.

Prompt engineering: where you probably started

Most teams are already doing this. Asking a model to summarize supplier risks, draft emails or explain trends in a spreadsheet is an early form of agent interaction. The value is quick and accessible, with no new systems integration or infrastructure required.

The limits show up fast. The agent relies entirely on how the question is framed, and it has no awareness of company-specific policies, contracts or historical decisions unless someone pastes them in. Prompting steers behavior. It does not embed institutional knowledge.

RAG: adding business-specific context

As organizations want more reliable outputs, they add retrieval-augmented generation, where the agent pulls from internal documents and databases you have specified. That grounds responses in real, trusted information. It still does not guarantee accuracy, so keeping a person in the loop remains essential.

This is the stage best suited to working with student teams, since connecting document repositories to an agent is a much lower barrier than fine-tuning and does not change the underlying model. The result is a more context-aware agent that delivers real business value while staying flexible and lower risk.

Fine-tuning: predictability at scale

Further along, it can be worth working with computational scientists to fine-tune a model. Fine-tuning shapes behavior so that tone, format and decision logic are applied consistently, and done properly it further reduces the chance of an incorrect output. For manufacturing this matters wherever outputs must be standardized, auditable and aligned to internal frameworks, supplier risk scoring and regulatory reporting being the obvious cases.

The same problem at all three stages

Take procurement. With prompt engineering, a purchasing manager asks a model about supplier risks or negotiation strategies and gets useful but generic best practice. With RAG, the agent reaches approved supplier lists, contracts and performance data, so the answers are grounded in the company's own context. With fine-tuning, the agent consistently applies internal risk logic and preferred negotiation language, and stays dependable as it is extended across the rest of the supplier portfolio.

From the first group meeting

The group met for the first time on December 16, covering AI adoption at Notre Dame, an overview of how the group will work, the student project clinics and the business analytics capstone. Our thanks to everyone who attended and shared input that shaped the agenda for the year ahead.