How LLMs and agents actually work, and three patterns you can try today

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

Recap: how LLMs and agents work

Notre Dame's Center for Research Computing presented on how these systems actually behave, and how to keep them from misbehaving once you point them at your own data. The takeaways:

• An LLM is a next-word predictor: strong on patterns, weak on precise recall. A bare model will confidently invent facts. That is its default behavior, not a glitch. Grounding it in your own documents and giving it real tools is what fixes that.
• What you put in the context window, the chat thread included, effectively reprograms the model each turn. The discipline of feeding it the right context and nothing more is the real lever, not clever one-off prompts.
• An agent is just tools running in a loop: gather context, act, verify, repeat. Reliability comes from the harness around the model, the verification and the guardrails, not from the model alone.
• The late-2025 jump in agentic tools was real. A controlled study found that early-2025 tools made experienced developers about 19% slower, despite expectations of a speed-up. With early-2026 tools, the same group ran about 18% faster. The tools matured, and the operators learned.

Three AI use cases you can try today

Three patterns our teams are already putting to work. None require a deeply technical team, and all keep a person in the loop on the output.

1. Grounded chatbots over your own documents

If your knowledge lives in documents, you can build a chatbot that draws on them and avoids general-knowledge hallucinations. Staying inside the Microsoft environment, with SharePoint and Copilot Studio, is convenient for companies with CMMC or DoD compliance requirements. Two proven examples: an HR assistant that answers employee questions directly from policy documents, freeing the HR team from routine work; and, at larger firms, job-code segmentation, where a proposed job description is compared against existing job codes covering duties, FLSA status and EEO-1 category, and the agent proposes where it should roll in.

2. Agentic workflows across systems

When a workflow spans several systems, you can chain agents to handle the repetitive first pass with tools such as n8n or UiPath alongside an LLM. In sales lead triage, one agent analyzes new leads as they arrive from the CRM and runs a quick qualification pass, then a second drafts a tailored response for a salesperson to review and send. In invoice processing, optical character recognition pulls PDFs out of email and populates the ERP with the invoice data.

3. Better-engineered custom applications

Vibe coding draws mixed opinions in the developer community, for good reason: speed without discipline produces fragile systems. If your team is building custom applications, tooling that reinforces planning, architecture, testing and deployment, rather than code generation alone, raises the floor without lowering your engineering standards.

Shared by group members

Token cost management

For teams worried about token costs now or later, IBM Bob is a coding agent that directs tasks among several models and provides cost estimates and usage transparency across the development and maintenance lifecycle, so AI spend can be budgeted the way cloud or CI/CD is. It was reportedly inspired by large companies, IBM included, that burned through an annual AI budget before the end of Q1 for lack of visibility into per-application cost.

From prototypes to production code

A group member shared 2389.ai, which helps build new products and transform existing ones by testing the latest AI technologies, and publishes research from applied projects monthly.

AI platforms are maturing into infrastructure

Anthropic confidentially filed a draft S-1 with the SEC on June 1, signaling a potential IPO. The takeaway is that the major AI platforms are maturing from experiments into the kind of mission-critical infrastructure you can plan multi-year operations around.

By industry sector

• RV, marine and manufactured housing. The industry is putting AI governance on the record, with an industry safety seminar featuring a panel on the responsible use of AI in regulatory and recall compliance, drawing leaders from Forest River, Brinkley RV, Winnebago, THOR, ATC Trailers and Lippert.
• Machining and metal fabrication. Instant quoting is the 2026 dividing line for job shops. Tools now read a CAD file and return programming, setup and run-time estimates in seconds, with CAM software consolidating into a single quoting, simulation and DFM hub that leads the job rather than following it.
• Banking and finance. AI is now table stakes for community institutions. It is the top technology priority for the third straight year, with 85% of leaders saying adopters gain a real edge, even as AI-driven fraud and deepfakes become the leading security concern.
• Healthcare. Ambient documentation has moved from pilot to standard, with Indiana systems running virtual nursing and ambient scribing as operational programs, a close model for any local provider weighing where to start.