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: Spring 2026 AI Projects Showcase
On July 15, AI Leaders of Tomorrow fellows presented the pilots they built last spring with regional partners. What each team delivered:
- GSM ran three pilots: plain-English querying of their Plex ERP, so a question returns a downloadable answer with no SQL required; a multilingual HR assistant that answers policy questions in five languages with citations, grounded strictly in GSM's own documents; and a financial planning workflow that builds statements and variance analysis inside Excel.
- MTI built an agent workflow that verifies, triages and drafts personalized responses to inbound sales leads, taking manual lead review off the engineers' plates while keeping a person on the final send.
- Kem Krest assessed how forecastable a hard-to-predict OEM accessories portfolio really is, and built a reusable approach for demand forecasting and inventory planning.
- Vista Manufacturing stood up an AI-assisted project management setup: a capacity dashboard, a cleaned-up task template and an agent that surfaces critical-path tasks quickly.
- Hoosier Crane automated the aggregation of apprentice reports across several systems into a single dashboard.
AI works best as an assistant with a person in front of it, and the process and data groundwork matters as much as the model.The theme running through all five projects
RPA versus agentic AI
Robotic process automation and agentic AI get talked about as rivals. In practice they are converging into one system, and the useful question is which job belongs to which.
What RPA does well
RPA is a rules-follower. You script the exact steps, log in, read this field, copy it there, generate the report, and a bot repeats them precisely, at volume, every time. For stable, high-volume, well-defined tasks such as moving data between an ERP and a portal, or pulling a recurring report, it is fast, cheap to run and dependable.
Where RPA breaks
RPA is brittle. Because it leans on fixed screen positions and field tags, a vendor changing a screen or an unexpected document format can break the whole workflow, and any deviation from the happy path stops it. At scale, keeping the bots running quietly becomes the automation team's real job, a fragility tax that erodes the return already harvested.
What agentic AI adds
An agent works from a goal rather than a script. You state the outcome; it reasons about how to get there, picks the tools it needs, reads unstructured inputs such as emails, PDFs and drawings, and handles exceptions by finding another route instead of stopping. That is the move from automating a task to automating an outcome.
The practical answer is hybrid
The framing holding up in the field is not rip out RPA and replace it with agents. It is a division of labor: let the agent be the brain that reasons and decides, and let the RPA bot be one of the tools it calls to execute a stable, repetitive step. The real question is less should we replace RPA and more where should reasoning live, and where should execution stay.
Key takeaways
- Use RPA for stable, high-volume, rule-based steps. Reach for agents where judgment, unstructured data or constant exceptions are the norm.
- Budget for maintenance, not just build. Brittleness is where RPA programs quietly lose money as they scale.
- Keep a person in the loop. Agents reason probabilistically, so they should propose and act within guardrails, with people owning the final call on anything that matters.
Sourcing AI talent
The role drawing the most attention right now is the forward-deployed engineer, who pairs existing expertise in AI platforms with a detailed understanding of a company's specific process challenges to deliver custom solutions, then takes those use cases back to the AI platform to decide whether they belong in the standard product.
Hiring a large AI firm will likely deliver strong results, but it takes significant resources and can lock a business into that firm's platform. iNDustry Labs cannot promise platform-scale results through student projects; what it offers instead is a more platform-agnostic way to approach AI development, and a path to scaling it in-house. Because the fellowship runs at limited capacity each semester, several member companies have asked for help identifying full-time hires who can lead and scale AI beyond prototype projects.
What we are reading
The Amazonification of metal fabrication
The Fabricator profiles SendCutSend, a web-first sheet metal shop where a customer uploads a design, gets an instant quote, and the order flows automatically into nesting and onto the floor. In one cut-and-bend example, 27 minutes passed between accepting the quote and a boxed part. Worth reading for anyone with low-volume, high-mix work, where parts spend minutes being cut and bent but days sitting in the queue. The leverage is in automating the front end, quoting, order processing and design-for-manufacturability checks, while keeping people on the judgment calls.
AI platforms are moving from selling tools to sitting in your building
Microsoft announced a $2.5B unit that embeds its own engineers inside customer operations to build and run AI systems on site, with contracts tied to measurable outcomes, following similar moves by Amazon, OpenAI and Anthropic. The large players appear to have concluded that most of the value in AI is lost in the last mile between a pilot and production, and that closing it takes people on the floor rather than another platform license. For most companies in the region, that is an argument for investing in internal AI talent, since hiring from those firms carries a high price tag.
A policy fight over open AI models
A coalition led by NVIDIA, now dozens of companies including Microsoft, Meta, IBM and Dell, published a letter urging policymakers not to restrict downloadable, open-weight AI models, arguing they are strategic infrastructure that will diffuse AI into factories and every other sector. It is a window into the open versus closed debate that will shape whether a company can run AI on its own data and hardware. Worth noting it is a lobbying position rather than neutral analysis: the counter-view questions whether access to the weights alone delivers the transparency and security it promises.
The plumbing that lets agents talk to your systems went enterprise-grade
The Model Context Protocol, the emerging open standard for connecting AI models to tools, data and applications, shipped its biggest revision since launch on July 28. The headline is a stateless design that lets these connections scale on ordinary IT infrastructure, with hardened security and authorization. The connective tissue behind the agents in vendor demos is maturing quickly toward something IT can deploy and govern at scale.
iNDustry Labs runs the AI Executive Group alongside the AI Leaders of Tomorrow fellowship at the University of Notre Dame.