I spent most of last year rolling my eyes at LLM demos. Another chatbot that answers HR questions. Another support assistant that paraphrases the manual. Don't get me wrong, those are fine, but if that's all you think large language models are good for, you're leaving the actual value on the table. The projects that moved the needle for my clients had nothing to do with conversation.
One of the first real wins came from a manufacturing client drowning in engineering change orders. These are dense, half-scanned PDFs that suppliers send in whatever format they feel like that week. A team of three engineers spent Mondays just triaging them, figuring out which ones needed immediate attention versus which could wait. We wired an LLM into that intake pipeline. It reads the document, extracts the part numbers, affected assemblies, and urgency signals, then drops a structured summary into the ERP. It's not flashy. It doesn't talk to anyone. But it cut Monday morning triage from four hours to about twenty minutes, and the engineers stopped hating their jobs.
I made a mistake on that project though, and it's worth mentioning because I see other teams walking into the same trap. We started by feeding the LLM every change order with no filtering, assuming it would learn context on its own. It didn't. It confidently misclassified a supplier's request to delay a shipment as a critical design flaw because the word 'critical' appeared in the boilerplate legal footer. That caused an escalation to the plant manager over nothing. I still cringe thinking about it. What fixed it was boring: we added a lightweight rules layer before the LLM even sees the document. Kind of like a bouncer deciding who's sober enough to enter. The LLM handles the fuzzy reading, but the deterministic checks catch the obvious stuff. You need both.
Another place LLMs shine is legacy code. I've been on more than one modernization project where the original developers retired and the comments are in a language nobody on the team speaks. Not a joke, one client had a COBOL system with inline comments in Danish. Rather than spend months trying to hire a Danish speaking COBOL developer, we had an LLM translate and annotate the subroutine logic. It wasn't perfect. We still had senior engineers review every output. But it collapsed the understanding phase from what would have been nine or ten months down to about six weeks. That's the kind of unglamorous work that rarely makes the keynote stage.
There's also document comparison and contract analytics. Legal teams hate reviewing vendor agreements. We built a tool that pulls out payment terms, termination clauses, and liability caps from hundreds of PDFs and flags discrepancies against standard templates. The general counsel told me it felt like going from reading tea leaves to getting a spreadsheet. That's the real enterprise pattern, by the way. Not replacing people with chatbots, but giving them a first draft, a structured extraction, or a highlighted discrepancy. The human still makes the call. The LLM just makes the call ten times faster.
If you're evaluating LLM use cases, skip the demo that pretends to understand your company culture. Look for the places where your team spends hours reading, summarizing, comparing, or translating. That's where the return hides. And start with a simple rules layer, so you don't cause a plant manager to yell at you over a legal footer.