In theory there is no difference between theory and practice. In practice, there is.
I love building technology in the enterprise because it's the very definition of practice. It's the toughest proving ground there is. In large, complex companies the textbook examples almost never show up, the supposed edge cases are business as usual, and the long tail of real edge cases goes on forever.
The data nobody cleaned
Take enterprise data. There's clean, structured data, maybe 20% of it. The rest lives in emails, documents and PDFs, if you're lucky. If you're not, it's scanned documents, faxes, images, handwriting, voice recordings, ZIP files, and file types that feel like fossils from an earlier era. And don't forget the notes and diary entries tucked away in every system.
Every insurance company has mountains of this. Email chains with three attachments. A fax of a handwritten note. Signals are buried in there, and they're easy to miss. A fraud indicator in a scanned fax. A coverage gap spread across an email thread and two attachments. Liability exposure scattered across documents in four different formats. In a large insurance operation, these aren't edge cases. They're the norm.
The forever project
For years, the tools weren't up to it. RPA made a much smaller dent than anyone expected, because it couldn't handle unstructured data and every integration was fragile and expensive to build. Intelligent document processing was supposed to fix that, but it needed too much training data, a custom model per document type, and still couldn't generalize across formats.
So the conventional wisdom became: clean up your data first. "We need to get our data in order before we can do anything with AI." It's a quixotic quest if there ever was one, and entire AI programs have stalled behind it. Yes, clean the data. But while you're cleaning, the risks are sitting there in plain sight.
Extraction isn't the point
Even when you can get at the data, extraction is only the beginning. Pulling fields out of a document would have seemed magical five years ago. Today it's table stakes.
Data doesn't volunteer and line itself up to be classified, extracted and summarized. It's notoriously misbehaved. It has to be gathered or accepted from many places: email, file shares, document management systems, applications. And it arrives in many-to-many relationships. One email can start a dozen business processes, and one business process may have to wait on three emails and a fax.
From an automation point of view, getting this far is a big win. From a business point of view, getting this far has created no value at all.
The value comes after the data. Understanding it. Spotting the risks and the opportunities. Deciding what to do. Putting the results into the systems that run the business. None of that happens if the data is wrong, and none of it matters if all you ever do is clean the data.
What I've learned is that you only create value when you can gather the data, process it, and act on it. Each step depends on the others, and no single part is worth much alone. For years the tools to connect all three didn't exist. Each part was too hard on its own.
What changed
LLMs understand semantics. No training, no custom model per document type. Point one at messy, unstructured data across formats and it understands what it's looking at. Before 2022, every kind of unstructured data needed its own strategy. Now they converge: get it to text, whether by OCR, image description or transcription, and let the model do the rest.
That makes it possible to span the whole chain, from intake to action, without brittle rules or endless cleanup. And here's the twist: the dirty data that used to block everything is suddenly the most interesting data you have. Waiting for the cleanup to finish is no longer the smart move, because the systems can finally work with the mess itself.
What's worked for us is multi-agent workflows with a hefty set of tools for processing data, calling APIs, searching, and driving a browser. Back them with memory and good context engineering, and the unstructured data tucked away throughout the enterprise becomes reachable. Add monitoring, escalation, traceability and scale, and it's ready for production.
It's an amazing time to be in enterprise technology. Problems that sounded simple and never were are suddenly solvable. I'd love to hear which old problems you're unlocking.