The question I hear from engineering teams has changed, and I think it's the most important shift in the stack right now. In 2024 it was "why would I need the extra tenth of a percent?" Their parser did a pretty good job transforming a file into JSON or markdown, the retriever found the chunk, and the generated response looked fine. Good enough was good enough. In 2026 nobody asks me that. They ask whether we can tell them exactly where on the page a value came from and how sure we are, because an agent is going to act on it without a human in the loop. The math behind that is simple. With multi-step agentic systems, errors at the context generation phase (when raw data is converted into JSON or markdown) compound downstream. We ran an experiment to demonstrate this point: 200 scanned invoices through the same 25-step invoice-payment workflow, using three differing parsing engines. Field accuracy looked close. Whole invoices were not close at all. GPT 5.6 Sol only parsed 1% of invoices correctly, but most of these errors were invisible if you aren't an agent. I wrote a short ebook on what changed over the past couple of years and what bar you should hold for your preprocessing engine, ours included. RAG needed mediocre data. Agents need extraordinary data. unstructured.io/blog/rag-nee…

Field accuracy is the number everyone reports when they talk about parsing forms. It's also the one most likely to mislead you. We ran 40 grant applications through Unstructured Transform and GPT-5.6 Sol. On field accuracy they looked almost identical: 99.5% against 98.7%. You'd call it a tie. But field accuracy counts individual fields. A form can get 99% of its fields right and still have one wrong value sitting in it, which means the whole form still needs a human to catch it. So we measured what actually decides your human workload: how many whole forms came back with zero errors. Transform returned 34 of 40 clean while the frontier model got about half. The more accurate system turned out to be the cheaper one too, at roughly 7.5x lower cost per page. A wrong value that still looks plausible slips past review and ends up in a customer record or an automated decision, where it's far harder to catch and far more expensive to fix than it would have been on the page. Read more in this week's blog: unstructured.io/blog/how-to-…

Last week we wrote about why better frontier models haven't made document processing go away. So this week we tested it, on forms. If you're evaluating a parser, field accuracy is the first number most people look at, and it's also the most misleading. We ran 40 grant applications through Unstructured Transform and GPT-5.6 Sol. 99.5% vs 98.7%. Basically a tie. But count the forms where every value is right and it's 34 out of 40 for Unstructured vs about 20 for GPT-5.6 Sol. At 10,000 forms, that's roughly 15% for a person to check vs ~50% while costing 7.5x more per page. The mistakes are small, like a name with one letter wrong or a number in the wrong column. One wrong field means the whole form is wrong, but these are easy to miss because the answer still “looks” right. So when you test any vendor, score whole documents, not fields, and check what each one costs you per page. We wrote about this here: unstructured.io/blog/how-to-…