Give an LLM the text of a PDF document. Ask the model to extract values in the document or in tables. Input the values into a spreadsheet. This is at minimum a task which costs companies around the world Hundreds of Millions of dollars a year.

Sure. An example is classification. Take suppose a company that provides a platform for other companies but only if their businesses are supportable under a variety of regimes, such as industry (no cannabis say), sanctions and export controls, etc. Typically this is done by training up an army of people that poorly review websites and collateral by hand and make some determinations with a very low precision and recall, and a very high expense profile including training, benefits, HR, etc, with high attrition. Building these operations is complex, time consuming, and often fails to produce meaningfully useful results for years.

With LLM at GPT4 level ability and a modest investment in scraping, context, building an agent based tool you can automate the entire thing with a pretty high precision and recall (such as a 0.9 precision with a 0.8 recall, vs 0.4 / 0.5 for humans). Using classical NLP, schema matching, and classification techniques you could triage and improve human performance by as much as 20%. But that requires a lot of effort and provides important but modest improvements. The LLM approach however is almost effortless from an engineering effort (note all of the infrastructure is the same as the NLP/schema matching/classification/etc effort).

This may not seem earth shatteringly important to you, but for an enterprise this is monumental. It dramatically reduces risk, complexity, cost, and all the while improves outcomes by an enormous amount.

Another example is incident management. Having LLM agents monitoring incidents on slack as reduced our MTRR dramatically. They provide summaries for new joiners, broad cast management updates, etc, and take care of all the glue out response teams typically did but with a high degree of quality, low degree of variability, and extremely low latencies compared to humans.

Another example is policy management. Many enterprises are saddled with thousands of pages of external requirements from regulators, and those in turn create tens of thousands of pages of policy, and hundreds of thousands of LoC of control implementations. By indexing them into a Context+RAG LLM you can create a policy oracle, allowing you to create coverage and effectiveness testing, analyze for policy or control gaps, draft responses to auditors and regulators with total policy knowledge, answer legals questions, advise management on the implications of a business driven exception, etc. The state of the art before was a search engine dumping phrase based matches into a grid of thousands of rows of results, which was effectively useless for all of these use cases.

I just started using one of OpenAI's Assistants to turn unstructured job listings into structured JSON with inferred scores on dimensions that are aligned with my preferences and my attached resume.txt file.

Right now, I have to copy+paste my input and outputs, but my next task will be to write a Tampermonkey script to automate it.