In full
Most business use of AI assistants stops at the same wall. The answer is good. The answer is in a chat window. Somebody now has to turn it into the thing that was actually needed, which was a spreadsheet, a report, or a deck. The assistant did the thinking and handed back homework.
Pick one document your team assembles by hand on a schedule: the monthly pack, the reconciliation, the exception report. Before automating any of it, write down where its inputs come from, who signs it off today, and what a wrong figure would cost. Then have the tool produce one cycle as a draft, and have the usual reviewer check it the way they would check a junior's work. If it survives that, run the next cycle. If it does not, you have learned something cheaply and early.
What changes when the output is a file
Document skills close that gap. They give the model the ability to produce real artefacts: an Excel workbook with formulas that calculate, a Word document with the structure a reviewer expects, a PowerPoint deck laid out to be presented, or data pulled out of a PDF that arrived as a scan.
This sounds like a convenience feature. In practice it changes which problems are worth pointing AI at. A capability that produces text is useful where text is the deliverable. A capability that produces a working file is useful everywhere a document is the deliverable, which in most businesses is a much larger share of the work.
An answer you have to retype has moved the work, not removed it.
Where it lands first
- Recurring reporting. The monthly pack that is assembled by hand from three systems, every month, by somebody senior enough to know where the numbers live.
- Reconciliation. Comparing two sources and producing an exception list, which is structured work with a structured output.
- Document extraction. Turning invoices, statements, or scanned forms into rows, which is the least enjoyable task in most finance and operations teams.
- First drafts of formal documents. Proposals, engagement letters, and board papers, where the structure is stable and the content changes.
The part vendors skip
A generated file carries an authority a chat message does not. A spreadsheet looks finished. It has formulas, formatting, and totals that add up, and that presentation makes it easy to accept without checking. This is a real hazard, and it is worse in a regulated setting where the document ends up as evidence.
So the discipline matters more, not less. A generated model still needs someone to check the assumptions. An extracted dataset still needs a sample verified against the source documents. The right posture is that these tools remove the transcription, not the review.
The questions worth asking before you deploy one
- Where does the file get produced? If the source data contains personal information, the processing location is a data-protection question, not a technical one.
- What did it read to produce this? A deliverable you cannot trace back to inputs is difficult to defend in an audit.
- Who checks it, and against what? Name the reviewer and the check before the first output, not after the first mistake.
- What happens when it is wrong? The interesting question is not the error rate. It is whether an error is caught before it reaches somebody who acts on it.
The honest framing
Document generation is not the impressive part of modern AI. It is the plumbing that makes the impressive part usable. Most of the value we see in client work comes from removing the mechanical step between a good answer and a filed document, because that step is where competent people spend hours that produce nothing new.
It is also the easiest capability to evaluate honestly. Either the workbook opens and the formulas calculate, or it does not. That is a refreshingly concrete test in a field with a lot of demonstrations and fewer measurements.
Document skills let an assistant produce the artefact instead of describing it: working spreadsheets, formatted documents, decks, and structured data pulled out of PDFs. That moves AI from tasks where text is the deliverable to the much larger set where a document is. The trade is that a generated file looks finished, so decide who reviews it and against what before you put one into a process.