You bought the seats. People use them for email and summaries. The report that eats three weeks every quarter still gets built the same way it always did.
The obvious move is to paste the source material into the assistant and ask for a draft. If you have tried it, you know what came back. Something that reads fine and is wrong in ways you have to check line by line.
That result is not a prompting failure. It is the wrong shape of tool for the job, and no amount of prompting fixes the shape. This note lays out what the right shape is and walks it through one report end to end.
Where the assistant breaks down
The tool is not the problem. It is doing exactly what it was built to do, which is answer a question you ask it. Your report is not a question.
It waits to be asked. Between prompts, nothing happens. Every step forward takes a person deciding what to ask next, which means a person is still running the process.
It does not go get anything. The data lives in your cost system, your shared drive, a vendor portal, and four inboxes. Someone still collects it and pastes it in. That collection is most of the three weeks.
It starts fresh every quarter. It has no record of how you framed the last one, so the structure drifts and this quarter stops being comparable to last quarter.
It will not show its work. No source trail means your reviewer verifies every number by hand. Checking a draft you cannot trust costs more than writing it did.
One long prompt does not hold. Give it everything at once and it skips things. It does not tell you what it skipped, so the gaps show up in review, or after the report is out.
What works: a workflow that runs the steps
The same models can do this. They need to be pointed at the job differently.
An agentic workflow lives outside the chat window. It connects to the systems the source material already sits in, pulls what it needs, and runs the steps in order without anyone starting each one. Each step has a defined output, so the next step is working from structured data instead of a wall of text. Where the work needs judgment, the workflow stops and routes it to the person who has that judgment, then carries their answer forward.
Nothing about it is a black box. Every figure in the finished draft traces back to the document it came from. Your reviewer checks the ones that matter instead of all of them.
Worked example: a quarterly market analysis report
To make that concrete, take a deliverable most people in construction and real estate will recognize.
A national contractor publishes a market analysis report every quarter. It draws on material cost indices, regional labor data, subcontractor bid history, supplier pricing, and published forecasts. Assembling it pulls senior people off client work for weeks, and by the time it publishes, some of it is already stale.
Step one, separate the mechanical work from the judgment work
Walk the report backward from the finished document and mark every step as one or the other. Gathering the sources, normalizing them into a consistent format, pulling the figures, and drafting the regional sections are mechanical. They are repetitive, they follow rules, and doing them by hand is where the weeks go.
Deciding what the numbers mean for the coming quarter is not mechanical. Neither is knowing that a spike in one region reflects a single project rather than a trend. That is why the senior people are on it, and that part stays with them.
Most reports split roughly this way, and the split is the whole design. Automate the mechanical steps, route the judgment steps to a person, and connect the two so neither waits on the other longer than it has to.
Step two, build it in layers
Each layer does one job and hands a defined output to the next. That is what keeps the whole thing checkable, and it is the difference between a workflow and one very long prompt.
Sources come in and get normalized. Figures get extracted with a citation attached to each one. Analysis runs on the structured result rather than the raw documents, so it is reasoning over clean numbers instead of re-reading PDFs. A human reviews at the gates that matter. The draft assembles from parts that have already been checked.
Step three, put the review gates where the risk is
Not every step needs a human. Two usually do.
The first is extraction, because a wrong figure that gets past this point contaminates everything downstream. The second is analysis, because that is where interpretation happens and interpretation is the thing your name is on.
Everything else runs unattended. Reviewers see the handful of items flagged as uncertain instead of the whole document. That is what turns review from a week into an afternoon.
What you get out
What you end up with is a process that runs, rather than another tool your team has to operate.
The report drafts from live sources, with every figure linked to where it came from. The structure holds quarter over quarter, so the comparison across periods is real. Your experts spend their time on the judgment calls instead of collecting spreadsheets. And the next cycle runs the same way, without rebuilding it.
Applying this to your own report
The report above is one shape. A grant report, an owner report, a board packet, a regulatory filing, an annual review: same anatomy. Recurring, source-heavy, mostly mechanical, with a thin layer of judgment that matters more than everything around it.
The work of adapting it is mapping your version. Where the sources live, how much of the output is rules, and which two or three decisions actually need your people.