The situation
A 40-page report lands in your inbox with a meeting on it this afternoon. You don't have an hour to read it properly. This is one of the things AI genuinely does well — but only if you ask in a way that gets you a brief you can actually rely on, rather than a bland paragraph that misses the one figure that mattered.
Step one: give it the document, don't describe it
Upload the actual PDF into ChatGPT, Claude, or Gemini rather than asking about the report from memory. All three let you attach a file. Working from the real document is what keeps the summary honest — it has the text in front of it instead of guessing from the title. If your tool can't take the file, paste the relevant sections in as text.
Step two: ask for structure, not a blob
A one-line "summarise this" gets you a vague paragraph. Ask for a brief shaped the way you actually need to use it — the decisions, the numbers, the risks, the things you'd get asked about in the meeting.
Step three: the check that catches the miss
A summary's real danger isn't a wrong fact — it's a quiet omission. The report's one awkward finding gets smoothed over and you walk into the meeting not knowing it exists. Catch that with a second, deliberately adversarial prompt.
Step four: spot-check before you rely on it
Pick the two or three figures you'd actually quote and find them in the original document. It takes a minute with the search function (Ctrl+F on Windows, Cmd+F on Mac) and it's the difference between "the AI said" and "I checked." If the numbers match, you can trust the shape of the rest.
When one pass isn't enough
For a routine report, the brief-then-check routine is plenty. For anything high-stakes — a document a real decision hangs on, or something you'll be quoted on — do one extra thing: ask a follow-up question that tests whether the tool actually understood the material. "What's the argument the author is making, and what evidence do they give for it?" A summary can list the right topics while missing the point entirely, and this question surfaces that fast. If the answer is vague or circles back to headings without real substance, the tool has skimmed rather than understood, and you should read the key sections yourself. It's a thirty-second test that tells you whether you can trust the brief or need to go to the source.
Next time a long document lands before a meeting, upload it and run both prompts above — the structured brief and the adversarial check. Then Ctrl+F two figures against the original. You'll walk in with a one-page brief and, more importantly, knowing the bit everyone else skimmed past. The whole thing takes under ten minutes.
Why this matters
Summarising is one of the safest, highest-value ways to use AI at work, because you keep the source in front of you and you're not asking it to invent anything. The skill is in asking for the brief in the shape you'll use it, then checking it didn't sand off the one detail that counts.