How to summarise a long report with AI without losing what matters
“Summarise this” is the most-used instruction in the world with an AI, and also the one that works worst when the document actually matters.
The problem isn’t that it summarises badly. It’s that it summarises well what usually matters in documents like that, which is not the same as what matters to you. And because the result is accurate, tidy and well written, you have no way of noticing that your paragraph is missing.
Why a generic summary betrays you
A summary is a decision about what gets left out. If you don’t state your criterion, the model picks it — and it picks what’s statistically central: the headlines, the conclusions, whatever takes up the most pages.
What almost always gets dropped is what you were after: the exception, the condition, the small print, the figure that breaks the trend.
Ask for a summary with a purpose
The difference between a useless summary and a good one fits in one sentence. Compare:
“Summarise this report.”
“Summarise this report so I can decide whether to renew this supplier’s contract. I care about missed deadlines, cost overruns and penalties.”
The second isn’t a summary: it’s a filter. And it filters by yours.
Whenever you can, say what you’ll use it for and what decision sits behind it. It’s the one piece of information the document doesn’t contain, and it completely changes what’s worth keeping.
Four ways to ask, depending on what you need
The map. “Give me the real table of contents, section by section, one line each.” Use it to know where to look before reading anything. It saves the most time and almost nobody asks for it.
The executive summary. “Ten lines, as you’d tell someone who has to decide and won’t read the report.” Careful: it’s the prettiest and the most dangerous, because it sounds like a closed conclusion.
What changes something. “What does this report say that isn’t what you’d expect? Surprises, contradictions, figures that break the trend.” This question rescues exactly what a normal summary throws away.
What’s missing. “What questions does this document leave unanswered?” It turns a summary into a to-do list.
The ten-second check
Before trusting it, ask about a specific detail that only exists inside: a figure from a table, what a numbered clause says, how many sections a chapter has.
If it gets it right, it read the document. If you get generalities — “section four sets out the service requirements” — it hasn’t seen it: it’s describing what documents like that usually contain. It’s more common than you’d think, and we cover it in what happens when you give an AI a PDF, including the page limits almost no provider publishes.
And if the report has charts, ask about one. That’s where the model that only received text falls apart.
What a summary will never give you
Tone. A report can say “it is recommended to evaluate” and mean “don’t do it”. That’s lost.
What isn’t written. If the document avoids a subject, the summary will avoid the same subject without flagging it. That’s why “what’s missing” is worth so much.
Responsibility. If the decision is yours and it goes wrong, “the summary didn’t mention it” is not a defence.
When the report decides something expensive
This is where a single opinion falls short, and not out of distrust in the machine: two careful readers don’t summarise the same way. One keeps the deadlines, another the risk, another the money. With one AI you inherit its criterion without knowing what it was.
That’s what The Judge does: several AIs read the same document separately and a fixed judge tells you where they agree — that’s probably what genuinely matters — and where they don’t. In a long report, disagreement between readers usually points at exactly the ambiguous paragraph that needed a second look.
In one line
Don’t ask for a summary. Ask for the summary you need for what you have to decide, and check with one question that it actually read it.
Fifteen extra seconds, and it’s the difference between a nice piece of text and something you can lean on.