When the Summary Becomes the Decision

Summaries are one of the most useful things AI does. A thirty-page report becomes five bullets. A two-hour meeting becomes a paragraph. A year of customer emails becomes a short list of themes.

That saves real time, and most of us would not give it up.

But there is a moment in every summary’s life that gets very little attention. It is the moment someone stops reading the source and starts acting from the summary. When that happens, the summary is no longer a convenience. It has become the thing the decision is made from.

Five bullets

Picture a leadership team preparing a decision about where to invest next year.

They have a thirty-page report drawn from twenty-four customer interviews. Nobody has time to read it, so an AI summary goes into the pre-read. Five bullets.

The first reads: Customer satisfaction improved from 7.6 to 8.1 this year.

The fourth reads: Some customers raised concerns about support responsiveness.

Both statements are true. Neither bullet contains an error.

In the original report, eleven of the twenty-four customers raised the support concern without being asked. Several said it would affect whether they renewed. Anyone reading the interviews would have met it again and again. In the summary, it appears once, in the fourth position, with the word “some.”

The leadership team reads the five bullets, sees the improving score, and decides to put next year’s investment into new product features. It is a reasonable decision based on what they read.

Nothing in the summary was false. And the team walked away with a different sense of what matters than the report would have given them.

What compression actually does

We tend to think of a summary as the original with the length removed. It isn’t. To make something shorter, a summary has to change its shape, and each change alters how the information will be read.

It changes prominence. The first bullet reads as the main point, whether or not the writer intended it to.

It changes proportion. A theme that took up a third of the interviews now takes up a fifth of the bullets.

It changes recurrence. Eleven separate warnings, each in a different voice, become one line. The reader loses the thing the source made obvious: that this came up again and again.

It changes language. “Some customers” and “eleven of twenty-four” describe the same fact, but they lead to different decisions.

It changes context. Points that were separate get grouped, and points that were connected get split apart, so that the result reads tidily.

And it removes things. What disappears does not appear as a gap. It simply isn’t there to be missed.

None of this makes a summary wrong. It means a summary is a new arrangement of the evidence, built before the reader arrives, usually without the reader knowing what was rearranged.

A summary can preserve the facts and still change the judgment.

Why a better summary doesn’t fully solve it

The natural reaction is that the summary should have been better. It should have said eleven of twenty-four. It should have put support higher.

And you can do better. You can tell the AI what the decision is: summarize this report for a choice between investing in support capacity and investing in new features. That is a real improvement, and it is how these tools should be used.

But a decision-specific summary still reflects assumptions about what that decision depends on. The assumption might be right. It might also miss the factor nobody thought to name, such as the way support problems might undermine confidence in the new features.

And in many cases, the decision isn’t yet defined when the summary is requested. The summary goes into a pre-read, and the question is sharpened in the meeting.

Unless the decision is part of the summarization task, there is no reason to assume the summary preserved what that decision depends on. Even when it is, the summary reflects someone’s judgment about what mattered. It is not a substitute for yours.

So the question is not whether the summary was accurate. Accuracy is the minimum a summary owes you. The question is whether it preserved what mattered for this decision.

Compression compounds

There is a further step that is easy to miss.

Once a summary is good enough to act from, people begin to treat it as the record. It gets forwarded. It gets pasted into the next deck. Someone summarizes the summary for the board. Eventually, the five bullets are the only version of the report that anyone remembers.

At each step, a little more is lost, and each loss is invisible. Every version can be an accurate compression of the one before it while sitting further from the original evidence. Nobody in the chain did anything wrong.

Over time, the summary stops describing the decision and starts determining it.

You won’t see the cost in the meeting. You’ll see it a year later, when renewals fall short and someone pulls out the original report and finds the warning on page nine, raised in eleven interviews, compressed into a single word.

Match the summary to the decision

The answer is not to distrust summaries or to read every source in full. That would give up the very thing that makes AI useful.

The better practice is to match the effort to the consequence of the decision.

For decisions that are low-stakes or easy to reverse, a summary is exactly what you need. Use it.

For decisions that are costly, hard to undo or likely to be repeated, ask three things before acting.

What does this decision actually depend on? Find that part of the source and read it directly, not through the summary.

What was compressed? Look for what recurred, what carried the strongest language and what was grouped together to make the summary tidy.

And what would change the decision if it turned out to matter more than it appears here?

Sometimes that will take real time. That is what the consequence of the decision justifies.

AI will keep getting better at telling us what a document says. That is valuable.

But whether the summary kept what this decision depends on is still a question the reader has to answer.