Practical AI · 3-minute read
There is something deeply reassuring about five tidy bullet points. A sprawling document goes in. A confident little executive summary comes out. Finally, the organization has achieved alignment, or at least a more attractive way to disagree.
Then somebody notices the sentence that did not make the cut.
A summary has to leave things out. That is its job. The trouble starts when the missing detail changes what a reader is allowed to do, what happens next, or whether the cheerful conclusion is true.
The exception is part of the answer
Consider this fictional equipment-rental policy: “Standard equipment can be reserved online. Specialty equipment requires written confirmation from the rental desk before the reservation is final.”
An illustrative bad summary says: “Customers can reserve equipment online.” Shorter? Absolutely. Ready to guide someone booking a specialty lift? Not quite. The exception has wandered off, presumably to attend a more important meeting.
The useful version keeps the distinction: “Standard equipment can be reserved online; specialty reservations are not final until the rental desk confirms them in writing.”
This is a constructed example, not a recorded tool test. It shows the review problem: a statement can sound reasonable while dropping the condition that makes it useful.
Tell the tool what decision the summary supports
“Summarize this” leaves a lot of room for interpretation. “Summarize this for a coordinator deciding whether a reservation is confirmed” gives the task a purpose. Include the audience, the decision and the details that must survive compression.
Ask for exceptions, deadlines, dependencies, exclusions and unresolved questions separately. If the source contains conflicting statements, the summary should surface the conflict. It should not settle an internal disagreement because one paragraph sounded more senior.
Using only the supplied document, prepare a short summary for [audience] deciding [decision]. Preserve conditions, exceptions, exclusions, deadlines and uncertainty. For each key point, give the source heading or page and a short exact supporting passage. Add separate sections for “Details that could change the decision” and “Unresolved or conflicting information.” If the source does not answer something, say so. Do not invent a citation, permission or conclusion. Source: [approved or fictional material].
Check what is missing as well as what is there
Put the summary beside the original. Check the passages supporting the decision, then look for the source's qualifications: words such as “unless,” “except,” “subject to” and “only after” are useful starting points. They are not a complete search strategy; important limits can be phrased other ways.
Open the cited section yourself. A page number supplied by an AI is a pointer to inspect, not proof. NIST's Generative AI Profile identifies confidently incorrect output and fabricated citations as risks. A second confident answer from the same tool does not resolve them.
Use material permitted in your organization's approved system. For decisions with serious consequences, the responsible specialist needs to review the source, not just the convenient version.
The goal is a shorter route to understanding. If a summary removes the detail that changes the decision, it has saved reading time by borrowing trouble.
Your turn: What small exception in your work tends to become a very large problem when somebody leaves it out?
Source: NIST Generative AI Profile, section 2.2. The rental example, review method and prompt are original editorial examples, not NIST findings or a guarantee of accuracy.
