09:10 — the pack
An investment committee received a 140-page diligence pack. An analyst used an approved enterprise assistant to create a two-page summary. The output was clear, balanced and cited sections of the source.
11:30 — the decision
The committee focused on customer concentration and accepted the technology risk as manageable. A conditional approval was issued.
Three weeks later — the contradiction
During legal drafting, counsel found that a critical licence could terminate on change of control. The summary had mentioned the licence under “standard third-party dependencies,” while the source described termination risk in a footnote to an annex.
What failed
The model did not hallucinate. It ranked. Compression changed prominence, and prominence changed the decision. The analyst checked whether statements were supported, but not whether the summary preserved the source’s risk hierarchy. The committee treated an aid as the evidence layer.
The interruption control
- Define questions the summary must answer before generating it.
- Require citations to the exact passages behind consequential claims.
- Show omitted high-risk clauses, not only selected themes.
- Make a person attest that the source’s qualifications and exceptions survived compression.
- Keep the original pack accessible at the point of decision.
The compression audit
Choose the five statements in the summary most capable of changing the decision. For each, open the source and ask: what qualification, exception or opposing evidence sits within two pages? If the answer changes the force of the statement, the summary is not decision-ready.
A summary can be factually defensible and institutionally misleading. The risk lives in what became small.