A General AI Tool Can't Tell You If Your Investigation Is Complete. We Can.
The output looked complete. The language was confident. There was a cause, some recommendations, a summary. Nothing flagged as missing. So you filed it and moved on.
But the general AI tool that produced that output had no way of knowing whether it was complete. It produced a response to your input. Completeness is a different question — and one it cannot answer.
This article is part of the series Why a Chat Window Isn't an Investigation.
What Completeness Actually Means in an Investigation
A complete investigation is one where the conclusions — the causes identified, the corrective actions recommended — are supported by the evidence that was actually collected. Completeness is not about length. A ten-page report can be incomplete. A four-page report can be thorough.
Completeness means the evidence collected covers all the relevant categories for the incident type, the root cause analysis is grounded in that evidence rather than in assumptions or general knowledge, gaps between what the evidence shows and what the conclusions claim are identified and noted, and the investigation record would hold up if someone reviewed it with fresh eyes.
A general AI tool cannot assess any of this. It produces text. The text is based on what you described. Whether the evidence behind that description is sufficient to support the conclusions being drawn is a question the tool was never designed to ask — and cannot answer.
The Problem With an Unchecked Output
When a general AI tool produces an investigation summary, its output is shaped by what sounds reasonable and coherent, not by what the evidence actually shows. If you described the incident in a way that suggested a clear cause, the response will reflect that cause confidently — regardless of whether the evidence supports it.
This creates a specific kind of risk: an investigation that appears complete but has invisible gaps. The report says maintenance records were not a contributing factor — but the maintenance records were never collected, so there's no basis for that conclusion. The report identifies training as adequate — but training records were never reviewed.
These gaps don't make the report look wrong. They make it look right — until someone asks for the underlying evidence.
What a completeness check surfaces: A systematic check against the full evidence set can flag evidence categories that are under-represented or missing entirely — gaps that would not be visible in the final output without that check, and that an insurer or regulator would look for when reviewing the file.
The Worked Example: No Signal That Anything Was Missing
A warehouse worker slips on spilled oil near a loading dock in a small manufacturing business and loses time from work. The owner describes the incident to a general AI tool and receives a well-structured output. There is no indication that anything is missing. The response doesn't say "you should also check whether maintenance records are complete" or "I notice you haven't mentioned a training history." It produces a confident, coherent summary based on the description provided.
Weeks later, the insurer requests the investigation file and asks specifically about equipment maintenance records and training documentation. The owner realises neither was collected. The investigation report made no reference to either gap — not because they weren't relevant, but because the general AI tool had no mechanism to detect the gap in the first place.
A structured investigation process performs a completeness check — assessing whether the evidence collected is sufficient to support the conclusions being drawn, and flagging where gaps remain before the investigation is closed.
Want to see how a structured investigation handles this? Start your free trial
What a Completeness Check Actually Does
A systematic completeness check compares the evidence collected against what a thorough investigation of this incident type typically requires. It identifies evidence categories that are under-represented or missing entirely. And it flags those gaps visibly — before the investigation is closed and before anyone external has a chance to find them.
This is related to the evidence collection problem described in A General AI Tool Doesn't Know What Evidence to Collect — but it's a distinct layer. Evidence collection happens at the start. Completeness checking happens after evidence is in, as a quality gate before conclusions are finalised.
Together, these two layers mean that an investigation can only be closed when there's an explicit, documented basis for the conclusions it draws.
If a regulator or insurer asks to see your investigation and finds that key evidence was never collected — and that the report made no mention of this — the gap works against you. A completeness check is how you close that loop before someone else does.