Analysis
What we set out to understand.
Our perspectives on AI in business: what we see in the field, what we measure, and what the news means for decision-makers. Each piece states what is established and where the evidence stops.
01
An AI that can say it does not know
Never promise zero hallucination. The defensible promise: a system that can assert nothing outside your data, cites its sources, computes its figures, and stays silent below the threshold.
02
A knowledge base in text files
Sending three targeted notes rather than a whole corpus costs 71 times less in tokens and answers better. The explicit link graph, the overhead of the format at identical content, and the limits of the method.
03
Where the errors really come from
On a real prototype, more than half the failures came not from the system but from rules absent from the reference framework.
04
A score that measures memory, not capability
Measured on three models with identical code: 22 points of spread on the construction corpus, 5 points on unseen cases, and a ranking that was entirely wrong.
05
The threshold must filter the answer, not the display
A threshold that filters the sources displayed lets irrelevant material influence the answer while giving the opposite impression.
06
Four failures that raise no error
What holds up an AI project is almost never the model. Four failures we lived through, with no error message: a preview flag, an access hardening that leaves a table mute, a variable read at start-up, a…