SBL Insights
The blog
-
Why AI Pilots Fail When Nobody Designs the Operating Model
Most failed AI pilots do not fail because the model is weak. They fail because nobody designed the handoff, exception path, audit…
-
The Real QA Problem in Retail Computer Vision Annotation
Retail vision models do not break only on bad models. They break on taxonomy drift, label ambiguity and weak reviewer loops.
-
How Herbarium Specimens Become Machine-Readable Data
A scanned herbarium sheet is not yet a scientific asset. The real work is preparing specimen context and taxonomy records.
-
Why Legislative AI Needs Record Governance Before Summarisation
AI can summarise debates and bills. But in a legislature, the record is a legal object — public memory in perpetuity.
-
Taxonomy Drift Is a Data Governance Problem, Not a Labelling One
When the label set changes faster than the guideline, every downstream metric quietly stops meaning what it used to mean.
-
What an Audit Trail Actually Has to Record
“We log everything” is not an audit trail. An audit trail answers a specific question: why did this system produce this result,…
-
Digital Twins Are Only as Good as Their Asset Register
A twin that models assets the register does not know about is a simulation, not an operational tool.
-
Public Platforms Need an Exception Path Before They Need a Model
In a public service, the interesting cases are the ones automation cannot close. Design for those first.
-
Annotation Guidelines Are a Product, Not a Document
Guidelines that are written once and never revised become folklore. The reviewer loop is what keeps them true.