Glossary
AI abstention
AI abstention is a platform's ability to decline to produce a conclusion when the available sources do not sufficiently support one. Instead of generating a plausible guess, the platform reports "no sufficient source." Abstention is what separates a controlled decision platform from a generic tool that responds fluently whether or not it has any basis.
Why is refusing to respond a feature?
Because in high-stakes work, a confident wrong output is the most expensive thing an AI can produce. A generic engine asked about a company’s standards will produce something either way; the output reads identically whether the basis is strong, weak, or absent. Every professional who has caught one of these fabrications afterward trusts the tool less, and every one that slips through costs real money.
Abstention removes that failure mode at the root. If the knowledge library does not support a conclusion, no conclusion is produced. The reader gets a precise, useful fact instead: the company’s knowledge does not cover this, which is often exactly what a decision-maker needs to know before proceeding.
What does abstention require underneath?
It only works when outputs are grounded to begin with. A platform can abstain honestly only if it knows which sources an output would rest on, which is the same machinery that produces source references and calibrated confidence. In that sense abstention is the end of the confidence scale: as support thins, confidence drops, and below sufficiency the platform declines. All three behaviors come from the control half of a knowledge and control layer operating over the company’s decision DNA.
The trust effect
Abstention is also what makes the platform’s confidence believable. When people see the platform say “no sufficient source,” they learn that its high-confidence outputs mean something, and delegation follows. A tool that always responds teaches people to verify everything, which returns the bottleneck to the senior experts it was meant to relieve.
See how Praxiron handles insufficient sources, or read the category pillar: What is a knowledge and control layer?