AI & Automation
Abbreviation
HITL
Also known as
Human review
Definition
Related terms
Abstention
A system's ability to decline to answer when it lacks sufficient grounds, rather than producing its best guess. In high-stakes finance work, a well-calibrated system that abstains reliably outperforms a nominally more accurate one that always answers, because the cost of a confident wrong answer is not symmetric with the cost of a flagged one.
Exception Queue
The holding area for transactions that could not be processed automatically, awaiting human resolution. Queue design — prioritization, evidence pre-assembly, routing to the right skill — determines whether exceptions are cleared in minutes or days, and is where most of the realized value of automation actually sits.
Confidence Score
A model's self-assessed reliability for a specific output, used to route work between automatic processing and human review. The score is only useful if it is calibrated: a system confident at 95% must be right about 95% of the time at that level, or the threshold is a fiction.
More in
AI & Automation
Abstention
A system's ability to decline to answer when it lacks sufficient grounds, rather than producing its best guess. In high-stakes finance work, a well-calibrated system that abstains reliably outperforms a nominally more accurate one that always answers, because the cost of a confident wrong answer is not symmetric with the cost of a flagged one.
Agentic AI
Software that pursues a goal across multiple steps — reading context, calling systems, making decisions, and acting — rather than producing a single output on request. In finance operations the distinction that matters is not autonomy but accountability: an agent that acts on the ledger must produce evidence a reviewer can follow.
Agentic Search
Retrieval that plans and executes several dependent queries to answer a question no single lookup can resolve — for example tracing a disputed charge across a contract, a purchase order, three invoices, and a remittance file. Evaluated on end-to-end answer correctness rather than on document relevance.
Confidence Score
A model's self-assessed reliability for a specific output, used to route work between automatic processing and human review. The score is only useful if it is calibrated: a system confident at 95% must be right about 95% of the time at that level, or the threshold is a fiction.
Exception Queue
The holding area for transactions that could not be processed automatically, awaiting human resolution. Queue design — prioritization, evidence pre-assembly, routing to the right skill — determines whether exceptions are cleared in minutes or days, and is where most of the realized value of automation actually sits.
Ground Truth
A verified, human-adjudicated set of correct answers used to measure system performance. Without a maintained ground-truth set, accuracy claims cannot be checked, and the absence of one is the most reliable tell that a vendor benchmark is not reproducible.
Low-risk start
Watch it run on your real documents and your own systems, before you commit to anything.