AI & Automation
Also known as
Citation · Attribution
Definition
Related terms
Hallucination
A model producing fluent, plausible content unsupported by any source. The specific danger in finance operations is that hallucinated output is indistinguishable from correct output at review speed, which is why grounding and abstention are structural mitigations rather than optional features.
Audit Trail
The complete, tamper-evident record of what happened to a transaction: who or what acted, when, on what input, under what rule, and what changed. For automated and agentic finance processes this is the load-bearing artifact — it is what makes the process auditable, and in outcome-based commercial models it is the contractual source of truth for attribution.
Retrieval-Augmented Generation (RAG)
Grounding a language model's output in documents retrieved at query time rather than relying on parameters alone. In finance operations, retrieval quality is usually the binding constraint on answer quality — the model rarely fails because it cannot reason, but because it was handed the wrong document.
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.