Glossary & Notecards

All 47 key terms from the course. Drill them as shuffled notecards (by module or the whole deck) or browse the full reference below. Terms you mark “knew it” stay marked across visits.

Notecards

All modules · 47 terms

0 of 47 marked known

Card 1 of 470 knew it · 0 to review

Browse the Full Glossary

Week 0 · Finance AI Foundation18 terms

Adversarial check: Having a fresh AI instance review a draft as an external skeptic would, checking every factual claim cold. Its findings become the human-review agenda.
Agentic tool: An AI tool that works across a whole folder or environment rather than a single chat message, ingesting files and producing finished deliverables.
Assumptions ledger: A written record of every assumption behind a number, so a reviewer can trace and challenge each one rather than accepting a total on faith.
Deterministic split: The habit of separating a workflow's deterministic parts from its non-deterministic parts, so AI builds the machinery for the former and performs only the latter.
Deterministic work: Work with one correct output for a given input, such as an amortization schedule or a reconciliation. The reliable move is to have AI build the formula or code, which then runs deterministically.
Folder as context: Briefing an agentic tool with a clean folder: the inputs you normally receive, one example of the output you expect, and nothing else.
Hallucination: A confident but false output, such as an invented citation or figure. It is the reason every source and number is verified before anything ships.
Human review checklist: The five-point close on any AI-assisted deliverable: verify every source, own every conclusion, strip AI writing patterns, trace every number, and reconcile.
Jagged frontier: The uneven boundary of a model's competence. Inside it, output is strong; just beyond it, output can be confidently wrong, so an expert still chooses the destination.
Minimum context principle: Give a tool exactly what the task needs and leave the rest out. Oversupplying context tends to lower quality and raises data-governance risk.
Model tier: The capability level of a model, from an economy tier for high-volume well-defined tasks to a premium tier for complex, ambiguous, or judgment-heavy work.
Multi-model orchestration: Having models from different vendors review each other's work, so triangulation surfaces blind spots that no single model reveals about itself.
Non-deterministic work: Work with many acceptable outputs, such as a memo, a summary, or a narrative. Language models fit this well because they are non-deterministic by construction.
Prompt injection: A security risk in which untrusted content in a file or page manipulates an agent into unintended actions. Curating and scoping the folder is a defense.
Red lines: The bright-line limits on AI use: what data cannot go in, which tool tier is allowed, what needs escalation, and what the audit trail must capture.
Red-lines check: A short gate run before a workflow starts, confirming the right data is going into the right tool with scoped access and approvals in place.
Skill: A packaged set of prompts, instructions, and connectors that encodes a repeatable workflow so an expert process can be run on demand.
Three-anchor journey: Framing a task as where you are (current materials), where you are going (a concrete end state, ideally a worked example), and how to get there (process, tool, and model).

Week 1 · Close & Reporting6 terms

Close-ready: Output finished to the standard a controller would send onward: reconciled, materiality-screened, correctly toned, and free of invented drivers.
Flux commentary: The written explanation of why financial line items changed period over period, expected to name real drivers and tie to the underlying numbers.
Materiality threshold: The size below which a variance is not worth explaining. It keeps commentary focused on the drivers that matter rather than on immaterial noise.
Tie-out: Confirming that every figure in a narrative traces back to a source number, so the commentary reconciles rather than merely sounds plausible.
Trial balance: A listing of every general-ledger account with its balance, where total debits equal total credits. The starting input for close and flux work.
Variance: The difference between two figures, such as current period versus prior period or actual versus budget, usually shown in both dollars and percent.

Week 2 · Variance & Driver Trees7 terms

Alternative hypothesis: A competing explanation for an observed result, offered alongside the primary one to guard against a confidently wrong single story.
Decision-quality layer: The step that turns an explanation into a recommendation by adding alternative hypotheses and the evidence that would distinguish them.
Driver tree: A structured breakdown of a result into the underlying drivers that moved it, such as splitting a revenue change into price, volume, mix, and cost effects.
Falsifiable: Framed so that specific evidence could confirm or rule it out. A hypothesis you cannot test is not decision-useful.
Mix effect: The part of a change caused by a shift in the proportion of higher- or lower-value items sold, separate from total volume.
Price variance: The part of a revenue change caused by a change in price, measured as the price change times the actual volume.
Volume variance: The part of a revenue change caused by a change in units sold, measured as the volume change times the budgeted price.

Week 3 · Research & Benchmarking5 terms

Comparability caveat: A flag that a metric is not directly comparable across companies because of a definition or disclosure difference, so it should be noted rather than estimated.
Extract-verify-cite: A durable research pattern: extract figures from sources, verify each one resolves, and cite it, so no number is unsupported.
Fact versus inference: The discipline of tagging every claim as either a fact (traceable to a cited source) or an inference (reasoned, and labeled as such).
Peer set: The group of comparable companies chosen as the basis for a benchmark, selected for similarity in business model, size, or sector.
R&D intensity: Research and development expense divided by revenue, a benchmark of how much of the top line is reinvested in development.

Week 6 · Anomaly Detection6 terms

Audit trail: A record of what was tested, the thresholds used, the findings, and the disposition of each, so a review can be reperformed.
Disposition: The proposed next step for a flagged entry, such as investigate, reclassify, or confirm with the vendor.
Duplicate payment: The same vendor and amount appearing more than once, a common accounts-payable error that exception testing flags.
Exception-based testing: Reviewing a data set by applying rules that surface only the entries needing attention, rather than reading every line.
False positive: A clean entry that a test flags incorrectly. Minimizing false positives keeps an exception report trustworthy.
Round-dollar entry: An amount posted in exact whole thousands, which can signal an estimate or a manual entry worth reviewing.

Week 8 · Contract Review5 terms

Clause extraction: Pulling the specific contract clauses that drive an accounting outcome, quoting the language verbatim rather than paraphrasing it.
Confidence rating: A high, medium, or low tag on an extracted implication, signaling to the accounting team where judgment is most needed.
Issue, not conclusion: Framing an accounting implication as a question to resolve rather than a settled treatment, so a first-pass read does not overstep the accounting team's judgment.
Performance obligation: Under ASC 606, a distinct promise in a contract to transfer a good or service, which drives how and when revenue is recognized.
Variable consideration: Contract payment that is not fixed, such as usage fees, royalties, rebates, or refunds, which must be estimated and constrained under ASC 606.

Back to all modules