[ Finance ]
Is Your Finance Data Ready for AI? A 12-Point Checklist
Use this 12-point checklist to test whether finance data is AI-ready across source scope, period logic, mappings, definitions, lineage, access, and ownership.
On this page
Consider a synthetic close-day test: one vendor bill, dated July 2, posted to June. A date filter assigns it to July; the general ledger assigns it to June. Both queries run cleanly. Only one answers the close question.
Finance data is ready for AI when a workflow can name its source, posting period, entity, currency, definitions, mappings, tests, permissions, and approver, then stop when one is missing. This checklist tests those conditions against a single deliverable.
What does AI-ready finance data actually mean?
AI-ready finance data is fit for a named financial task, not immaculate in the abstract. A ledger extract may be ready for a one-off vendor review and unfit for a consolidated budget-versus-actual report. Readiness depends on whether the data carries enough evidence to answer one question consistently and stop when the evidence fails.
Microsoft Purview's current data-quality guidance tests accuracy, completeness, conformity, consistency, timeliness, and uniqueness. Finance needs one more test: whether each field means what the deliverable assumes. “June” can mean transaction date or posting period; USD can mean local, reporting, or budget currency.
The gap between clean data and financial meaning explains how exact arithmetic produces a bad answer. The control boundary for ChatGPT financial analysis uses a synthetic example in which recognized revenue and booked revenue produce two correct gross-margin calculations that differ by 176 basis points. Arithmetic does not choose the governing definition. Finance does.
NIST AI RMF 1.0 Core asks teams to document task scope, data selection and suitability, knowledge limits, roles, and human oversight. The framework is voluntary and sector-neutral. It does not prescribe a finance checklist. It does support the operating principle behind one: scope and ownership belong in the system before the output reaches a decision.
Which 12 checks should finance run before using AI?
Run the 12 checks against one recurring deliverable, such as June budget versus actuals, a cash forecast, or close commentary. Each check needs evidence, not a confident “yes.” A failed material check should stop the run; a non-material failure needs a documented exception and repair owner. The checklist is an editorial control aid, not an audit standard or a validated scoring instrument.
| Check | Evidence test | What breaks | Pass condition |
|---|---|---|---|
| 1. Account identity | Map every material local account, plus any dimensional split the reporting rule requires, to an approved treatment; tie the mapped trial balance to source. | The same economic account splits, or unrelated accounts combine. | A versioned crosswalk has an owner, explicit split rules, no material orphan mappings, and ties to source. |
| 2. Posted scope | Inventory transaction types and statuses; reconcile the posted-only population to the GL. | Excluded statuses or non-posting records, such as purchase orders in NetSuite, pollute actuals. | A controlled GL-impact rule produces an extract that ties to the authoritative posted balance. |
| 3. Period and calendar | Run the same cutoff by transaction date and posting period; sample late and backdated entries. | Rows move between months and variance drivers point at the wrong period. | The authoritative period field, fiscal calendar, and exception treatment reproduce the requested period. |
| 4. Entity and consolidation | Reconcile the entity register, ownership scope, books, and intercompany elimination status. | Entities disappear, double count, or retain intercompany balances. | Governed membership, counterparty keys, and elimination logic produce consolidated totals that tie. |
| 5. Currency and FX | Reperform sample translations; record local and reporting currency, rate type, period, and version. | Mixed currencies create exact-looking but incomparable values. | Versioned rate tables and policy govern transaction, consolidation, and budget conversion, including when Current, Average, or Historical rates apply. |
| 6. Dimension hygiene | Profile null, invalid, inactive, conflicting, and duplicate cost center, customer, product, and vendor values. | Drivers fragment, collapse, or land in an “other” bucket. | Required dimensions pass approved completeness and validity rules using canonical IDs. |
| 7. Budget-to-ledger mapping | Resolve every budget row to account, entity, period, version, currency, and required dimensions. | Budget versus actuals drops rows or compares different scopes and versions. | The controlled schema and crosswalk leave no material orphans and identify one active plan. |
| 8. Metric definition | Have two operators calculate the metric from its documented definition and compare results. | Revenue, gross margin, ARR, or free cash flow changes meaning between runs. | A metric ID, formula, scope, owner, and effective date produce one governing answer. |
| 9. Relationship keys | Trace a sample customer through contract, invoice, GL entry, and cash using stable IDs. | The workflow guesses joins from names and loses the business relationship. | Canonical identifiers and an explicit relationship map remove material name-only joins. |
| 10. Freshness and completeness | Check load time, row counts, key uniqueness, nulls, duplicates, outliers, and source totals. | A stale or partial population still produces plausible commentary. | Freshness, completeness, uniqueness, and source-total rules meet the task's approved thresholds. |
| 11. Reconciliation and lineage | Replay a closed period; trace sampled output figures to source IDs, logic version, and change history. | Bad joins, signs, mappings, or amendments remain invisible. | Known-answer tests and drill paths reproduce every sampled material figure from source. |
| 12. Access and ownership | Test the service identity against role, entity, row, field, and period restrictions; name the approver. | Sensitive data travels too far, or an unsupported answer reaches distribution. | Least-privilege access and a named finance owner govern scope, exceptions, and final use. |
How do you test source scope and time?
Pick one closed period and ask the source system for the posted population, control totals, entity membership, and fiscal calendar. Preserve the query or report parameters. A screenshot of a total is not a source contract.
ERP semantics matter. Oracle's NetSuite documentation distinguishes posting transactions such as vendor bills and bank deposits from non-posting records such as purchase orders and sales orders. A broad transaction export can therefore include records that do not belong in general-ledger actuals. The right rule varies by ERP, but the test is stable: the approved population must tie to the GL.
Time needs its own contract. NetSuite's reporting guidance states that date filters use transaction date while period filters use posting period. The July 2 invoice in the opening scene may belong in June reporting. A query that filters calendar dates can be technically flawless and financially late.
Multi-entity scope asks which subsidiaries belong, which book governs, what counterparty pairs exist, and whether eliminations ran. Do not ask AI to infer membership from company names. Store the entity key, parent relationship, reporting book, base currency, and elimination state.
How do you test mappings and financial meaning?
Account 4010 is only useful when its meaning is stable within the task. Oracle's chart-of-accounts guidance shows that NetSuite account numbers appear in transactions and financial reports.
Microsoft's Dynamics 365 guidance documents several approaches. For global consolidation from separate local charts, Dynamics 365 requires mapping each local chart to the consolidation company's global chart; it also warns that external mapping is prone to error when financial reports are created. The checklist makes a narrower point: the chosen identity and mapping rule must be explicit.
A grain mismatch duplicates money. Microsoft's Power BI star-schema guidance says fact tables should load at a consistent grain. A monthly budget by department cannot be joined directly to invoice-level actuals by vendor and day without choosing an aggregation rule. If the workflow mixes grains, it can duplicate values or imply unsupported precision.
NetSuite's budget model shows that, depending on enabled features, budgets can be keyed by account, period, subsidiary, book, fiscal year, category, currency, customer or project, item, class, department, location, and custom segment. The exact fields differ by system. Every budget row must resolve to the actuals grain and one active plan version.
“Gross margin” needs a formula, scope, currency basis, exclusions, owner, and effective date. Hallucinated inputs are one failure. Ambiguous business meaning is another: a system can calculate exactly what it was told and still answer the wrong question.
How do you test comparability, controls, and ownership?
Test comparability with a versioned currency policy, controls by forcing known failures, and ownership by verifying least-privilege access plus a named approver. NetSuite documents currency-exchange rates for transactions, consolidated exchange rates for actuals, and, when the relevant features are enabled, budget exchange rates. Its consolidated exchange-rate documentation records Current, Average, and Historical rate types by accounting period and subsidiary pair. One unlabeled `fx_rate` column is not enough evidence for a consolidated result.
Then break the workflow on purpose. Remove a mapping. Duplicate a transaction. Delay a source load. Change a sign. Revoke access. The run should fail visibly at the right boundary.
PCAOB AS 1105 says that when an auditor uses company-produced information as audit evidence, the auditor must test its accuracy and completeness or test the controls over accuracy and completeness, and assess whether the information is sufficiently precise and detailed for the audit. AS 1105 does not automatically govern an internal management report that is not used as audit evidence. Whether or not AS 1105 applies, another operator should be able to test the population, method, and detail.
Access controls must survive the AI layer. A finance analyst who cannot open payroll in the ERP should not gain payroll access by asking a model. Test the AI service identity, retrieval layer, exported files, and output surface against the same role and entity boundaries. Log the access decision with the run.
What can AI not repair, and what must stay human?
AI should not own approval of which revenue definition governs, whether an unresolved reconciliation is accepted, which FX policy applies to equity, or whether a variance is material to the board. A controlled system can retrieve recorded policy, flag missing evidence, route exceptions, and draft language. Finance remains accountable for the choice.
Human review also fails when the reviewer sees only polished prose. A finance owner needs the source population, mapping version, formula, tests, exceptions, and changes since the last run. Otherwise the approval is theater with a timestamp.
A finance AI pilot stalls when data, logic, context, controls, or ownership are missing. The checklist turns that diagnosis into twelve places where the workflow either produces evidence or stops.
How should a finance team use the checklist this week?
Choose one recurring output due within 30 days. Name its finance owner. Run the 12 checks against the last closed period and attach evidence to each answer. Mark a check green only when the evidence is documented and reproducible, yellow when a named operator can supply it manually, and red when the scope or meaning remains disputed.
Do not average the colors into a readiness score. One red item can invalidate the entire output. A missing product dimension may be tolerable for a cash report and fatal for a product-margin analysis. The owner decides materiality by use case, records the exception, and sets the stop rule before the next run.
Start with a one-page source contract: business question, authoritative systems, posted-status rule, period field, entity scope, currency policy, mappings, metric definition, tests, exception path, and approver. Centralizing every byte can wait. The contract cannot.
How does Pluvo make finance data ready for repeatable AI work?
Pluvo stores finance definitions, source relationships, and controls outside the language model. Its ontology connects a question to approved entities and definitions. Lineage keeps the source records and transformations. Controls apply scoped permissions and audit logging to people and agents and can require human approval before a sensitive report or workflow proceeds.
The 12-point checklist is Pluvo's synthesis of its published pilot diagnostic, number-handling contract, data-fluency scorecard, and source-contract guidance. It is a control aid, not an audit standard or certification.
For one practical finance-AI control pattern each week, subscribe to the AI Finance Playbook.
The July 2 invoice may still belong in June. An AI-ready workflow does not guess. It points to the field that proves it.
Frequently asked questions
What is AI-ready finance data?
AI-ready finance data gives a named workflow authoritative sources, accounting meaning, reproducible tests, permissions, lineage, and a human approver. The workflow can reproduce its answer and stop when required evidence fails.
Does finance data need to be centralized before using AI?
Not necessarily. A recurring AI workflow needs explicit source authority, stable identifiers, mappings, definitions, relationships, tests, access rules, and ownership. Those contracts can span approved systems while a longer-term data-platform program continues.
How is AI-ready finance data different from clean data?
Clean data passes format and quality checks. AI-ready finance data also encodes accounting meaning, lineage, permissions, and approval.
Should finance teams calculate one readiness score?
Usually not. One material failure can invalidate an entire output, while the same defect may be immaterial for another task. Evaluate each check against the deliverable, record evidence and exceptions, and define stop rules before the run.
Which finance data issues should stop an AI workflow?
Stop when material source totals do not reconcile, scope or definitions are disputed, mappings are orphaned, period or currency treatment is unknown, required data is stale or incomplete, access is unauthorized, or no named owner can approve the result.
What finance decisions must remain human-owned?
Under this checklist, finance professionals remain accountable for metric definitions, materiality, accounting and FX policy, unresolved reconciliations, exception acceptance, recommendations, and final distribution. A controlled AI workflow can assemble evidence and draft explanations after those decisions are recorded.
