Finance AI Should Interpret Metrics, Not Recalculate

Where interpretation stops and calculation stays governed

Finance leaders face a seductive proposition: connect an AI system to actuals and budgets, ask it to explain performance, and receive executive-ready commentary in seconds. The risk is not that the commentary will be poorly written. The risk is that the system will silently recalculate the numbers while writing the sentences.

For CFOs, FP&A directors, and controllers, that distinction is not technical nuance. It is the difference between a governance-ready finance capability and an audit liability dressed in natural language.

The single most important design decision in enterprise finance AI is the boundary between interpretation and recalculation. That boundary must never move.

Inspect one published case in the Finance proof explorer. The full inputs, contract, and downloads stay in the worked example.

The problem: fluent commentary on ungoverned numbers

Finance organizations already operate rigorous processes for how variance is calculated, how materiality is determined, and how management-attention classifications are assigned. These processes exist because financial numbers carry legal, regulatory, and fiduciary weight. They are reviewed, reconciled, and tied to source systems that auditors recognize.

An AI system connected directly to raw ledger extracts or loosely modeled budget tables can produce variance figures that look correct. The model applies arithmetic. It selects columns. It compares periods. The output reads confidently. But nobody can explain — in audit terms — which materiality rule was applied, which budget version was referenced, or why this variance differs from the official management report.

A forecast that is slightly wrong is a planning problem. A financial number silently recalculated by a model, with no visible audit trail, is a governance problem.

Why common approaches fail

Letting the model "validate" finance math. When teams invite the AI layer to check whether variance "looks right," they open the door to silent recalculation. The model may adjust rounding, re-aggregate at a different grain, or apply an inferred materiality threshold that differs from the official rule set.

Conversational access to raw tables. Unconstrained SQL generation against general-ledger tables can change grain, joins, filters, and definitions without passing through the approved close. That is the failure mode to prevent. A validated tool, governed view, or controlled query can still execute a calculation Finance has already approved. The requirement is controlled metric authority: the agent may run approved queries and read approved fields, and it may not invent a formula, grain, or threshold at question time.

Duplicating FP&A logic inside the AI project. When finance AI teams rebuild variance logic in a notebook or prompt because "it was faster," they create a second unofficial close. The AI path diverges from the report the CFO already signed. Reconciliation becomes a manual exercise after every demo.

Treating narrative generation as low risk. Governance reviews often focus on data access while treating generated text as cosmetic. In finance, narrative and number are inseparable. An executive summary that states a material variance figure is a financial output, regardless of whether it appeared in a table or a paragraph.

The governing principle

In Finance Intelligence, the AI layer interprets governed outputs. It does not:

• recalculate variance,
• redefine materiality,
• invent a KPI,
• override the source of truth, or
• generate unsupported financial values.

This is not a limitation to work around. It is the reason finance teams can trust the output at all.

Governed fields — actual amount, budget amount, variance amount, variance percentage, material flag, variance category, management-attention classification — are calculated deterministically before the AI layer produces a single sentence of commentary. The model explains what the numbers mean for decision-makers; it does not become an alternate calculation engine.

That principle aligns with the broader Orisdale reference architecture and the constraints applied to every governed finance agent.

Implementation architecture

Finance intelligence follows a strict top-to-bottom sequence:

Actuals and Approved Budgets
        ↓
Governed Finance Logic
        ↓
Semantic Finance Intelligence
        ↓
Enterprise Data Platform
        ↓
Agentic AI Runtime
        ↓
Executive Finance Experience

Each layer has a distinct responsibility:

Source reconciliation — actuals tied to the approved close; budgets tied to the sanctioned planning version, not ad hoc extracts.

Governed finance logic — variance calculations, threshold rules, materiality flags, and management-attention classifications expressed as versioned rules that Finance can test against the official report.

Semantic finance intelligence — business context that tells the AI layer what a cost center represents, how departments roll up, and which comparisons are valid at executive grain.

Platform publication — governed outputs exposed inside Snowflake, Databricks, or Google Cloud with the same access controls Finance already operates.

Constrained AI runtime — Snowflake Cortex, Databricks Genie Agents, or Gemini Enterprise Agent Platform experiences that read governed fields and produce narrative, ranked exceptions, and conversational answers — without write access to calculation logic.

A practical example: the CFO's Monday review

A CFO opens a finance intelligence experience and asks: "What are the largest material variances this period, and which departments require management attention?"

In a recalculation architecture, the model queries source tables, computes variance on the fly, applies an inferred materiality concept, and generates a ranked list. The commentary is fluent. The numbers may not match the package the controller prepared for the leadership meeting.

In an interpretation architecture, material variances and management-attention classifications were computed during the governed close process — using the same rules the official report uses. The AI layer reads those fields, ranks them, explains primary drivers in plain language, and identifies what changed from the prior period. When the CFO compares the AI summary to the controller's report, the figures align because they were never separate paths.

Matching the official variance amount does not make the commentary correct. A model can repeat a governed number and still invent a cause, misstate the sign, or describe a hypothesis as a fact. Permitted behavior is to read the governed fields, rank them with the approved rule, and keep observed drivers separate from labeled hypotheses. Review has to check both the returned numbers and the claims made about them.

Sample questions this architecture supports include: "What changed from the previous period?", "What are the primary unfavorable drivers?", and "What should the CFO review first?" Each answer is commentary on governed outputs, not a parallel close.

Worked example

The boundary is easier to judge on numbers you can check. The example below uses three invented departments and a small, versioned rule set. A deterministic calculation produced the results, and every input, the rule contract, the calculation source, the results, and the validation record are published so a reader can reproduce them. It shows what a governed calculation establishes, and what it leaves for a person to review. The same boundary sits inside the reference architecture.

Controls and review boundaries

Finance AI requires tighter boundaries than most enterprise use cases:

• Immutable calculation layer — AI agents have read access to governed outputs, not mutate access to rules or source transactions.
• Version pinning — commentary references a specific budget version and close period; retroactive restatements trigger re-generation, not silent updates.
• Scope disclosure — responses identify the reporting period, currency, consolidation level, and any exclusions applied in the governed logic.
• Escalation paths — when data is missing or a rule fails, the system reports the gap rather than inferring a substitute value.

Governance boundary: If an AI-generated variance figure disagrees with the official management report, the AI output is wrong by definition — regardless of how reasonable the model's alternative calculation appears. Investigation belongs in the logic layer, not in prompt tuning.

Human review remains in place for consequential finance decisions. AI commentary accelerates review; it does not replace sign-off. Monitoring covers access patterns, output quality, and any attempt by agents to invoke unauthorized calculation tools — as described in the Architecture monitoring boundary.

Platform considerations

Finance data typically already lives on an enterprise platform with mature access controls. Finance AI should extend that boundary, not exit it.

Snowflake secure views and Cortex experiences can serve governed finance fields without exporting sensitive ledger detail to external systems. Databricks Unity Catalog governs both pipeline logic and the datasets agents consume. Google Cloud BigQuery governed datasets and Gemini Enterprise Agent Platform controls keep finance outputs inside existing IAM policies.

The platform choice matters less than the architectural invariant: governed logic executes inside the platform's production boundary; AI reads the result. Alteryx One can support upstream preparation when actuals and budgets require harmonization before publication, but the authoritative variance logic should not exist only in a preparation sandbox disconnected from the platform Finance trusts.

Questions leaders should ask

Before approving finance AI for executive consumption, CFOs and controllers should require answers to:

1. Which governed fields does the AI read — and can we list them alongside their rule definitions?
2. Can the AI system modify, re-derive, or override any of those fields through tools or generated SQL?
3. Will AI-generated variance figures match the official management report for the same period — and how do we verify that?
4. What budget version and close period does each response reference, and what happens when a restatement occurs?
5. Who owns the materiality rules — Finance policy or the AI project team?
6. What audit trail exists for AI-generated commentary shown to leadership?

If question three cannot be answered with a deterministic yes, the program is not ready for executive finance review — regardless of narrative quality.

Next step

See this boundary applied in the Finance Intelligence solution — including sample executive questions and governed fields — then review how finance logic fits the full stack on the Architecture page. For agent-level constraints, see Enterprise AI Agents, Governed by Design.

Related reading: What an Enterprise AI Agent Needs in a Business Logic Contract uses the worked example on this page as the calculation an agent is allowed to read. Human-in-the-Loop AI is the review design for the packet that calculation produces.