Financial Consolidation with AI: What to Automate, What to Control and How to Audit It

Consolidation is, first and foremost, a data problem governed by strict rules

Anyone who has been through a consolidated close knows where the hours go: not into eliminations, but into chasing data. Artificial intelligence can significantly reduce that time and the risk of error, but only if it is applied to the right tasks and within a control framework that an auditor can review.

Consider a group with subsidiaries in several countries. Each month, it receives trial balances using different charts of accounts, currencies and closing calendars, and with widely varying levels of data quality. Before applying a single consolidation rule, the finance team spends a substantial part of the close standardising data, reconciling balances and chasing discrepancies. This preparation work—not the accounting elimination itself—is where the hours and manual adjustments accumulate and, consequently, where errors arise.

That is why consolidation is particularly well suited to AI: it involves large volumes of semi-structured data, recurring patterns from one period to the next, and an end result that can be checked. It also explains why this area requires greater rigour than others: consolidated financial statements are prepared under the responsibility of the directors, audited, and used as the basis for decisions by investors, lenders and regulators. A model that “gets it right almost every time” is not good enough.

The regulatory framework cannot be delegated to a model

AI does not change the consolidation rules; it changes how the data to which those rules are applied is prepared and verified. It is important to keep in mind which standards govern each decision, because none of them permits a probabilistic approach:

Area	Reference	Requirement
Scope and control	IFRS 10; Articles 42 et seq. of the Spanish Commercial Code; NOFCAC (Royal Decree 1159/2010)	Determine which entities must be consolidated and which method applies, exercising judgement on control.
Business combinations	IFRS 3	Fair value of acquired assets and liabilities, goodwill and non-controlling interests.
Foreign currency translation	IAS 21	Functional currency, closing and average rates, and translation differences recognised in equity.
Intragroup transactions	IFRS 10 (B86); NOFCAC	Elimination of balances, transactions and unrealised profits.
Related parties and disclosures	IAS 24; IFRS 12	Information about interests in other entities and related-party transactions.
Presentation	IAS 1; IFRS 18 from 2027	Structure of the financial statements and new subtotals in the income statement.

 

This leads to a design rule: decisions governed by accounting standards—consolidation scope, method, functional currency and elimination criteria—must be configured and documented as explicit, approved rules. AI can suggest, detect and explain, but the calculations that feed into the consolidated financial statements must be deterministic, reproducible and traceable back to the source data.

Where AI adds value in the consolidation cycle

The greatest returns come in the stages before eliminations: data capture, standardisation and reconciliation. These are high-volume tasks with stable patterns, where human error can silently propagate into the consolidated figures.

Stage	AI application	Typical technique	Human control
Data capture	Read trial balances and reporting packages in different formats (Excel, PDF and ERP exports).	Extraction using language models, validated by checking totals.	Review extractions with low confidence scores.
Account mapping	Suggest mappings between each subsidiary’s local chart of accounts and the group chart.	Supervised classification using historical mappings and account descriptions.	Approve every new mapping before it goes into production.
Intercompany reconciliation	Match balances and transactions between entities, even when references, dates or amounts do not align exactly.	Fuzzy matching and matching models.	Validate discrepancies above the materiality threshold.
Currency translation	Check the exchange rates applied and detect items translated using an incorrect rate.	Rules and comparison against official exchange-rate sources.	Approve the period’s exchange-rate table.
Adjustments and eliminations	Suggest recurring journal entries and detect eliminations omitted compared with previous periods.	Comparison with historical patterns.	Approve and sign off every consolidation journal entry.
Analytical review	Detect unusual movements by entity, account and period.	Statistical anomaly detection and time-series models.	Document an explanation for every significant alert.
Notes and disclosures	Draft explanatory notes and commentary on movements based on consolidated data.	Language models using verified data as their sole source.	Technical review and approval of the final wording.

 

In practice, intercompany reconciliation is often the best starting point in terms of effort versus results. It is one of the most frequent causes of delays in multinational group closes, and its outcome is binary and verifiable: balances either reconcile or they do not.

One important warning: language models are excellent at reading documents and drafting text, but they should not perform calculations. Any figure appearing in a generated draft must come from the consolidation engine, never from the model itself.

Architecture: a deterministic engine and a supervised AI layer

The architecture that stands up best to an audit separates two functions: AI suggests and the engine calculates. No model-generated proposal should enter the consolidated figures without human approval and explicit accounting rules.

Supervised AI consolidation architecture · 5 layers

Three principles underpin this design:

• Separation of duties. The model does not write directly to the consolidation system. Its outputs are proposals that a responsible person accepts, corrects or rejects.
• Reproducibility. The same data and rules always produce the same consolidated figures, regardless of the version of the AI model used to prepare them.
• Full traceability. Every figure can be traced back to the original balance, the mapping applied, the elimination performed and the person who approved it.

Risks, controls and audit evidence

Introducing AI into the consolidated close adds a new component to the internal control system over financial reporting. It must be treated with the same rigour as any other system that feeds into the annual financial statements.

Specific risks

• Plausible errors. A model may suggest an incorrect mapping or match that looks reasonable. Unlike a broken formula, it may not produce an obvious error.
• Automation bias. When the success rate is high, reviewers may approve outputs without checking them. The control exists on paper but does not operate in practice.
• Model drift. An acquisition, an ERP change at a subsidiary or a new business line can alter data patterns and degrade performance without warning.
• Confidentiality. Group trial balances before publication are sensitive information. Sending them to external services requires assessing where they are processed, who can access them and whether they are reused to train models.
• Opacity. If the reason for a proposed adjustment cannot be explained, it cannot be defended to the auditor either.

Minimum controls

1. Confidence and materiality thresholds. Proposals with low confidence scores or amounts above a defined threshold must always receive explicit human review.
2. Complete logging. Each proposal is recorded with the input data, model version, result, approving user and date.
3. Change management. Rules, mappings and models are version-controlled. A model change must be validated through regression testing on previous closes before being used in production.
4. Task-level performance metrics. Mapping accuracy, the percentage of automatically matched items and false positives in anomaly detection should be measured at every close.
5. Sampling of automatically approved items. Periodically review a sample of proposals accepted without intervention to detect drift and automation bias.

What the auditor will need

Spanish auditing standard NIA-ES 315 (Revised) requires the auditor to understand the entity’s information technology environment and its general IT controls. NIA-ES 500 requires the auditor to evaluate the reliability of information produced by the entity itself. In group audits, NIA-ES 600 adds requirements for coordination with components. In practical terms, the auditor will expect to be able to rerun the consolidation from the source data, inspect the approval trail and verify that AI-related controls operated throughout the year, not only at year-end.

From a regulatory perspective, the EU Artificial Intelligence Act, Regulation (EU) 2024/1689, does not list accounting consolidation among the high-risk uses in Annex III. Nevertheless, the AI literacy obligation for staff operating these systems (Article 4) has applied since 2 February 2025, and a proportionate governance policy is a reasonable way to demonstrate due diligence. For groups with particularly sensitive data, deploying models on their own infrastructure or in private environments is a viable alternative that eliminates much of the confidentiality risk.

How to get started: a phased implementation

Experience shows that projects fail when they start with the model. Successful projects start with the data and controls.

1. Assess the current close. Measure the duration of each stage, identify where manual adjustments are concentrated and determine which intercompany differences recur. Without this baseline, improvement cannot be demonstrated.
2. Establish data governance. Maintain a stable group chart of accounts, documented mapping tables, a closing calendar and named owners for each entity. AI amplifies data quality—both good and bad.
3. Run a limited pilot. Select one high-impact application, usually intercompany reconciliation or account mapping, and run it in parallel with the existing process for at least two closes. Compare the results and adjust the thresholds.
4. Industrialise with controls. Integrate the solution into the production process with the controls described above, an approved AI-use policy, team training and documentation prepared for the auditor. Only then extend it to other stages.

The question that really matters

AI does not consolidate accounts: it prepares, reconciles, detects and explains, allowing the finance team to spend its time on professional judgement rather than mechanics. Real value emerges when a deterministic consolidation engine, explicit accounting rules and a supervised AI layer with a complete audit trail are combined. The question for a group is not whether it should use AI in its close, but whether it can demonstrate to its auditor and directors that it uses AI under control.

At Audalia Auditores, we support business groups in designing consolidation processes and assessing controls over the systems involved in financial reporting, including those based on artificial intelligence. If your group is considering this step, please contact us.