AI Earnings Release Automation: Non-GAAP Reconciliation Practitioner Walkthrough

If your team spends the first week after quarter-end pulling trial balance data, computing adjustments, and cross-checking reconciliation tables line by line, you're not alone. For most mid-to-large finance teams, the non-GAAP reconciliation is the most manual, error-prone, and legally exposed section of the entire earnings release. AI can change that workflow materially. But the same automation that cuts your timeline from 12 days to 7 can also generate a reconciliation table that violates Regulation G before anyone notices.

This walkthrough maps every step of the non-GAAP reconciliation process to its actual AI risk level, explains the SEC rules the AI must get right, and gives you the governance controls to deploy safely.

Key takeaway: AI can safely automate the mechanical steps of non-GAAP reconciliation. It cannot safely make the judgment calls. The line between those two categories is where most teams get into trouble.

What the SEC Actually Requires in a Non-GAAP Reconciliation

The rules are specific, and the AI output must satisfy them exactly. Regulation G (17 CFR 244.100) requires that any public disclosure of a non-GAAP financial measure include a presentation of the most directly comparable GAAP measure and a quantitative reconciliation to it. This applies to earnings releases furnished on Form 8-K, investor presentations, and any other public disclosure. There is no safe harbor for automation errors.

Item 10(e) of Regulation S-K adds three more requirements for SEC filings specifically:

• Disclose why management believes the non-GAAP measure provides useful information to investors.
• Disclose how management uses the measure internally.
• Never give the non-GAAP measure greater prominence than the comparable GAAP measure.

That third point, the prominence rule, is one of the most frequently violated requirements in practice, and it is particularly easy to violate when AI auto-formats a document. If an AI-generated earnings release places Adjusted EPS before GAAP EPS in the headline table, the company has a problem before the press release goes out.

The SEC's Division of Corporation Finance updated its Compliance and Disclosure Interpretations (C&DIs) on non-GAAP measures in May 2016, tightening guidance across six categories: prominence, individually tailored accounting principles, per-share non-GAAP measures, income tax effects, non-recurring labels, and operating expense exclusions. A December 2022 C&DI update (Questions 100.01 through 100.06 revised) added a specific warning: labeling an adjustment as "non-recurring" when similar charges have appeared in prior periods is misleading and violates Reg G. That 2022 update is critical for AI deployments, as explained below.

Non-GAAP comments have consistently ranked among the top SEC comment letter topics for S&P 500 companies, with prominence violations and incomplete reconciliations being the most common deficiencies, according to Audit Analytics data.

The Non-GAAP Reconciliation Workflow: Step-by-Step AI Risk Map

Here is the full workflow, with an honest assessment of where AI helps and where it creates risk.

Step 1: Data Ingestion and Roll-Forward (Low Risk, High Value)

This is where AI delivers the most unambiguous time savings. Purpose-built platforms like Workiva have integrated AI-assisted roll-forward of non-GAAP tables from prior quarters, pulling figures directly from ERP systems and automatically populating the current-period reconciliation structure. The mechanical work that once took a senior analyst two days can run overnight.

Early AI adopters have compressed the post-quarter-end earnings release timeline from an average of 10 to 14 days to 6 to 9 days, according to IR advisory firm data published in 2025. Critically, the time savings are concentrated here, in data ingestion and first-draft generation. Human review time has not compressed proportionally.

What to check: verify that the AI has pulled from the correct GL period, that intercompany eliminations are applied before the data enters the reconciliation, and that the opening balance ties to the prior period's filed reconciliation.

Step 2: Arithmetic Validation and Anomaly Detection (Low Risk)

AI excels at cross-checking that the sum of adjustments equals the difference between the GAAP and non-GAAP figures, that per-share calculations use the correct share count, and that subtotals roll up correctly. These are deterministic checks with no judgment involved.

Platforms with anomaly detection can also flag when a line item deviates materially from the prior-quarter trend, prompting the team to investigate before the release goes out rather than after. This is genuinely useful and carries low regulatory risk.

Step 3: Selecting the Most Directly Comparable GAAP Measure (High Risk)

This is a judgment call the AI cannot make reliably, and getting it wrong is a documented SEC comment trigger. Deloitte's DART non-GAAP roadmap notes that the SEC has questioned registrants when multiple adjustments are needed to reconcile to a GAAP income measure, but only one or two adjustments would reconcile to a cash flow measure. In those cases, the SEC may challenge whether the registrant chose the right GAAP comparator.

An AI system that auto-selects the GAAP comparator based on the company's historical practice could perpetuate a wrong choice that was never challenged. The controller or VP of Financial Reporting must own this decision explicitly, with documentation of the rationale.

Key takeaway: "A registrant must reconcile a non-GAAP measure to the most directly comparable GAAP measure. Such reconciliation should be quantitative and is generally expected to include each significant reconciling item." Deloitte DART, Section 3.2.

Step 4: Adjustment Classification (High Risk, Non-Negotiable Human Review)

This is the single highest-risk step for AI automation. The December 2022 C&DI update is explicit: if a company labels a restructuring charge as "non-recurring" in the AI-drafted narrative, but similar charges appeared in the prior two or three years, the disclosure is deficient under Reg G.

The problem is structural. An LLM trained on the company's historical earnings releases will learn to describe restructuring charges the way the company has always described them. If the company has historically used aggressive non-recurring language, the AI will reproduce it. As the Harvard Law School Forum on Corporate Governance noted in 2024, the risk of AI-assisted errors in non-GAAP disclosures is asymmetric: efficiency gains are incremental, but a single material error in a reconciliation table can trigger enforcement, restatement, and securities litigation.

KPMG's 2024 Non-GAAP Handbook is direct on this point: AI-assisted classification of adjustments as recurring versus non-recurring requires human judgment and cannot be fully delegated to automated systems given the SEC's fact-specific, pattern-based analysis.

Practical control: build a look-back check into your review process. Before finalizing any adjustment labeled non-recurring, require the preparer to confirm in writing that the charge did not appear in the prior four quarters. The AI can surface the prior-period data; the human must make the call.

Step 5: Income Tax Effect Calculation (High Risk, AI-Assisted)

The income tax effect of non-GAAP adjustments must be calculated and disclosed. The SEC has objected to companies that apply a single statutory rate to all adjustments without explaining why that rate is appropriate, per EY's 2024 Financial Reporting Developments series.

For a multi-jurisdiction company with complex tax structures, this is a computationally intensive step where AI can genuinely help, running the adjustment through the applicable jurisdictional rates and producing a draft tax-effect schedule. But the output requires review by someone with ASC 740 expertise. The AI cannot assess whether a deferred tax asset valuation allowance changes the effective rate, or whether a particular adjustment is deductible in the relevant jurisdiction. For a deeper look at AI in the ASC 740 tax provision workflow, see AI Tax Provision ASC 740 Reporting: A 2026 Practitioner Walkthrough.

Step 6: Narrative Drafting (High Risk, Legal/IR Sign-Off Required)

Item 10(e) requires the company to explain why management believes the non-GAAP measure provides useful information and how management uses it. As of early 2026, several Fortune 500 companies are using LLM-based tools, including Microsoft Copilot for Finance and purpose-built tools from Workiva, to draft this language, with legal and IR teams reviewing and editing, according to Bloomberg Tax reporting.

The risk flagged by Bloomberg Tax is real: AI-drafted boilerplate that is not tailored to the company's actual use of the measure can be flagged by the SEC as inadequate disclosure. The "management uses this measure to evaluate segment performance" language that appears in thousands of earnings releases is exactly the kind of generic output an LLM produces. It may satisfy the letter of Item 10(e) or it may not, depending on how the company actually uses the measure.

For a full framework on reviewing AI-drafted financial narrative before it becomes public record, see Reviewing AI-Drafted Financial Commentary: A CFO's Process Guide.

Step 7: Prominence Check (Medium Risk, Mandatory Before Publication)

The prominence rule is violated more often in AI-formatted documents than in manually formatted ones, because layout is auto-generated and the AI has no inherent understanding of regulatory sequencing requirements. A formatted earnings release where Adjusted EBITDA appears in the headline table before GAAP operating income is a comment letter waiting to happen.

Build a specific prominence checklist into your pre-publication review:

1. Does any non-GAAP measure appear in the headline or opening paragraph before its GAAP equivalent?
2. Is the reconciliation table placed in a location of equal or lesser prominence than the GAAP financial statements?
3. Does the document title or subject line reference a non-GAAP figure without also referencing the GAAP figure?

PwC's 2024 non-GAAP guide specifically flags non-GAAP EPS appearing before GAAP EPS in earnings releases as a recurring SEC comment trigger, per PwC Viewpoint.

Purpose-Built Platforms vs. General-Purpose LLMs

This distinction matters for vendor decisions. General-purpose LLMs (ChatGPT, Copilot in isolation) have no awareness of Reg G, no access to your prior-period reconciliations, and no mechanism to enforce prominence rules. They can draft narrative and perform arithmetic, but they have no compliance guardrails specific to non-GAAP disclosure.

Purpose-built financial reporting platforms, Workiva being the most widely deployed, embed the reconciliation workflow into a structured environment where prior-period data is linked, cross-references to the 10-Q are automated, and anomaly detection runs against defined thresholds. The AI features in these platforms are constrained by the data model, which reduces (but does not eliminate) the risk of a hallucinated adjustment or a broken cross-reference.

The practical recommendation: use purpose-built platforms for the reconciliation table itself and for cross-referencing to the 10-Q. Use general-purpose LLMs, if at all, only for narrative drafting, with mandatory human review before any output touches a public document. For a broader comparison of AI financial reporting platforms, see AI Financial Reporting Software Comparison 2026: Enterprise Buyer's Framework.

SOX Controls for AI-Assisted Non-GAAP Reconciliation

SOX Section 302 certifications require the CEO and CFO to certify that disclosure controls and procedures are effective. Under SEC Rule 13a-15, the disclosure controls framework must encompass the AI workflow, including data inputs, model outputs, and human review steps. There is currently no SEC guidance specifically addressing AI in disclosure controls, which means companies must design their own frameworks.

Matt Kelly at Radical Compliance has been direct about the consequence of not doing so: "AI tools used in financial reporting must be treated as automated controls under the COSO framework and subjected to the same testing and documentation requirements as any other IT general control. An LLM-generated non-GAAP reconciliation that is not validated by a human with appropriate expertise could constitute a material weakness in ICFR."

A 2025 FEI survey found that 67% of large-cap finance teams were piloting or actively using AI tools in at least one phase of the earnings release process, but only 23% had formally updated their disclosure controls documentation to reflect AI use. That gap is the single largest governance risk in current AI-assisted earnings release workflows.

The PCAOB's 2024 Staff Spotlight on technology in auditing confirmed that auditors must understand and evaluate the controls over any automated tool used in the financial reporting process. Companies without documented AI controls will face harder auditor questions, not easier ones.

Minimum Control Documentation for AI in the Earnings Release

1. Define the AI tool as an automated control in your SOX control inventory, with the control owner, the input data sources, and the output it produces.
2. Document the human review step as a separate manual control, specifying who reviews, what they check, and how they evidence the review (sign-off in the platform, email confirmation, or a formal checklist).
3. Maintain an audit trail of the AI-generated output and the reviewed/approved version, with timestamps and reviewer identity.
4. Test the automated control at least annually, confirming that the AI tool produces accurate output when given known inputs.
5. Update your disclosure controls memo to reflect AI use before the next SOX 302 certification.

For a detailed walkthrough of audit trail requirements for SEC filers using AI, see AI Audit Trail Requirements for SEC Filers: 2026 Practitioner Walkthrough.

The ESG Parallel: ISSB and CSRD Adjusted Metrics

The same reconciliation risks that apply to financial non-GAAP measures are migrating into sustainability reporting, and almost no one is talking about it.

The ISSB's IFRS S1 and S2 standards, effective for annual periods beginning on or after 1 January 2024, require disclosure of sustainability-related financial information, including metrics that companies may present on an adjusted basis. A company disclosing "adjusted Scope 1 emissions excluding acquired assets" faces structurally the same classification and prominence risks as a company disclosing Adjusted EBITDA. The ISSB has flagged that non-GAAP-style adjusted sustainability metrics create comparability and reliability risks analogous to financial non-GAAP measures.

EFRAG's ESRS implementation guidance requires companies to reconcile sustainability metrics to auditable underlying data, a requirement structurally similar to GAAP-to-non-GAAP reconciliation. AI tools being deployed for CSRD compliance reporting must therefore incorporate reconciliation logic, and the same governance controls that apply to financial non-GAAP automation should apply here.

If your ESG team is using AI to produce CSRD or ISSB disclosures with adjusted metrics, they need the same human-in-the-loop review framework your financial reporting team uses for Adjusted EPS. For a full walkthrough of AI in ESG reporting, see AI ESG Reporting Automation: A 2026 Practitioner Walkthrough.

Briefing the Audit Committee on AI in the Earnings Release

Audit committees are asking hard questions about AI use in financial reporting, and most CFOs do not yet have a clean answer. The framework below gives you a starting point.

What to tell the audit committee:

• Which specific steps of the earnings release process use AI tools, and which platform.
• What human review controls wrap each AI-assisted step.
• How the AI workflow is documented in the SOX control inventory.
• What testing has been performed on the automated controls.
• Whether the external auditors have been briefed and what their assessment is.

What not to say: that AI is just a drafting aid and therefore outside the disclosure controls framework. SEC Chair Gary Gensler was explicit in 2024 public remarks: "existing disclosure obligations apply to AI-assisted processes, and companies cannot disclaim responsibility for AI-generated financial disclosures." The audit committee needs to understand that the SOX 302 certification covers the AI output once it enters the earnings release.

For a broader framework on briefing the audit committee on AI, see AI Board Reporting and Audit Committee Oversight in 2026.

Pre-Publication Checklist for AI-Assisted Non-GAAP Reconciliation

Before the earnings release goes out, run through these checks. Each one maps to a known SEC comment trigger or governance gap.

• GAAP measure appears before non-GAAP measure in every table and in the headline (prominence rule).
• Every non-GAAP measure has a quantitative reconciliation to the most directly comparable GAAP measure.
• The GAAP comparator selection is documented and reviewed by the controller.
• Every adjustment labeled non-recurring has been confirmed as not appearing in the prior four quarters.
• The income tax effect of non-GAAP adjustments is calculated at the appropriate jurisdictional rate, not a single blended rate, with the rationale documented.
• The Item 10(e) narrative is tailored to how management actually uses the measure, not generic boilerplate.
• The reconciliation table ties to the 10-Q/10-K GAAP financials on a line-by-line basis.
• The AI tool's output and the reviewed version are both retained in the audit trail.
• The SOX disclosure controls memo reflects AI use in this workflow.
• The audit committee has been briefed on AI use in the earnings release process.

The earnings release is the most analyst-scrutinized document your company publishes each quarter. AI makes the mechanical work faster. The judgment calls, and the legal exposure, remain yours.