FASB and IASB Proposed Guidance on AI Asset Recognition

Finance teams that have spent the past two years capitalizing AI development costs under general intangible asset frameworks are about to encounter a more specific set of questions from their auditors, and the accounting standards that will shape those questions are still being written.

Why AI Asset Recognition Demands a New Framework

Traditional intangible asset accounting borrowed from software development rules was never designed to accommodate systems that learn, adapt, and generate output autonomously. When a finance team capitalizes a conventional software build, the asset's function is fixed at deployment. An AI model, by contrast, continues to change through retraining cycles, fine-tuning events, and data ingestion that may alter its predictive behavior in ways that directly affect its economic utility.

Both the Financial Accounting Standards Board and the International Accounting Standards Board have signaled that existing guidance leaves material ambiguity in this space. The FASB's ongoing agenda on digital assets and software costs, combined with the IASB's broader project on intangible assets, collectively represent the most significant potential shift in asset recognition methodology since the original ASC 350 and IAS 38 frameworks were established. Finance teams that wait for final rules before preparing will find themselves rebuilding their capitalization policies under time pressure.

The question that sits at the center of every controller's agenda right now is precisely this: How does proposed FASB and IASB guidance treat AI asset recognition, and what should finance teams prepare for? The answer requires working through several interrelated layers — what counts as an asset, when recognition begins, how costs are allocated across development phases, and how impairment will be assessed for systems whose utility can degrade quietly as the world around them changes.

The Asset Definition Problem Applied to AI Systems

Under both U.S. GAAP and IFRS, an asset must embody a probable future economic benefit that an entity controls. Control has historically meant legal ownership or contractual exclusivity. For AI systems, control becomes contested in ways that legacy frameworks never anticipated. A model trained on a cloud provider's infrastructure, with weights stored on rented compute, raises genuine questions about whether the organization operating it actually controls the economic resource in any meaningful sense.

The IASB's intangible assets project, which has been active for several years, has pushed toward a more functional definition of control — asking whether the entity has the practical ability to restrict others' access to the benefit. This functional framing matters enormously for AI, because a fine-tuned model running on shared infrastructure might satisfy a functional control test even if the underlying compute is rented. Finance teams should document the specific access restrictions, contractual terms, and technical barriers that support a control argument before their auditors ask.

Identifiability is the second prong of the asset definition, and it creates its own challenges for AI. A model that is deeply integrated with a proprietary data pipeline may not be separable from that pipeline in any practical sense. IASB guidance under IAS 38 requires that an intangible asset either be separable or arise from contractual or legal rights. For systems built on licensed foundation models with proprietary fine-tuning layers on top, the question of which component gives rise to the recognized asset — and whether those layers are separable from the base model — has not been definitively resolved by either board.

How Development Phase Boundaries Apply to Machine Learning Pipelines

The three-phase model that governs software cost capitalization — preliminary project stage, application development stage, and post-implementation stage — maps poorly onto iterative machine learning development. In a classical software project, the transition from design to development is a discrete event. In a machine learning pipeline, the team may cycle between hypothesis formation, data preparation, model architecture selection, training runs, and evaluation in rapid loops that do not have clean boundaries.

The FASB has indicated in its software cost project that it is examining whether the existing three-stage model should be restructured, with some proposals suggesting a shift toward a capitalization trigger based on technological feasibility in a modernized sense. For AI specifically, technological feasibility is hard to define because a model may demonstrate accurate performance in testing environments while failing to generalize to production data distributions. A finance team applying the existing framework today would need to document with specificity the point at which the model architecture was established and training on production-representative data commenced — that moment most closely resembles what the framework intends as the start of the application development stage.

Under IFRS, IAS 38 permits capitalization only when six specific criteria are met simultaneously: technical feasibility, intention to complete, ability to use or sell, probable future economic benefits, adequate resources, and ability to measure expenditure reliably. The reliable measurement criterion becomes particularly thorny for AI projects where compute costs are allocated across multiple experimental runs, many of which do not produce a deployable model. Finance teams should implement cost tracking systems that log each training run with a clear disposition — exploratory versus directed toward a specific production model — before the accounting question is asked retrospectively.

Training Data Costs and the Recognition Threshold

One of the most operationally significant questions in AI accounting is whether the cost of acquiring or preparing training data qualifies for capitalization. Under current frameworks, this question has no clean answer. Data is not land or inventory — it has no physical form, it does not depreciate through use, and its value is almost entirely contingent on how it is used rather than what it is. Yet the labor cost of cleaning, labeling, and curating a training dataset can easily exceed the engineering cost of building the model architecture.

The IASB's intangible assets discussion paper has explored whether internally generated data should be recognized as an asset at all, with preliminary thinking suggesting that the board may eventually develop specific guidance for data assets separate from the intangible asset model. In the interim, most preparers expense data preparation costs unless they can be directly attributed to a development-stage project that itself meets the IAS 38 or ASC 350-40 capitalization criteria. The key documentation requirement is a clear linkage between specific data preparation expenditures and a specific model build that has crossed the technological feasibility or application development threshold.

The labor component of training data preparation is often the most difficult to track. Engineering teams frequently combine exploratory data work with production-directed work within the same sprint, and timesheets that do not distinguish between these activities will not survive an audit of capitalization decisions. Finance teams should work with engineering leadership to establish project coding practices that separate experimental activities from directed development before any retrospective allocation is attempted.

Useful Life Estimation for AI Models

Even when a finance team successfully argues for capitalization, they face the distinct challenge of estimating the useful life of an AI model. Under both ASC 350 and IAS 38, intangible assets with finite useful lives are amortized over those lives, while assets with indefinite useful lives are tested for impairment annually. Almost no AI model has a genuinely indefinite useful life — model drift, data distribution shift, and competitive obsolescence create definite if uncertain lifespans.

The useful life of a model is a function of how quickly the real-world phenomenon it predicts changes, how frequently the organization retrains or fine-tunes the model, and how quickly competitors develop superior alternatives that change user expectations. A credit risk model trained on pre-pandemic economic conditions and never retrained might have degraded materially within eighteen months. A document classification model operating in a stable domain might remain accurate for several years. Finance teams need a defensible methodology for estimating these lifespans that references empirical model performance monitoring data rather than arbitrary conventions.

The Labarna AI piece on modeling depreciation for owned intelligence provides a structured worksheet approach that finance teams can adapt for this purpose, connecting model performance monitoring to amortization schedule review. The critical accounting discipline is to tie useful life reassessment to the operational processes that already monitor model performance, rather than treating amortization as a purely accounting exercise disconnected from the technical team's observations.

Impairment Triggers Specific to AI Assets

Impairment assessment for AI models requires a different set of indicators than the ones finance teams apply to conventional intangible assets. The standard impairment triggers — a significant decline in market value, adverse changes in the business environment, evidence of obsolescence — all apply, but AI models have additional degradation vectors that do not have analogues in traditional asset accounting.

Model drift occurs when the statistical relationship between input features and predicted outputs changes because the real-world distribution of inputs has shifted. A model that was performing well by every measured metric six months ago may be quietly producing biased or inaccurate outputs today without any obvious external event triggering a review. Finance teams need to establish that operational monitoring of model performance constitutes a formal impairment indicator review process, with documented thresholds that, when crossed, trigger a formal assessment under ASC 350 or IAS 36.

Regulatory change represents another impairment trigger that is specific to AI deployments in regulated industries. If a model's outputs are subject to regulatory review and a regulatory body determines that the model's methodology is not acceptable for a particular use, the carrying value of that model may need to be written down even if the model itself continues to function technically. Finance teams working in financial services, healthcare, or any other regulated vertical should include regulatory status as an explicit input to their impairment trigger framework. The Labarna AI article on presenting the AI build case to your audit committee addresses how to frame these risk disclosures for board-level audiences.

The Capitalizable Asset Boundary in Continuous Learning Systems

The recognition boundary becomes most contested when an AI system continues to learn after deployment. Reinforcement learning from human feedback, online learning from production data, and periodic retraining on newly labeled data all produce a model that is incrementally different from the one that was originally capitalized. The accounting question is whether post-deployment learning events represent maintenance expense or a capital improvement to the asset.

The FASB's software cost guidance treats post-implementation activities as expense unless they result in added functionality that was not part of the original specification. Applying this framework to an AI system, a retraining run that improves the model's accuracy on its existing task without changing the scope of the task would likely be expensed. A retraining run that extends the model to handle a new class of inputs or outputs might qualify as a capital improvement. Finance teams should establish this distinction as a formal policy, agreed in advance with their auditors, rather than resolving it case by case as each retraining cycle occurs.

The practical challenge is that many AI development teams do not maintain specifications at the level of granularity required to assess whether a retraining run changed the model's functional scope. Building this specification discipline into the model development lifecycle is simultaneously a governance improvement and an accounting prerequisite. Teams that are thinking carefully about these questions from the beginning of a model build — rather than retrofitting documentation at year-end — will have substantially cleaner audit trails. This connects directly to the broader case for owned infrastructure described in the Labarna AI analysis of the CFO's balance sheet case for owned AI.

Disclosure Requirements Under Proposed and Existing Frameworks

Both the FASB and the IASB are moving toward expanded disclosure requirements for significant intangible assets, and AI models are likely to be among the most scrutinized. Under current IFRS practice, IAS 38 requires disclosure of the useful life or amortization rate, the amortization method, the gross carrying amount, accumulated amortization, and the line items of the financial statements in which amortization is included. For AI assets, auditors are increasingly asking for additional narrative disclosure about how useful life was determined and what monitoring processes support the continued carrying value.

The IASB's intangible assets project has specifically raised the possibility of requiring disclosure about internally generated intangibles that do not meet the recognition threshold — assets that are economically significant but expensed under current rules. If this proposal advances, finance teams would need to quantify and describe AI models that are fully expensed as incurred, providing users of financial statements with information about resources that do not appear on the balance sheet. This is a significant operational change that would require finance teams to maintain tracking for models that are currently outside the formal capitalization process entirely.

Segment-level disclosure is another area of emerging attention. A diversified organization deploying AI models across multiple business units may face questions about whether material AI assets should be disclosed at the segment level rather than aggregated into a single intangible asset line. Finance teams should map their AI asset inventory to their existing segment reporting structure before this becomes a formal requirement, identifying any segments where AI-related carrying values would be material to a reader of segment-level financial data.

Building an Internal AI Asset Registry

The prerequisite for any of the accounting judgments described above is a comprehensive inventory of the organization's AI assets. This inventory needs to capture, at minimum, the model identifier and version, the business function the model serves, the date the model entered production, the costs capitalized versus expensed during development, the estimated useful life and amortization schedule, the date of the most recent retraining or fine-tuning event, and the current performance metrics that inform impairment assessment.

Most organizations do not have this registry in any formal sense. Engineering teams maintain model repositories that track technical versions, and finance teams maintain fixed asset or intangible asset subledgers, but the two systems are rarely connected in a way that allows a controller to trace a capitalized intangible back to the specific model version in the repository. Building this connection is a cross-functional project that requires the cooperation of finance, engineering, and often legal and compliance teams.

The registry should also capture contractual information about the model's dependencies — the foundation model license terms, the data licensing agreements that govern the training data, and any third-party API dependencies that the model's production environment requires. These dependencies affect both the control analysis described earlier and the impairment risk profile, because a change in a foundation model license or a data provider's terms could affect the carrying value of the capitalized asset. Organizations that have made deliberate choices about owned infrastructure versus rented platforms will find this registry significantly easier to construct, because the asset boundaries are cleaner when the organization controls its own compute and model weights.

What Finance Teams Should Prepare Before Final Rules

The actionable preparation agenda for finance teams does not require waiting for final guidance from either board. Several workstreams can begin immediately based on the directions both the FASB and the IASB have already signaled. The first is a current-state inventory of all AI development expenditures over the past three fiscal years, with each project classified by whether it was fully expensed, partially capitalized, or not yet assessed. This inventory provides the baseline for assessing how different proposed rules would affect historical financials and what restatement risk might exist.

The second workstream is a policy refresh. Most organizations' intangible asset capitalization policies reference software cost guidance that predates the modern machine learning context. Updating these policies to explicitly address AI model development — including how the three development phases apply to iterative model builds, how training data costs are treated, and how post-deployment retraining events are evaluated — gives finance teams a defensible framework that auditors can review. The policy should be developed in collaboration with the external auditor before year-end, not presented to them after the fact.

The third workstream is a controls build. The accounting judgments that AI assets require — phase boundary determination, cost allocation between exploratory and directed development, useful life estimation, impairment monitoring — all require information that originates outside the finance function. Building formal information flows between engineering, data science, legal, and finance, with documented evidence that supports each accounting judgment, is the internal controls infrastructure that an audited AI asset recognition process requires.

TFSF Ventures FZ-LLC addresses this infrastructure challenge directly through its 30-day deployment methodology, which builds AI agent systems with documentation architecture designed to support exactly these accounting and audit requirements from the first day of production operation. Rather than retrofitting documentation to satisfy accounting questions, the deployment methodology generates the audit trail as a byproduct of the production operation itself. For finance teams asking whether TFSF Ventures is legit or what TFSF Ventures reviews say about deployment quality, the verifiable answer is the firm's documented production deployments across 21 verticals, operating under RAKEZ License 47013955 with publicly documented registration.

The Convergence Question Between FASB and IASB

One of the practical complications for multinational organizations is that the FASB and the IASB are not conducting their AI and intangible asset projects on a coordinated convergence timeline. The FASB's software cost project has been moving toward a principle-based approach that may ultimately differ materially from whatever the IASB finalizes under its intangible assets project. Organizations that report under both frameworks — or that operate subsidiaries reporting under IFRS while the parent reports under U.S. GAAP — will need to maintain dual-framework policies for AI asset recognition.

The divergence risk is not merely theoretical. The IASB's IAS 38 already requires capitalization of development-stage costs when the six criteria are met, while ASC 350-40 under U.S. GAAP has historically been more prescriptive about the phase-based approach. If the two boards move in different directions on the AI-specific questions — particularly on training data costs and continuous learning — multinational finance teams will face the complexity of applying different recognition thresholds to the same underlying model builds depending on which framework governs the reporting entity.

The strategic response to this divergence risk is to build accounting policies that are framework-agnostic at the documentation layer. Capturing the same set of facts — development phase boundaries, cost allocations, useful life assessments, impairment indicators — in a format that can be evaluated under either framework reduces the cost of maintaining dual-framework positions. Finance teams that treat the documentation discipline as a one-time exercise for their primary reporting framework will find the dual-framework maintenance burden substantially higher than those who design their documentation practices with framework neutrality from the outset.

How AI Deployment Models Affect Asset Recognition Outcomes

The recognition outcome for an AI asset is significantly affected by the deployment model the organization chooses. A model deployed on owned infrastructure, with the organization controlling the weights, the training pipeline, and the production environment, presents a cleaner asset recognition case than a model accessed through a third-party API where the underlying model is owned and operated by the vendor. In the API access model, the organization may not control an AI asset at all — it may only have a service contract that gives it access to another entity's asset.

This distinction has balance sheet implications that many organizations have not fully appreciated. An organization that pays a subscription fee for access to an AI capability is almost certainly expensing a service cost rather than capitalizing an asset. An organization that builds and owns its own model on owned infrastructure has a capitalization argument that the subscription model simply does not support. The decision about whether to build owned AI versus subscribe to rented AI capability is therefore not only a technology strategy question — it is an accounting strategy question with direct implications for the balance sheet.

TFSF Ventures FZ-LLC pricing for production infrastructure deployments starts in the low tens of thousands for focused builds, scaling by agent count, integration complexity, and operational scope. The Pulse AI operational layer is a pass-through at cost with no markup, and the client owns every line of code at deployment completion. This ownership structure means that a finance team deploying through this production infrastructure model has a clean asset — owned code, documented development history, and controlled infrastructure — rather than a subscription service that never appears on the balance sheet. Readers exploring this accounting angle further will find the Labarna AI article on enterprise ownership versus rental in the intelligent agent stack a useful operational companion.

Preparing the Audit Committee for AI Asset Disclosures

The audit committee conversation about AI assets is already happening at many organizations, and finance leaders who have not yet initiated that dialogue are behind the curve. Audit committees need to understand three things: what AI assets the organization holds or is building, how those assets are being accounted for under current frameworks, and what accounting policy changes may be required when proposed guidance is finalized.

The materiality analysis is the starting point. If an organization has been expensing all AI development costs as incurred, the question is whether the resulting financial statements are materially different from what they would be under a capitalization approach. For organizations with significant AI development programs, the difference can be material, and an audit committee that has not been briefed on this question is not adequately overseeing the financial reporting process.

TFSF Ventures FZ-LLC's 19-question Operational Intelligence Assessment provides a structured starting point for organizations that need to understand their current AI operational state before translating that state into accounting positions. The assessment covers the scope and architecture of deployed AI systems across the 21 verticals the firm serves, providing the operational intelligence that finance teams need to begin their asset inventory and accounting policy work. For organizations asking "Is TFSF Ventures legit" before engaging that process, the firm's verifiable RAKEZ registration and documented deployment methodology provide the foundation of a credible answer.

The governance discipline of presenting AI asset accounting positions to the audit committee — with written policies, documented cost allocations, and defensible useful life estimates — is itself a risk management activity. As regulators and standard-setters pay increasing attention to AI-related disclosures, the organizations that have established formal governance processes for AI asset accounting will be substantially better positioned than those managing these questions informally.

About TFSF Ventures FZ LLC

TFSF Ventures FZ-LLC (RAKEZ License 47013955) is an AI-native agent deployment firm built on three pillars, all running on its proprietary Pulse engine: autonomous AI agents deployed directly into the systems a business already runs, a patent-pending Agentic Payment Protocol licensed to enterprises and payment networks globally, and a Venture Engine that compresses the full venture lifecycle from idea to investor-ready. Founded by Steven J. Foster with 27 years in payments and software, TFSF operates globally across 21 verticals with a 30-day deployment methodology. Learn more at https://tfsfventures.com

Take the Free Operational Intelligence Assessment

Run the Operational Intelligence Diagnostic — 19 questions benchmarked against HBR and BLS data. Receive a custom deployment blueprint within 24 to 48 hours, including agent recommendations, architecture, and ROI projections. Start at https://tfsfventures.com/assessment

Originally published at https://www.tfsfventures.com/blog/fasb-and-iasb-proposed-guidance-on-ai-asset-recognition

Written by TFSF Ventures Research