Should You Capitalise Your AI Costs? A Practical Guide
Most AI investment lands on the P&L as expense. But if your AI project meets the right criteria—identifiability, control, technical feasibility—the costs can be capitalised as an asset. This guide walks through IAS 38, ASU 2025-06, and the decision tree every CFO needs.
The Accounting Gap: Why Most AI Spend Is Expensed
Trillions in AI investment worldwide, yet most of it flows straight to the income statement. This is because AI systems are notoriously difficult to recognise as distinct assets under existing accounting standards.
The core problem is identifiability. Traditional intangible asset standards (IAS 38 or ASC 730) were written for patents, licences, and acquired software—assets you could point to, separate, and sell. An in-house trained AI model embedded across your operations? That fails the identifiability test in most regulatory frameworks.
IAS 38 and US GAAP treat AI development costs differently. IFRS creates a high bar for capitalisation; US companies have more flexibility under ASU 2025-06. Understanding which standard applies to your business is the first step.
The result: your P&L is inflated with what should arguably be capital investment. And from a strategic perspective, you lose visibility into the true cost of building AI capability.
IAS 38: The Six-Criterion Test
Under International Accounting Standards, an intangible asset—including AI systems—can be capitalised only if it meets all six criteria:
1. Identifiability
The asset must be separable from the business or arise from contractual or legal rights. Is your AI model distinct enough to be ring-fenced, licensed, or sold separately? Most integrated AI systems fail here.
2. Control
Your business must have control: the ability to use it, restrict others' access, and extract benefits. Ownership of training data and model weights generally satisfies this.
3. Probable Future Economic Benefits
Quantifiable benefits from the AI system: revenue uplift, cost reduction, or improved decision-making. Speculative benefits do not count.
4. Technical Feasibility
Proof that the AI system can function as intended. Completed prototypes, MVP results, or pilot data satisfy this criterion.
5. Ability to Measure Cost Reliably
You must track development costs with precision: labour hours, compute infrastructure, training datasets, and external consulting. Vague cost buckets disqualify the asset.
6. Available Resources
Proof that you have or can secure the financial, technical, and human resources to complete development and deploy the asset.
The identifiability hurdle is the killer. If your AI system is tightly woven into your core product or operations, regulators will argue it is not separately identifiable. Standalone AI tools (chatbots, recommendation engines, internal analytics platforms) have a better chance.
A pharmaceutical firm spent £8.2M building a proprietary drug discovery AI. The model can be licensed to partner companies and withheld from competitors. IAS 38 auditors approved capitalisation: identifiability clear, control proven, future benefits quantified via royalty projections. Contrast this with a retail chain's AI demand forecaster, embedded in inventory management across all stores—integration defeats separability, so the £2.5M cost is expensed.
ASU 2025-06: US GAAP Takes a Different Path
In March 2025, the FASB released ASU 2025-06, specifically addressing internal-use software and AI. The US standard is deliberately more permissive than IFRS.
IAS 38 (IFRS)
- Identifiability test is mandatory and strict
- AI systems must be separable or contractually bound
- Integration into operations generally disqualifies asset recognition
- Most AI spending stays on the P&L
ASU 2025-06 (US GAAP)
- Internal-use software presumed capitalisable if developed internally
- AI systems classified as software; no separability requirement
- Identifiability assumed for tracked development costs
- More AI spending can move to the balance sheet
For a US company, ASU 2025-06 unlocks significant balance-sheet optionality. If you can trace your AI development costs—labour, infrastructure, training data acquisition—the standard presumes you can capitalise them. For an IFRS reporter, the burden is far higher.
What Can Be Capitalised? A Cost-by-Cost Breakdown
Development Phase: Capitalise These Costs
Labour: salaries and benefits for engineers, data scientists, and product managers actively building the AI system (not general IT support). Infrastructure: compute (GPU/TPU clusters), storage, model training runs, version control. Data: cost to acquire, clean, label, or generate training datasets. External services: consultant fees for architecture or model development.
Once the AI system moves into production and is available for use, the capitalisation phase ends. Any costs after that point—maintenance, monitoring, retraining, infrastructure for live inference—are expensed.
Non-Capitalisable Costs
Research and feasibility studies (pre-development): literature reviews, proof-of-concept experiments where you are still evaluating whether to build. General IT maintenance: server upkeep, security patches, database administration. Regulatory and compliance reviews: AI audit, fairness testing, regulatory filing fees. Retraining and fine-tuning: ongoing model updates post-deployment.
The dividing line is when the asset becomes available for use. IAS 38 and ASU 2025-06 both require this gateway. Once your AI system is live, development stops, and operational costs begin.
The Capitalisation Decision Tree
Walk through these questions in order. If you answer 'no' to any question, the costs are likely not capitalisable:
- Are you tracking costs separately? If labour, compute, and data costs are mixed into a general IT budget, capitalisation is off the table. Audit trail required.
- Can you prove technical feasibility? Do you have prototype results, MVP metrics, or pilot data showing the AI works? Speculation does not count.
- Are probable future benefits quantifiable? Revenue, cost savings, or decision-quality improvement tied to the AI? Vague 'strategic value' is insufficient.
- Is the AI separately identifiable (IFRS) or internally developed software (US GAAP)? If embedded and inseparable from your core business, IFRS auditors will likely object. US reporters have more latitude.
- Have you reached 'available for use'? Is the system deployed in production? If still in pilot or development, capitalisation may be deferred to a later period.
Impact on Financial Statements
Capitalising AI development costs moves expense from the P&L to the balance sheet, affecting four key metrics:
Profitability
If you capitalise £5M in AI development, reported net profit improves in Year 1 (no expense), but then deteriorates in Years 2–5 as amortisation kicks in. The total income statement impact over the asset's life is the same; only the timing shifts.
Return on Assets (ROA)
Capitalisation inflates the asset base, potentially reducing ROA. However, if the AI system generates revenue or cost savings, the effect may be offset.
Debt-to-Assets Ratio
More assets on the balance sheet can improve leverage ratios, making the business appear less risky—useful for covenant compliance or credit ratings.
Cash Flow
No impact. Capitalisation is purely an accounting treatment; the cash outflow occurs when the asset is built.
CFOs and boards should review the assumptions underlying capitalisation—particularly the expected useful life and likelihood of future benefits—with their auditors before committing. A capitalised AI asset that fails in Year 2 requires an immediate impairment charge, creating volatility and credibility damage.
Amortisation: How Long Does an AI Asset Live?
Once capitalised, the AI system is amortised over its expected useful life. For AI, this is typically 3–5 years for general-purpose models (large language models, commodity NLP systems) and 5–10 years for proprietary, domain-specific AI (drug discovery, oil reservoir forecasting, bespoke decision engines).
The shorter amortisation reflects the rapid pace of AI advancement and the risk of technological obsolescence. A chatbot deployed today may be superseded by a radically better model in 18 months.
Both IAS 38 and ASU 2025-06 require annual impairment testing if facts and circumstances suggest the asset's value has fallen (loss of competitive advantage, regulatory restriction, market shift). A poorly adopted or deprecated AI system triggers an impairment write-down.
The Bottom Line
The choice to capitalise or expense AI costs hinges on your accounting standard, your auditor's stance, and the asset's separability. US companies have a genuine capitalisation option under ASU 2025-06; IFRS reporters face a higher bar. Either way, rigorous cost tracking, proof of technical feasibility, and quantified future benefits are non-negotiable. Work with your CFO and auditors to establish the capitalisation policy upfront—retroactive adjustments are messy and weaken credibility.
Explore More
AI Valuation Methods
Learn how to value AI assets using Relief from Royalty, MPEEM, Cost, and Market approaches—essential for M&A.
Explore →AI ROI Framework
A step-by-step framework for measuring what AI investments actually create, with worked examples and board-ready metrics.
Learn More →Intangible Asset Categories
The CHS framework and seven categories of intangible assets, including the new assets AI creates.
View Taxonomy →Subscribe to our newsletter
Get the latest insights on intangible asset growth and productivity delivered to your inbox.
Get clarity on your AI capitalisation options
Book a consultation with our team to assess your AI costs, accounting standard requirements, and balance-sheet strategy.