AI cost of goods sold, the new accounting discipline of the generative era

AI cost of goods sold is the new accounting category that finance leaders are beginning to formalise as generative AI moves from a supportive overhead into a direct input of the product or service being sold. When an AI call is the reason a specific deliverable exists, its cost stops behaving like a SaaS overhead line and starts behaving like cost of goods sold: variable per unit of output, deducted above gross margin, and directly attached to revenue recognition. This article describes what AI cost of goods sold means, why it does not fit into existing IFRS, GAAP, or Dutch RJ categories cleanly, and what the practical implications are for gross margin, unit economics, and pricing decisions for software companies, agencies and any organisation whose deliverable is now partly produced by an LLM.

The audience is CFOs, controllers, finance leads and accountants at European organisations where AI has recently crossed the line from experimental tool into a production input. If your P&L still lists AI under "software subscriptions" but the AI call is what actually generates the invoiceable output, this page is the framework for what to do next.

Why AI cost of goods sold is a genuinely new category

For a century, cost of goods sold (COGS) has meant the direct costs of producing the goods or services a company sells: materials, direct labour, and the factory overhead that could be reasonably allocated to specific units. Everything else, from office rent to sales commissions to SaaS subscriptions, lived above the line as operating expense. That framework worked because the boundary between "what makes the widget" and "what supports the company making widgets" was clear at the physical scale most accounting standards were designed around.

Generative AI breaks that boundary because a single API call to Anthropic Claude or OpenAI GPT is often the direct producer of the specific deliverable being invoiced. For a marketing agency, an AI-generated draft is not a supporting tool; it is the deliverable, or the substantial majority of it. For a SaaS company whose product runs prompts on behalf of end users, the LLM call is the product being consumed. In both cases the AI cost is variable per unit of output, directly attached to revenue, and impossible to reason about as overhead without hiding what is actually happening.

The three properties that qualify a cost as COGS in every serious accounting framework (variability per unit, direct attachment to output, causal necessity to production) apply to a large share of modern AI spend. And yet no accounting standard, from IAS 2 (Inventories) to Dutch RJ 270 (De winst-en-verliesrekening), currently has a category that maps onto per-request generative AI spend as a component of COGS. Standards were written when the concept of "the cost of producing one deliverable" implicitly meant physical inputs or human hours, not model tokens billed by a third-party inference provider.

That gap is what makes AI cost of goods sold a genuinely new category. It is not just a re-labelling exercise. The measurement problem is different, the recognition timing is different, and the disclosure implications are different from anything current standards were designed to handle.

What qualifies as AI cost of goods sold, in practice

The practical test for whether a specific AI cost belongs in COGS or in operating expense turns on three questions.

Is the cost variable per unit of output sold? An LLM call that runs once per client deliverable is variable per unit. A ChatGPT Enterprise subscription used by the whole team is not, even if it supports client work. Variability at the unit level is the first filter.

Does the cost disappear if the specific unit is not produced? If you did not produce this particular blog post, this particular report, this particular support-agent response, would this specific AI call have happened? Yes for a per-request generation, no for a monthly seat licence. Causal necessity is the second filter.

Is the cost attributable to the specific unit with reasonable effort? Attribution needs to work at the per-deliverable grain, which is where identity-based AI spend attribution and code-path attribution matter. If you can only attribute to the team level, the cost is closer to overhead in accounting terms; if you can attribute to the specific deliverable, the cost is COGS-eligible.

When all three tests pass, the cost belongs in AI cost of goods sold rather than in operating expense. When only the first two pass but attribution is coarse, disclosure typically improves by treating the cost as a distinct line inside COGS anyway, even if allocated by a defensible key (revenue-weighted, deliverable-count-weighted, or hours-weighted).

The accounting treatment question no standard has answered

The current accounting standards do not tell you how to book per-request AI cost, because the transaction shape did not exist when the standards were written. Three specific gaps show up when finance teams try to apply existing rules to AI spend.

Timing of recognition. Under IFRS 15, revenue is recognised when a performance obligation is satisfied. For an AI-produced deliverable, the corresponding cost should be recognised at the same moment, matching the cost to the specific revenue event. In practice, provider invoices arrive at the end of the month aggregated across all calls, which means finance teams either accrue at request time (requiring per-request cost data) or wait for the invoice and mis-match the timing.

Capitalisation vs expensing. Traditional software cost has capitalisation criteria under IAS 38 (Intangible Assets) and FASB ASC 350-40, but per-request LLM calls do not create an intangible asset. They are consumed on the spot. The natural treatment is expensing at the moment of consumption, but that requires cost data at the same moment as the consumption event.

Disclosure granularity. Regulators and auditors have not yet issued specific guidance on whether AI cost warrants separate line-item disclosure. Best practice emerging from IFRS filers in 2026 is to disclose AI cost of goods sold as a separate line within COGS when it exceeds five percent of revenue, or as a note when smaller. The Autoriteit Financiële Markten has flagged AI disclosure as an area of increased supervisory interest for listed companies without yet publishing a specific standard.

Until the standard-setters issue specific guidance, finance teams have to make defensible choices. The defensible choice consistently made by CFOs at AI-native companies in 2026 is: treat per-request AI cost as COGS when it passes the three-question test above, capture per-request cost data at the moment the request happens, and disclose separately when material.

Gross margin implications of AI cost of goods sold

Once AI cost migrates from operating expense to cost of goods sold, gross margin drops mechanically. A software company whose gross margin was 82% under the "AI is a SaaS subscription" treatment may report 71% under the "AI is COGS" treatment, without any operational change. This is not a deterioration in business quality, it is a reclassification, but the reported number moves.

The consequence is that gross margin comparisons across companies and across years become misleading unless the AI treatment is disclosed. A SaaS company that has always expensed AI as opex will show a higher gross margin than an equivalent AI-native company that treats it as COGS, even if the underlying economics are identical. Analysts, board members and acquirers who benchmark on gross margin need to know which treatment is in use.

The practical response is threefold. First, disclose the treatment explicitly, both in filings and in board reporting. Second, compute the gross margin both ways during transition periods so board members can see the like-for-like comparison. Third, budget the migration: if the treatment changes mid-year, both the historical restatement and the ongoing measurement infrastructure need budget lines that most companies do not currently have.

Unit economics with AI cost of goods sold

The upside of migrating AI cost into COGS is that unit economics finally become computable at the deliverable level. Once every deliverable carries its own AI cost, the fundamental questions become answerable directly:

• What is the gross margin on this specific client engagement?
• Which product feature has the lowest per-unit AI cost, and which has the highest?
• At what price point does a specific SKU cross into unprofitable territory?
• Which pricing tiers should be repriced because AI cost per user in the tier has grown faster than expected?

For subscription SaaS, this is the missing input for cohort economics that has been guessed at for years. For agencies, this is the input for defensible per-client margin analysis that survives a partner meeting. For any business shipping AI-touched output, this is the source of the CFO conversation that either justifies expanding the AI programme or pulling back specific workflows.

For the strategic framework that consumes AI unit economics as input, see strategic AI cost management for SME CFOs. For the attribution mechanisms that produce the underlying data, see AI spend attribution and attribute AI costs to teams.

The governance requirement behind defensible AI COGS numbers

Defensible AI cost of goods sold reporting rests on infrastructure that most SMEs do not currently have. Three concrete requirements have to be met for the numbers to survive an auditor question.

Per-request cost data at the moment of the transaction. Provider month-end invoices do not have the granularity to support COGS treatment because they aggregate across all calls in the billing cycle. Per-request capture requires a governance layer that records cost as each request happens, not after.

Deliverable-linked attribution. The per-request cost record needs to be linked to the deliverable (client engagement, product feature, subscription tier) whose revenue it supports. Identity-based attribution handles the team level; per-deliverable attribution requires either explicit deliverable metadata passed at the request site or a downstream join to your CRM, project system, or billing system.

Retention through audit window. For audit and inquiry, per-request records typically need to be retained for the length of the applicable audit window (in the Netherlands, seven years for tax records under article 52 of the Algemene wet inzake rijksbelastingen). Providers rarely retain per-request data for that period; the retention has to happen on your side.

An organisation without these three capabilities can still report AI cost as COGS, but with less defensibility if challenged. As the NBA and other professional bodies extend their guidance on AI in accountancy through 2027, the audit standard for defensibility is likely to tighten. Building the infrastructure now is cheaper than retrofitting it during an auditor conversation.

Talk to Ciralgo

If AI cost has crossed the line from tool to input at your organisation, and the question of how to book it defensibly is now on your controller's desk, a 20-minute call is enough to see whether Ciralgo fits. We build the EU-hosted per-request capture and attribution layer that produces the audit-defensible numbers a COGS treatment requires. We do not only advise, we build. Book a slot via our contact page, we reply within one working day.

Frequently asked questions

Is AI cost of goods sold recognised by IFRS or GAAP?

Not as a distinct standard as of 2026. IFRS 15 and IAS 2 provide the general framework for cost of goods sold, but neither addresses per-request AI cost specifically. FASB and IASB have both included AI disclosure in their 2026-2028 workplans. Until specific guidance issues, the defensible treatment is to apply general COGS principles (variability, direct attachment, causal necessity) to AI cost and disclose the treatment.

Should a SaaS company shift AI cost from opex to COGS?

If the AI call is part of the product being consumed by the subscriber (retrieval, generation, transcription that the user sees), yes, the natural home is COGS. If the AI use is internal only (support automation, internal tools that make employees faster), the natural home is opex. Many companies have both and split accordingly.

How does AI cost of goods sold affect SaaS gross margin benchmarking?

It typically reduces reported gross margin compared to the pre-AI treatment, without changing underlying economics. Analysts benchmarking gross margin across SaaS companies in 2026 and beyond will need to normalise for AI treatment differences. Disclosure of the treatment is becoming best practice for that reason.

Does the EU AI Act affect how AI cost of goods sold is disclosed?

Not directly. The AI Act deals with system safety, transparency to end users, and deployer obligations, not accounting treatment. However, the audit-log infrastructure that the AI Act Article 12 requires is the same infrastructure that produces defensible per-request cost data, so the two obligations naturally converge.

How is AI cost of goods sold different from cloud cost of goods sold?

Cloud COGS has been an established practice at SaaS companies for a decade, tied to per-tenant infrastructure allocation. AI COGS differs because the cost is per-request rather than per-tenant, the price basis (tokens) is not a fixed unit of infrastructure capacity, and the model routing decision itself is a lever that changes unit cost within the same output. Cloud COGS methodologies transfer partially but need adaptation.

What is the tax treatment of AI cost of goods sold in the Netherlands?

Under the Dutch Corporate Income Tax Act, cost of goods sold is deducted as an operating expense in the year of consumption. The tax treatment of AI cost is straightforward as an expense; the accounting question of whether it appears above or below the gross margin line is a presentation question, not a tax question. The Belastingdienst has not yet issued AI-specific guidance.

Further reading

• AI cost management for European SMEs: the 2026 practical guide
• Strategic AI cost management for SME CFOs: a four-part framework
• AI spend attribution: multi-model, code-path and identity-based mechanisms
• AI voor accountants: wat de NBA-regels toestaan
• AI cost attribution per client for agencies and consultancies
• How to attribute AI costs to teams: four methods for SMEs
• What is an EU-hosted LLM proxy
• Ciralgo pricing

Last reviewed 8 September 2026. This page is a forward-looking framework, not audit advice. Consult a chartered accountant for treatment specific to your organisation.