Usage Billing Review

Accrued Expenses Examples for AI and Cloud-Dependent Businesses

Unbilled AI and cloud usage creates liabilities finance must estimate.

· 9 min read
Cover illustration for “Accrued Expenses Examples for AI and Cloud-Dependent Businesses”
Accrued Revenue Accounting · November 10, 2026 · 9 min read · 2,014 words

Accrued expenses keep building up at AI and cloud-dependent businesses as their products use services before vendors bill for them. Accurate accounting depends on consumption data that finance can use at period-end.

Why AI and cloud businesses carry accrued expenses that traditional SaaS companies do not

A fixed, seat-based software contract gives finance a known amount and a renewal schedule. Consumption-based pricing removes that certainty because the amount owed changes throughout the contract lifecycle. The contract no longer supplies the complete answer. Finance may still plan around fixed entitlements, while use of the product changes costs between billing dates. The structural break is therefore not simply higher spending, but a liability whose measurement depends on operational activity.

Zylo’s 2026 SaaS Management Index found that 78% of IT leaders have experienced unexpected charges on a SaaS bill due to consumption-based or AI pricing models. Those surprises show a gap between the use buyers planned for and how much they actually used. For accounting, that gap becomes a cutoff problem. Costs belong to the period in which they were incurred, even when the information needed to measure them arrives later.

What an accrued expense is and why the definition matters more here than elsewhere

An accrued expense is a liability recognized when a cost is incurred, regardless of when an invoice arrives or cash is paid. At period-end, the company estimates unbilled costs, debits the relevant expense account, and credits accrued liabilities. It then reverses or adjusts the accrual when the invoice arrives. The accounting mechanics themselves have not become more complicated. The difficulty lies in determining the amount to record.

For a fixed software contract, the agreement usually supplies that amount. For token-based, GPU-minute, or API-call pricing, the agreement supplies a rate, while metering supplies the quantity. Finance needs both to calculate the liability. If consumption records are incomplete or inaccessible, the estimate rests on assumptions. That uncertainty increases the risk that the eventual invoice will require a material true-up, shifting reported expense away from the period that generated it.

Accrued cloud infrastructure costs as the baseline liability

Cloud infrastructure leads to a recurring accrual because use goes on while vendors bill in arrears. Unbilled costs include hosting for running instances, storage, data egress, and additional charges when usage exceeds reservations. Traffic and batch jobs change the total. A month-end estimate based on an earlier invoice can therefore leave out costs already incurred, especially when demand is growing fast.

Where infrastructure supports product delivery, an understated accrual can understate COGS and overstate gross margin. The error affects more than the balance sheet. It makes the services delivered in the period look cheaper than they were, though the vendor’s later invoice will show the difference.

Consensus Cloud Solutions’ Q2 2026 Form 10-Q lists accrued advertising expenses, compensation, sales and other taxes, interest, and other accrued expenses. The filing does not identify a separate cloud infrastructure accrual. Its relevance is the broader balance-sheet context in which cloud-dependent businesses report liabilities. Estimates of the costs of running the business must fit into that accounting structure, even if each type of consumption has no public line item of its own.

Accrued LLM inference costs as a new liability

LLM inference adds a cost that scales with model calls rather than simply with the infrastructure kept running. Each prompt, agent step, or automated workflow can use tokens that an upstream provider bills for later. By month-end, the cost of those tokens has already been incurred. The accrual depends on internal metering that may not yet agree with the provider’s count.

A product launch can introduce this liability within a single quarter. A company launching AI Inference Services mid-year may have no comparable LLM expense in the prior-year period, followed by a significant cost in the launch quarter. Past spending offers little help in that case. Nor can finance reliably use user counts to work out token consumption, because the choice of model, prompt length, and output length change what each interaction costs.

Upstream pricing adds another source of uncertainty because the buyer does not control the provider’s terms. Changes in those terms change how much the same activity costs. Where inference sits in COGS, the estimate also shapes reported unit economics. A product can appear to have stronger gross margins simply because its unbilled token consumption has not been fully captured.

Accrued GPU compute costs when jobs cross period boundaries

Training and fine-tuning create a different accrual pattern from continuous inference. A training run beginning on the 20th and ending on the 10th of the following month spans two accounting periods. Costs incurred through the 31st belong to the first period, even though the job is still running and there is no final invoice. Completion is not the accounting cutoff.

The estimate starts with instance type, hourly rate, and elapsed time. The cost can turn out differently if the job ends early, spot instance interruptions occur, the job restarts, or the final bill is adjusted. Discounted interruptible compute can therefore lower the estimate while complicating its calculation. Finance needs to work out the cost incurred through period-end rather than just assign a share of what the whole job is expected to cost.

Minimum-spend commitments introduce a further distinction. Where a GPU agreement requires payment for a period regardless of use, the applicable committed amount sets a minimum obligation. Job-level costs and contractual minimums therefore require different estimation methods.

Accrued API consumption costs embedded in the product

External APIs for search, vision, speech, translation, and geolocation create liabilities through small charges that accumulate across product activity. A feature may generate a per-call or per-unit charge each time it invokes an external service. The individual charge can look negligible. Across a growing product, however, those calls can produce a significant unbilled balance.

Vendors differ in their billing cycles, how late their invoices arrive, and how much usage they show in their developer consoles. Without a unified record of outbound calls, finance assembles the accrual from dashboards, usage emails, and earlier invoices. These sources do not necessarily describe the same cutoff. A prior-period invoice used as a proxy is especially weak during rapid growth because it reflects a smaller volume of activity.

Credit pools introduce a related accounting distinction. Purchased credits represent a prepaid asset for the buyer, rather than an accrued expense for services not yet paid for. Consumption metering determines how much of that asset remains. Finance therefore needs to distinguish prepaid use from unbilled use, even when both support the same product feature.

Accrued compensation and variable pay tied to usage-driven revenue

When the customer signs the contract, it may be too soon to work out a sales commission tied to usage-based revenue. Its amount depends on subsequent consumption and the revenue associated with it. The signed agreement is only the starting point. If the underlying usage total is uncertain, the commission base and resulting compensation accrual inherit that uncertainty.

Outcome-based pricing adds a further dependency. Where a product charges only when an agent resolves an issue without human intervention, period-end accounting depends on the number of qualifying resolutions. Those outcomes may also lead to commissions or support costs. An incomplete outcome count can therefore affect both revenue and the related compensation estimate, while variable pay can also change accrued payroll taxes and benefits.

The Consensus filing reports accrued compensation as of June 30, 2026. That disclosure establishes the presence of the liability. When pay depends on consumption or outcomes, the records supporting that pay must be measured with the same rigor as the underlying revenue.

Accrued sales tax and regulatory liabilities created by usage-based transactions

With usage-based pricing, the tax base can change too. A fixed subscription provides a known charge, while a consumption-based bill depends on activity totaled at the end of the billing period. Where that activity is taxable, an incomplete usage estimate affects the related tax liability. The uncertainty extends beyond operating costs. Finance may need to estimate both the revenue earned and the tax on it before billing is done.

Jurisdiction and classification add separate questions. The location associated with consumption may differ from the contract location, while the treatment of AI inference, API access, and token-based services may be unsettled. These questions make tax accruals dependent on both measured activity and the applicable treatment of that activity, with any tax obligation associated with routing through a location subject to that treatment.

The same filing reports accrued sales and other taxes as of June 30, 2026. It illustrates that tax liabilities sit within the accrued balance of a cloud-dependent business. It does not establish the tax treatment of any particular AI service or transaction.

Why these accruals are structurally harder to estimate than conventional software costs

The central difficulty is that the activity record and the financial obligation are produced by different systems. Internal meters count events, while vendors use their own records, billing cycles, and pricing terms. Agreement may emerge only after invoicing. Until then, finance must translate operational evidence into an accounting estimate without assuming the two records are identical.

Metering lag, duplicate events, and late-arriving records each distort that translation differently. Missing events can understate consumption, while duplicates can inflate it. Late records pose a cutoff problem even when the counts in them are right. Multi-vendor aggregation compounds the difficulty because a consolidated balance may combine sources with different reporting delays.

Zylo’s index reports that spending on AI-native applications surged 108% year-over-year across enterprises. As spending grows, errors in the estimate have a bigger effect on the accounts. A systematic understatement reduces reported liabilities and inflates income, while overstatement does the reverse. Finance therefore needs evidence supporting the amount recorded for the correct period.

How real-time metering infrastructure reduces accrual estimation error

Reliable metering narrows the gap between consumption and accounting evidence. Finance needs deduplicated event counts by customer, cost category, and vendor through the last second of the accounting period. A delayed dashboard leaves the cutoff incomplete. An ingestion layer captures events, an aggregation layer totals them for the period, and a pricing engine applies the relevant rates.

Accuracy also depends on how the pipeline handles exceptions. Idempotency keeps the same event from being counted twice, while late-arriving events and corrections must go into the period they belong to. Speed alone does not provide those controls. A fast but duplicated event stream can produce a less reliable accrual than a slower, reconciled one.

The same data must support distinct accounting balances when subscriptions, usage overages, and prepaid credits coexist. Unused customer credits relate to deferred revenue, earned but unbilled usage relates to accrued revenue, and unbilled vendor consumption relates to accrued expense. These balances are not interchangeable. A billing platform that unifies metering and pricing can make their shared consumption evidence accessible to finance, while a custom stack requires ongoing engineering resources for accuracy and uptime.

What CFOs and controllers at AI companies should do differently at period-end

Finance and engineering should agree on a cutoff procedure and a daily or hourly report that finance can obtain independently of vendor billing. That report becomes the basis for the estimate. Engineering should participate in the close because event logs and internal meters provide evidence that a vendor dashboard alone may not supply.

Accruals should then be grouped by estimation method. Fixed-rate commitments and closely reconciled usage can be accrued mechanically, while token and API estimates need documented methods and variance ranges. Amounts that depend on outcomes need an estimate backed by evidence and, where needed, disclosure of what is still uncertain. Each settled invoice should be compared with its original accrual. Repeated differences signal a metering gap or billing error that needs correction before the next close.

Real-time spend alerts should also serve finance, giving early notice when consumption diverges from the estimate. Historical accruals are not sufficient baselines during rapid growth. Zylo’s index reports that AI-native app spend grew 400% year-over-year for large enterprises, reinforcing the need to use current activity. The ability to close accurately and quickly ultimately depends on consumption data that is real-time, deduplicated, and accessible to finance.

Sources

  1. Consensus Cloud Solutions, Inc. - Form 10-Q - FY2026
  2. AXON ENTERPRISE, INC. - Form 10-Q - FY2026

More in Accrued Revenue Accounting