A finance team closes the books on January 5. The December invoice from an AI provider will not arrive until January 18, but engineering consumed millions of tokens in December. This is the practical question behind which AI costs need accruals: if a service was used before period-end and the company has an obligation to pay for it, the cost generally belongs in that closing period, invoice or not.

For organizations using multiple model providers, cloud AI platforms, gateways, and embedded AI tools, this is becoming a meaningful close issue. The challenge is not only estimating the total liability. It is assigning that liability to the business unit, product, project, or agent that generated it, so the P&L, forecast, and eventual vendor invoice tell the same story.

Which AI costs need accruals at month-end?

The core accounting principle is straightforward: accrue an expense when the company has received the goods or services, the amount can be reasonably estimated, and payment has not yet been recorded. For AI, consumption is often metered continuously while billing follows days or weeks later. That timing gap is where accruals matter.

The most common candidates are usage-based LLM API charges, cloud-hosted model inference, GPU or compute consumption, vector database usage, AI gateway fees, and managed AI services delivered during the period. If a team used those services in December, the related expense should generally be reflected in December, even if the invoice is dated in January.

Consider an internal support agent that processed 2 million customer interactions during the final week of the month. The company can obtain token records from its gateway, multiply them by the contracted input and output rates, and estimate $42,000 of unbilled model usage. If those services were consumed before month-end, an accrual is typically more faithful than waiting for the vendor invoice.

A simple entry could look like this:

```text Dr. AI Services Expense - Customer Support $42,000 Cr. Accrued Expenses Payable $42,000 ```

The expense account and cost center should reflect the economic owner, not merely the team that administers the vendor account. If product, sales, and support all consume the same provider account, putting the full amount in IT creates a distorted view of operating costs and AI ROI.

Costs that are usually accrued

Usage-based services are the clearest case, but the list can extend beyond model tokens. Organizations should evaluate the following at each close:

Model and inference usage: API calls, input and output tokens, image or audio generation, batch jobs, fine-tuning runs, and model-hosting charges incurred but not invoiced.
Cloud AI infrastructure: GPU instances, model endpoints, storage, data processing, networking, and serverless workloads that support production AI services.
Third-party AI operations services: Usage-priced observability, evaluation, guardrail, orchestration, or gateway services consumed during the period.
Professional and managed services: Work performed by an external provider before period-end, where the contract, statement of work, or acceptance evidence supports a reasonable estimate.

The accounting conclusion depends on the contract and the facts. A monthly platform minimum may be due regardless of usage, for example, while overage charges may depend on actual consumption. Both can require a close review, but the evidence and estimation method may differ.

Costs that do not automatically require an AI accrual

A purchase order, budget approval, annual commitment, or forecasted future workload is not itself an expense. Finance should not accrue the full value of a signed AI contract simply because procurement has committed to it. The question is what was delivered or consumed as of the reporting date.

Prepaid subscriptions are also different from accrued expenses. If the company pays $120,000 upfront in January for a 12-month AI platform subscription, the initial entry is generally a prepaid asset, then expense is recognized over the service period. Accruing the full annual amount would overstate current-period expense.

Unused credits need similar scrutiny. If a company buys credits that remain available for future use, the accounting treatment may be prepaid until consumption occurs, subject to the contract terms and the organization’s accounting policy. Credits that expire, are nonrefundable, or are tied to minimum commitments can introduce additional complexity. Material arrangements should be reviewed with the controller and, where appropriate, external auditors.

Internal employee costs are another area where labels can confuse the analysis. Salaries for an AI engineering team follow the company’s normal payroll and compensation accrual process. Calling the team “AI” does not create a separate type of accrual. The more relevant question is whether any portion of internally developed software qualifies for capitalization under the company’s applicable accounting policies. That decision should be made deliberately, not inferred from cloud invoices or project names.

Build the accrual from usage, not from a vendor bill

A single vendor statement may provide a total balance but not enough evidence for an accurate cutoff or allocation. The best accrual process starts with a defined data hierarchy: actual metered usage through the final day of the period is strongest; provider usage dashboards and cloud billing exports are next; contractual rates are used to convert units into dollars; and a documented estimate fills only the remaining gap.

For example, assume an organization has December usage from December 1 through December 29, but the final two days are unavailable at close. Finance can calculate actual charges through December 29 and estimate December 30-31 based on recent daily usage, adjusted for known releases, outages, or end-of-year traffic patterns. That is more defensible than accruing a flat percentage of the prior month.

The estimate should be retained with its source data, assumptions, preparer, reviewer, and reversal plan. Audit scrutiny tends to focus on the cutoff, the basis for the estimate, and whether the eventual invoice is compared with the accrual. A repeatable schedule answers all three.

Allocate before booking the entry

Allocation is not a reporting enhancement added after close. It determines whether expense ownership is credible in the first place.

The preferred allocation method uses direct identifiers: business unit, application, environment, project, agent, customer, or cost center attached to each request. A customer-support agent’s token usage can then be charged directly to customer support, while an internal coding assistant can be assigned to engineering.

When direct attribution is incomplete, use a transparent secondary driver. API calls, tokens, GPU hours, transactions, active users, or workload share can be reasonable drivers depending on the service. An equal split across departments is easy, but it is rarely defensible for a growing AI program because usage is almost never equal.

Suppose $100,000 of unbilled AI usage is accrued. Usage data shows that product consumed 55%, support 30%, and sales operations 15%. The entry should reflect those owners rather than one centralized technology cost center:

```text Dr. AI Expense - Product $55,000 Dr. AI Expense - Customer Support $30,000 Dr. AI Expense - Sales Operations $15,000 Cr. Accrued AI Services Payable $100,000 ```

This approach also improves forecasting. Department leaders can see the actual run rate they own, rather than being surprised later by an allocation that appears unrelated to their operating decisions.

Reverse, reconcile, and true up

Most usage-based AI accruals should reverse automatically in the following period. When the vendor invoice arrives, accounts payable records the actual invoice against the expense account. The reversal prevents the same period of service from being expensed twice.

Using the earlier $42,000 example, finance would reverse the accrual on January 1. If the January invoice for December usage is $43,200, the $1,200 difference becomes a normal true-up. That variance should not be ignored. It may reveal a price change, incomplete usage telemetry, a late-arriving service category, or a flawed allocation rule.

Over time, track accrual-to-invoice variance by vendor and cost type. A small, consistent variance may support a simplified method. Large or volatile variances signal that finance needs better source data or a shorter billing-data lag. Materiality matters as well: a $500 estimate error may be acceptable for one team, while a 10% error on a seven-figure AI spend line is a control problem.

Put AI accruals into the close calendar

The practical goal is not a perfect estimate at any cost. It is a controlled process that produces a materially accurate result within the close timetable.

Assign ownership across finance and technology. Platform or FinOps teams should provide metered usage and explain technical changes. Finance should define materiality thresholds, review assumptions, book the journal entry, and reconcile estimates to invoices. Procurement can supply rate cards, minimum commitments, and contract changes that alter the calculation.

As AI use expands, the most useful close control is a monthly AI spend package: usage through cutoff, estimated unbilled consumption, allocation by owner, journal entry support, and prior-month true-up analysis. That turns a vague cloud bill into evidence finance can book and operators can act on.

The closing question is not whether every AI charge is perfectly known on day one. It is whether the company can show what it consumed, who benefited, how the estimate was calculated, and what happened when the invoice arrived. That is the discipline that lets AI spending scale without turning the close into a recurring surprise.

Which AI Costs Need Accruals? A Practical Guide