Insights

How do I account for AI agent spend in the general ledger?

By Brian Diamond

Published October 5, 2026

Account for AI agent spend as a usage-based operating expense: accrue it at period end from metered usage, allocate it to the cost center the agent worked for, post it to a dedicated AI services expense account, and reverse and true it up when the provider invoice arrives. The mechanics are the same as any unbilled utility; what is new is that the usage is generated by software acting for several departments at once, so attribution has to happen before the entry, not after.

What makes agent spend different from a SaaS subscription

A SaaS seat is a fixed monthly charge with one obvious owner. Agent spend is variable, arrives from several sources, and is often unbilled at close:

  • Provider APIs (OpenAI, Anthropic, Google) invoice monthly in arrears, usually two to ten days after period end.
  • Cloud AI services (Bedrock, Vertex, Azure OpenAI) are buried inside the cloud bill under the infrastructure account.
  • Private inference on your own GPUs never produces an invoice; the cost is depreciation and power already booked elsewhere.
  • Agent-initiated purchases (a tool subscription, a data pull, a paid API) can land on a corporate card.

If you wait for invoices, the month closes with the spend missing, then lands lumpy a period later with no owner.

The five steps

  1. Create the accounts. One operating expense account, "AI services," with sub-accounts if volume justifies it (model inference, agent tools, private inference allocation). One accrued liability account, "Accrued AI services."
  2. Capture usage with attribution. Each usage record needs the agent identity, the task or workflow, and the cost object (department, project, or customer). This is captured at runtime by the gateway, the agent framework, or a system of record; it cannot be reconstructed from the invoice.
  3. Accrue at cutoff. Price the metered usage at the contracted rate, group it by cost object, and post the accrual.
  4. Allocate. Post each cost object's share to its department, or use a single entry with multiple cost-center lines if your ledger supports dimensions.
  5. True up. When the invoice arrives, reverse the accrual, book the actual, and post any difference to the same cost objects pro rata. Keep the difference visible; a persistent gap means the rate card or the meter is wrong.

Worked example

A company runs three agents in September. Usage metered at contracted rates: Support agent $4,200 (Customer Success), Research agent $2,800 (Product), Procurement agent $1,000 (Operations). No invoices received by September 30.

September 30, accrual

Account Debit Credit Cost center
AI services expense 4,200 Customer Success
AI services expense 2,800 Product
AI services expense 1,000 Operations
Accrued AI services 8,000

October 8, provider invoice arrives for $8,240 (rate card excluded a cached-input surcharge).

Account Debit Credit Cost center
Accrued AI services 8,000
AI services expense 126 Customer Success
AI services expense 84 Product
AI services expense 30 Operations
Accounts payable 8,240

The $240 variance is allocated pro rata and the rate card is corrected for October.

What the auditor will ask

Expect three questions: how usage was measured, how the rate was applied, and how the allocation to cost centers was determined. Each should trace from the GL line to the usage records that produced it. If the answer to any of them is "the engineering team sent a spreadsheet," the control is weak.

Related

What is the journal entry for AI token usage? · How do I do chargeback for AI usage? · Which AI costs need accruals?

Worked examples are illustrative. Account structure and policy should be confirmed with your controller and auditor.

Not sure which of your AI costs are being booked? Run the free Agent Spend Assessment.

Onaro Meridian is FinOps for agentic AI: the system of record that attributes, controls and books what AI agents spend.

Brian Diamond

Brian Diamond

Brian Diamond is a fractional Chief AI Officer and founder of Onaro. He has spent 30 years running infrastructure operations and founded LANStatus, a Connecticut managed services provider and Microsoft partner, in 2001. He holds a Chief AI Officer certification and writes the CAIO Brief on AI leadership for finance and operations.

LinkedIn · CAIO Brief · Author page

Markdown version