Usage-based billing changes the timing of your close
Historically, most software costs closed themselves. A seat licence is billed at a known amount on a certain date. Things are predictable, accruals and postings can be automated or included in the month-end checklist. However, as AI quickly became a significant cost item, we are moving into an environment API calls, inference and token consumption are usually billed based on usage and in arrears: you consume in September, and the invoice for September arrives later, usually sometime in October.
That gap is small while spend is small. But as you've probably already noticed, this type of spending is growing quickly, and it's very likely here to stay. If usage is not tracked and posted correctly, your P&L could tell a completely false story. How can you practically avoid this in a manner that is scalable as a bigger portion of spend falls under this category? The fix is simple and uses tools you already own— it just needs to become a deliberate part of the month-end routine, not an afterthought.
Why "book it when the invoice arrives" drifts
Booking AI costs on invoice receipt means each month's P&L carries the previous month's usage. When usage is flat, that roughly washes out. When it isn't, it doesn't.
On a growing usage line, which describes most AI spend right now, expense is consistently understated, because every month you are recognising a smaller, older month. If usage drops, say after a model switch or a feature change, the effect reverses and expense is overstated. Either way, gross margin, departmental budgets and unit economics all move for reasons that have nothing to do with the business. The numbers start to look like noise, and you can easily miss key warning signs or (possibly worse) opportunities. The feedback loop is broken.
Accrue from the vendor's own usage data
The good news is that usage-based vendors tell you what you owe before they bill you. Most provide a usage dashboard, a cost export or an API that reports consumption and cost to date. That is your accrual basis.
At month-end, pull the period's usage for each material usage-based vendor and book an accrual: debit the relevant AI expense account, and credit accrued liabilities. This is your opportunity to also apply the cost allocations, and any other rules you require. Set it to reverse on the first day of the next period. When the invoice arrives, it posts normally against the reversed accrual. Be sure to apply the exact same allocation rules as you did for the accrual. If your providers' usage data is reliable, which it typically is, the invoices should match the accrual to the cent. If not, the difference is likely going to be very small.
A few details help you line the accrual exactly with your invoice:
- Use the period's usage, not the running balance. Set the correct start and end dates at the period's boundaries, in the vendor's billing timezone if it differs from yours.
- Apply your contracted rates. Dashboard list prices may not reflect committed-use discounts, tiered pricing or credits. Understand what you're working with, adjust once if needed, document and reuse.
- Treat prepaid credits separately. If you bought credits upfront, that's a prepaid asset, it's different from the accrual approach. In this case the prepaid balance is drawn down by the usage. Ensure you're posting the asset upfront, and drawing against it to the P&L.
- Set a materiality threshold. Accrue the vendors that matter first, and expand to the small, stable ones.
Your existing system already handles this
None of this needs new software. QuickBooks Online, Xero and every mid-market ERP support reversing journals, accrued liability and prepayment accounts, as well as cost attribution and vendor-level coding. What's usually missing is the process: someone named to pull usage for each vendor, a fixed point in the close calendar when it happens, and an agreed way to post it.
Where it helps to automate the data pull, that's a reasonable next step. VP Spend, for example, pulls usage and cost from AI vendors on a schedule so the accrual starts from data that's already collected and coded. But the method is the same either way, and a well-kept spreadsheet with a clear owner gets you most of the way there.
Keep a trail back to consumption
Flat-fee costs are easy to evidence: the contract and the invoice agree. Usage-based costs move every month, and eventually someone, whether an auditor, a board member or a new controller, will ask how you know the number is right.
The answer should be a short trail: the usage export or screenshot as of the cutoff date, the rate applied, the resulting accrual, and next month's comparison against the actual invoice. Save the export with the close file rather than relying on being able to regenerate it later; vendor dashboards change, and historical views aren't always kept.
That monthly comparison is also your quality check. Track the accrual-to-invoice variance per vendor. If it stays small, the method works. If it keeps landing in one direction, adjust the method, not just the number. (How the underlying records are structured so this reconciles cleanly is its own topic, covered in our companion piece on the AI spend subledger.)
A lightweight month-end checklist
If you want to start with this close:
- List your usage-billed AI vendors and mark which ones are material enough to accrue.
- Name one owner to pull period usage from each vendor's dashboard or API on a fixed close day, cut off at month-end.
- Book reversing accruals at contracted rates, coded with each vendor's default dimensions.
- Save the usage export with the close file as supporting evidence.
- Review accrual-to-invoice variance each month and refine the estimation method when a pattern appears.
A close that reflects what actually happened
Usage-based AI spend isn't hard to account for. It just runs on a different clock than its invoices, so the close should follow the usage. Once that routine is in place, the monthly number reflects what the business actually consumed, and it holds a clear line back to what happened in the operation. You spend less time figuring out whether the numbers are right, and more time understanding what they're telling you.