Buyer-side · Reconciliation
How to reconcile AI agent usage with provider invoices
Your AI invoice and your usage export are two different documents, produced by two different systems, for two different purposes. They will not match — and each reason they don't match has a shape you can check. This post gives you the three-step method, the seven failure modes we see most, and a free tool that runs the arithmetic in your browser.
The three steps
- Get the billing rule text. Not the pricing page — the rule: what counts as one billable thing, at what granularity. Per resolution? Per operation, or per attempt? Which token buckets, at which multipliers?
- Get reconcilable records. An invoice line and a usage export you can actually join: same window, same keys, line-level. Aggregate totals cannot be reconciled, only compared.
- Recalculate under the vendor's own rules, then mark every row. Each line becomes one of three verdicts:
PASS(recalculation agrees),FAIL(discrepancy with a locatable cause), orUNPROVABLE(the evidence to decide is missing — never guessed).
The third step is where most teams stop early. The value is not the total; it is the row-level verdict list you can hand to a vendor conversation.
Seven failure modes, four classes
From our working checklist (full version with runnable examples):
| Class | Failure modes | Typical symptom | Owner |
|---|---|---|---|
| Precision & buckets | cache reads omitted from the cost export, or rounded per line; input priced as one bucket when it is four | naive tokens × price is off by 10–40% | Engineering |
| Commercial overlays | seats & enterprise contracts; credits, commitments, tax, amortization, FX | invoice lines with no usage counterpart | Finance |
| Timing | month-end revision window | the "final" number changes 7 days later | Finance / engineering |
| Allocation | currency normalization; shared/historical/personal keys | 30–60% of spend reads "unallocated" | Engineering |
Order matters: fix precision problems before arguing about commercial overlays. Otherwise you are comparing a wrong usage number to a right invoice number — and you will lose the conversation on arithmetic instead of winning it on substance.
What a reconciled row looks like
A finished check is boring on purpose. Every row: the invoice amount, your recalculation, the rule applied, the source of that rule, and the verdict. Rows that cannot be decided say UNPROVABLE with the missing evidence named. See a synthetic end-to-end example in the runnable example (cost.csv, usage.csv, reconcile.py — all local, no accounts).
Resolution-based support bills are the sharpest case
Per-resolution AI billing (Intercom Fin, Zendesk AI, and peers) bills outcomes that the vendor's own system marked as resolved. The buyer-side question is not "was the count right" but "who attested this resolution" — and the answer is in the billing rule text. Our rule cards for six AI support vendors live in the repository (registry/vendor-rules.json), each with its source and a confidence grade; the recount tool runs in your browser and uploads nothing. Start with who attested your resolutions for that specific method.
Boundaries, stated plainly
PASS means "recalculation under the stated rule agrees", not "the invoice is fair". Verovo does not recover money and does not promise to. If you want a human look, free measurement runs accept 20–100 line anonymized samples (first response usually within 24 hours).
Reproduce the method yourself
pipx install verovo
verovo recount --list-vendors # six vendors, each rule with source + confidence
verovo demo # synthetic end-to-end, no accounts, no upload
Or use recount.html — same rules, your browser, zero upload. The full seven-mode checklist with commands: ai-invoice-reconciliation.md.
Related
- Who attested your resolutions — the resolution-billing audit method
- FAQ · 常见问题(中文) — the same checks in Chinese, with sourced numbers
- A full synthetic buyer review — PASS / FAIL / UNPROVABLE on a real output