Ask five people in AI cost management what "reconciliation" means and you will get five answers, most of which describe estimation with extra steps. That is not carelessness — it is a vocabulary gap. The industry's standard glossaries define allocation, showback, chargeback, and forecasting in careful detail, and then stop short of the one operation finance actually closes the books on: tying spend back to the invoice document itself.
This page defines the term we use for that operation, precisely enough that you can check whether any tool — including ours — actually performs it.
Deterministic invoice reconciliation is the independent recomputation of AI spend from usage quantities and an effective-dated rate card, compared — by fixed rules, with no statistical model or language model in the calculation path — against the provider's cost report and the invoice document, at the finest grain the invoice discloses, with every unmatched amount surfaced as a first-class output.
Each phrase is load-bearing, so let's take them in turn.
Independent recomputation. The cost figure is derived from raw usage quantities (tokens, requests, cache reads and writes) multiplied by rates, not read back from the provider's own cost totals. This distinction matters because some tools ingest the provider's cost reports and display them back — a mirror of the provider's ledger. A mirror is internally consistent by construction: it agrees with the provider because it is the provider's number. Reconciliation verifies instead: agreement is a result, not an assumption, and disagreement is a finding.
Effective-dated rate card. Provider prices change, sometimes mid-month, which means one billing period can span two rate regimes. A reconciliation engine must know which rate was in force for each unit of usage, not just the current price sheet. Cache writes are a good example of why: on Anthropic, a cache write is billed at a multiplier of the base input rate that depends on the cache's time-to-live — 1.25× for five-minute cache writes, 2.0× for one-hour writes. A recomputation that ignores effective dates or multiplier structure produces an estimate, whatever it is labeled.
Deterministic. Same inputs, same output, every time — the comparison is arithmetic and rules, reproducible by a third party from the same data. No language model touches a number in the cost path. This is what makes a reconciliation result checkable rather than persuasive.
At the finest grain the invoice discloses. Providers bill in line items and buckets, not per request. A reconciliation claim is honest only at the grain the billing document itself supports; claiming finer precision than the invoice exposes is extrapolation. This is also where our own claim gets its wording: "Reconciled to the cent at the finest grain your provider's invoice discloses, with any residual shown — never hidden."
Every unmatched amount surfaced. Unrecognized billing lines, usage without a price, spend your telemetry never saw — these are outputs of the reconciliation, not embarrassments to be absorbed into a total. A reconciliation that silently folds its exceptions into "other" has quietly become an estimate again.
In practice the operation splits into two questions, and it helps to name them separately:
A tool can be strong at Level A and never attempt Level B. Both are needed before "reconciled" is the right word for the result.
A definition is only useful if it classifies the hard cases. In our implementation, every dollar of AI spend lands in exactly one verification tier: T1 — Reconciled (recomputed per request against effective-dated price tables and tied to the invoice total), T2 — Invoice-verified (confirmed against invoice totals but not attributable per request — fixed subscriptions, committed pools), T3 — Management-declared (company-asserted values, not independently confirmed), and Uncovered (spend seen but not placeable in any tier, disclosed as context and never presented as reconciled).
Three distinct quantities then describe how the total ties out: coverage (the share reconciled per request), residual (the reconciliation error on tracked spend), and untracked (invoice spend your telemetry did not capture). A near-zero residual does not imply full coverage — they answer different questions, and a serious reconciliation reports both.
A term earns trust by stating its own limits, so here are ours. First, reconciling an invoice bucket to the cent does not prove how cost is attributed to features or customers inside that bucket — offsetting errors can misattribute spend while the bucket total still ties out perfectly. Attribution is a separate discipline with separate evidence, and "to the cent" is never the right phrase for it. Second, the claim is provider-specific by nature: our evidence base is a real Anthropic invoice (June 2026, list prices, reconciled with the residual shown and explained on our methodology page). Our OpenAI reconciliation is implemented but has so far been verified only against a synthetic oracle — a test fixture with known correct answers — so the to-the-cent claim does not yet extend to OpenAI, and we say so rather than blur it.
When a vendor or an internal tool says "reconciled," test the word: Is the cost recomputed independently, or mirrored from the provider? Is the rate card effective-dated? Is the comparison deterministic and reproducible? Is the grain the invoice's own grain? Are exceptions first-class outputs? Five yeses and the word is earned. Anything less is estimation — possibly excellent estimation, but a different operation with a different evidentiary weight, and your close process deserves to know which one it is getting.
ToTheCent is our attempt to make this definition boring: a monthly artifact your finance team can rely on, rather than a term of art that means whatever the dashboard wants it to mean.