Skip to content

Cost attribution & chargeback

Use this guide to trace AI cost to the responsible team, project, identity, and budget; decide which records are safe for chargeback; and produce a reproducible FOCUS v1.2 export. Venturi builds the result from cloud billing, provider usage, identity, ownership, and optional request-level evidence, without relying on manual tags.

Start with coverage and unknowns if a large share of spend is unresolved. Use attribution detail to inspect the evidence behind one record before exporting a period.

How attribution works

Venturi links each invocation across the six-layer attribution graph: model invocation → service attribution → code ownership → identity resolution → organizational hierarchy → budget attribution. The engine stores those six layers as five graph node types plus budget edges: Invocation, Service, Project, Identity, and Organization are nodes, and budget responsibility is expressed through billed_to and budgeted_under edges on the Organization rather than as a sixth node. See The attribution graph for the full mapping. Platform Stage A resolves ownership deterministically; RAIL (Resource Attribution & Inference Layer) infers only eligible unresolved ownership relationships in Stage B. A separate Stage C subsystem allocates shared cost without changing ownership.

The ownership paths distinguish direct resolution, RAIL inference, and shared-cost allocation. Platform materialization caps coper at 0.95 before publishing the attribution record with its confidence and evidence.

  • Stage A (Venturi platform) resolves ownership deterministically wherever your data supports it: direct identity and org mappings, service deployments, and code-ownership signals.
  • Stage B (RAIL trained edge-existence inference) infers the most likely owner where deterministic resolution cannot reach, always with a confidence score and an evidence basis. Each Stage B call has a hard 20 ms wall-clock timeout inside the asynchronous processor’s 100 ms P99 Stage B budget; Stage B never runs on the synchronous interceptor path. If inference times out or is unavailable for any other reason, the heuristic baseline answers instead, so attribution processing keeps flowing and the affected results are labeled degraded.
  • Stage C (a separate allocation subsystem) allocates fractional cost across resolved owners, conserving 100% of spend. Allocation-derived attributions are capped at 0.70, or at 0.50 when there is no proportional allocation signal, so they never reach the chargeback band on their own.

Venturi builds the graph from the data your cloud and identity provider already hold. For the full model, see How Venturi works.

Chargeback confidence

Every attribution carries an operational confidence score, coper, between 0.0 and a deliberate policy ceiling of 0.95. Venturi never asserts certainty. Confidence renders both as a number and as a band:

The last column shows whether the band is included in chargeback. Excluded results remain visible for context.

Band Confidence Meaning Chargeback
Chargeback-ready coper >= 0.80 Clears the confidence floor Eligible for export only with qualifying output state, current evidence, and no open dispute or degraded fallback. Reconciliation and Finance acceptance govern final billing.
Provisional 0.50 <= coper < 0.80 One or more candidates, below the billing floor Excluded
Estimated coper < 0.50 Weak signal Excluded

The 0.80 floor governs internal chargeback

coper >= 0.80 is the canonical confidence threshold. It does not establish financial accuracy or replace the other gates: export_eligibility must be eligible, the output state and evidence must qualify, and no dispute or degraded result may remain. Reconciliation and finance acceptance govern publication of the exact export. The threshold also applies to underlying spend in verified-savings calculations, which require separate baseline and acceptance evidence. These checks do not set Venturi’s annual license fee. Stage C allocated attributions are capped at 0.70 where proportional allocation signals are present and at 0.50 where they are not, so they can never enter an accepted chargeback figure; they are shown, but never billed.

Because confidence is calibrated per edge type and per tenant, the score means what it says: a 0.85 is genuinely more likely correct than a 0.62. The distribution of your attributions across these bands is visible on the dashboard, so you can see how much of your spend is chargeback-eligible before you export anything.

Unknown spend is reported, not guessed

Venturi distinguishes six output states and surfaces them rather than guessing:

  • deterministically_resolved and strongly_inferred: confident ownership.
  • bounded and ambiguous: a constrained set of possible owners.
  • unknown and not_identifiable: Venturi can see the spend but cannot responsibly assign it.

Unknown spend is never silently allocated and never shown as $0.00; an unknown cost renders as an em-dash, not a fabricated zero. You investigate coverage gaps through the Unknown Spend Analyzer instead of discovering them mid-audit.

The chargeback export

When you are ready to bill, export the FOCUS v1.2-conformant chargeback file: 64 columns (57 base plus 7 Venturi extensions such as ownership state, capture feasibility, time basis, and carbon grams), required cost fields always present, cost computed on the model actually served, and only export-eligible rows included: chargeback-ready (coper >= 0.80) results that no degraded condition or heuristic-baseline fallback produced. A chargeback file never contains a sub-floor row, so you never bill a team on an attribution you could not defend.

The export’s full treatment (column reference, API request, scheduling, and formats) lives in Reporting & exports.

Reproducible audit trail

The export is built so your auditor can reconcile the chargeback figure to your cloud bill without Venturi’s involvement. Each row reconciles to a billing-period figure, carries its ownership state and confidence, and links back to the evidence that produced it. That audit independence is the point: finance accepts numbers it can verify itself.

For dashboards, scheduled delivery, and the full column list, see Reporting & exports.

Chargeback-line integrity

A chargeback only works if finance can defend it line by line. Venturi builds every exported line so it stands on its own under scrutiny: each line tells you where its cost came from, every dollar on the invoice is accounted for, and any line you bill on can be traced back to the evidence behind it. Below is what each line carries and how to read it.

Every line carries its cost basis

Each chargeback line states whether its cost is estimated or reconciled, and surfaces the controlling reconciliation factor for that line’s provider, model, and period, along with where that pairing sits in its reconciliation lifecycle (ESTIMATED, PARTIAL, or RECONCILED). You always know whether a number is a pre-bill estimate or a figure that has already been trued up against the provider invoice, so an early-month estimate is never mistaken for a settled charge. Reconciliation defines each of these terms.

Reconciliation

Venturi keeps two cost fields for every invocation and never merges them:

  • Estimated cost is computed from tokens and pricing within seconds of the invocation. A line that carries it is labeled ESTIMATED in x_TimeBasis.
  • Reconciled cost is the same spend trued up against the provider’s billing data. Reconciliation targets completion within 24 hours (Service objectives), and a line that carries reconciled cost is labeled RECONCILED in x_TimeBasis.

Each provider, model, and period moves through a reconciliation lifecycle based on its chargeback-eligible lines:

Lifecycle state Meaning
ESTIMATED No eligible line in the group carries reconciled cost yet.
PARTIAL Some, but not all, eligible lines carry reconciled cost.
RECONCILED Every eligible line carries reconciled cost.

Two checks use the same ±5% band:

  • Reconciliation factor. For each provider and model, the factor compares the reconciled invoice amount with the pre-invoice estimate over the monthly window. Its validation band is 0.95 to 1.05 (±5%), and a factor outside that band raises an investigation alert.
  • Control total. For each provider, model, and period, the export carries a control-total block: the chargeback-eligible lines, the named excluded cost, and the unattributed residual (the three buckets below) must sum to the reconciled invoice amount within ±5% of that amount. You can recompute the control total from the exported rows alone.

A chargeback report cannot be published while its reconciliation is still ESTIMATED or PARTIAL (see approval thresholds).

Complete invoice disposition

For each period, provider, and model, Venturi shows a complete disposition of 100% of reconciled spend in three explicit, labeled buckets:

Bucket What it is In the chargeback?
Chargeback-eligible Reconciled lines at coper >= 0.80 with export_eligibility=eligible, qualifying output state, and current evidence. No open dispute or degraded fallback may be present. Included
Excluded Reconciled known cost with a supported attribution candidate that fails an eligibility gate. Exclusions cover provisional confidence, ineligible output state, stale evidence, open dispute, or degraded fallback. Each line is named once with its exclusion reasons; not billed
Unattributed residual Known reconciled invoice cost without a supported attribution, including Estimated (coper < 0.50), unknown and not_identifiable records Never billed as attributed cost; named and disposed of by the residual-disposition policy you choose below

Each reconciled known-cost dollar belongs to exactly one bucket; Finance acceptance governs publication and final billing of the reconciled export rather than row eligibility. A line that fails several gates appears once with all applicable reasons. Unknown per-record cost remains unknown; the residual is the known reconciled invoice remainder, not a fabricated value for missing usage.

The eligible lines, the named excluded cost, and the residual sum to the reconciled invoice within the ±5% control-total band defined in Reconciliation, so the chargeback foots to the bill independently. You choose the residual-disposition policy: leave it unallocated and footnoted, route it to a named shared cost center, or split it by a rule you set. Your choice is recorded with the export and written to an append-only audit row.

No silent re-labeling

No disposition ever relabels excluded or unattributed cost as attributed, and no sub-floor or unattributed dollar can appear inside a chargeback-ready line. A finance reviewer who asks “what about the other 18%?” sees exactly where it went, by name, never a number that quietly covers only the easy spend.

Calculation receipts

Each chargeback-ready line links, in both the export manifest and the dispute surface, to a calculation receipt: the provenance behind its confidence. The receipt binds the line to the corpus and decision threshold that scored it, and for that line’s edge type and tenant it states the calibration method used, the size of the calibration set, and the band’s measured calibration error, alongside the two-layer evidence basis. It also says, in plain language, what the confidence band means. If a line’s calibration provenance cannot be assembled into a receipt, it is not exported as chargeback-ready; confidence you cannot trace is never billed.

Calibrated, and capped on purpose

The receipt is why a 0.85 is genuinely more trustworthy than a 0.62. Confidence is still capped at 0.95 (Venturi never claims certainty), and Stage C allocation-derived lines remain capped at 0.70, or at 0.50 when there is no proportional allocation signal, below the 0.80 billing floor, so they appear with full evidence but never enter a billed figure.

Served-model costing

When a request is routed to a different model than the one asked for, Venturi costs the line against the model that actually served it, never the requested one, and flags the route divergence on both the line and the dispute surface. A line whose cost basis is a requested model that differs from the served model is never emitted as a settled charge. So when a department head cross-checks the provider console, the model and the dollars match.

Dispute restatements

If a dispute on a line you already billed is overturned, Venturi does not quietly rewrite history. The originally delivered export stays immutable. Instead, the next export carries a superseding adjustment line: a signed delta that credits the previously-billed owner and debits the corrected one, linked to the original line, with the full override lineage: who changed it, when, and why. The correction to Venturi’s attribution graph and the correction to your bill always travel together, so a right answer inside Venturi never leaves a wrong charge standing outside it.

Fiscal-period finalization

You can mark a chargeback period finalized in step with your fiscal calendar. Before finalization, in-period corrections recompute freely and are clearly labeled provisional. After finalization, the exported figures and their manifest are immutable: any later change (an overturned override, an allocation refinement, a reconciled-cost true-up) is never silently folded back into the closed period. It is carried forward as an explicit, audited amendment line in the next open period, linked to the original period, record, and reason.

Why a closed period stays closed

The canonical accounting trust-breaker is a closed number that quietly changes months later. Venturi’s “March” figure does not move in May. If something needs correcting, you see an amendment with a visible marker and a clear reason; the two ledgers stay reconciled and every closed-period change remains visible.

Energy & carbon, attributed too

The same graph attributes energy and carbon alongside cost. Every priced invocation carries an estimated energy figure (Wh) and carbon figure (gCO2e) drawn from a catalog of covered models, exported in the FOCUS file as x_EnergyWh and x_CarbonGrams, and shown in the attribution detail and in the Sustainability domain’s Your energy & carbon view. For a model that is not in the catalog, Venturi returns null, never zero, so an unmeasured footprint is labeled unmeasured rather than counted as no footprint.

From connection to first chargeback

  1. Connect your cloud. The read-only connector alone produces account- and team-level attribution.
  2. (Optional) Send request-level events for per-call resolution: tokens, cost, latency, and the identity behind each call.
  3. Watch coverage build; review your confidence distribution on the dashboard.
  4. Export the FOCUS chargeback file and hand it to finance.

Reconciliation of attributed cost against provider billing targets completion within 24 hours (see Reconciliation and Service objectives). Beyond that, time to first reviewed chargeback depends on source coverage, invoice anchors, and your approval workflow.

Give feedback on this page

Sign in to send feedback. Opens in a new tab.

Copy page details 
Docs feedback: Cost attribution & chargeback
Page: https://docs.venturi.systems/capabilities/cost-attribution-and-chargeback/