Home / Standards / AIGS 200
Exposure Draft

AIGS 200 Cost and Token Accountability

Every unit of AI consumption is attributed to an owner, a purpose and a budget. Spend that cannot be attributed is measured and reported, not hidden.

Document
AIGS 200, Exposure Draft ED-2026-200
Status
Founding draft, open for member comment
Comment period
60 days from publication
Working Group
TWG-200 Cost

1. Objective

To make AI consumption, measured in tokens, credits, messages or currency, attributable and defensible, so that an organization can say who spent what, on what, and whether it was within budget.

2. Scope

All metered AI consumption: model API tokens, platform credits and capacity packs, per-seat AI licences, and consumption by agents and automated flows, whether billed directly or through a cloud provider.

3. Key terms

Consumption event. A recorded unit of AI usage reported by a provider or platform.

Attribution basis. What an event is attributed to: the AI system, the owner only (when the platform reports no system), or nothing.

Unaccountable spend. Consumption that cannot be linked to an AI system with an owner of record under AIGS 100.

Collection freshness. How recently consumption data was successfully collected, as distinct from when consumption last occurred.

Consumptiontokens, credits By AI systembest attribution By owner onlylabeled as such Unattributedcounted, not hidden Budget,period to date Unaccountablespend % Owner andboard report
Attribution of AI consumption under AIGS 200.

4. Requirements

200.1The organization shall collect AI consumption events from every material provider and platform, at a frequency of at least daily where the provider supports it.
200.2Each consumption event shall be attributed to the AI system that generated it. Where the platform does not report the system, the event shall be attributed to the identifiable owner and labeled "owner basis". Where neither is available, it shall be recorded as unattributed. The attribution basis shall be shown wherever the figure is reported.
200.3When events are grouped, they shall not be merged under an empty or missing name, and an aggregate shall never credit one owner with another owner's consumption.
200.4The organization shall report unaccountable spend: the share of consumption not linked to an AI system with an owner of record.
200.5Budgets shall be set per period and scored against consumption within the same period. Lifetime totals shall not be presented as period-to-date figures.
200.6Owners shall receive alerts at defined thresholds (proposed: 80% and 100% of budget) before the period ends.
200.7Consumption shall be converted to currency using a documented price per unit. Consumption without a known price shall be reported as unpriced, not as zero cost.
200.8Reports shall state collection freshness, so that "no recent consumption" can be distinguished from "collection has stopped".
200.9Per-seat AI licences shall be reconciled at least quarterly against active use, and inactive seats reported to their budget owner.

5. Evidence

  • Consumption report by AI system and owner, with attribution basis.
  • Budget register and alert history.
  • Unaccountable spend trend over at least three periods.
  • Price table with sources and effective dates.
  • Licence reconciliation records.

6. Metrics

MetricDefinition
Unaccountable spendConsumption not linked to an owned AI system, divided by total consumption, per period.
Attribution qualityShare of consumption attributed by system, by owner only, and unattributed.
Budget adherenceOwned systems within budget, divided by owned systems with budgets.
Seat utilizationActive AI seats, divided by licensed AI seats.

7. Basis for conclusions

Total consumption is already visible in most provider consoles; what organizations lack is accountability for it. The draft therefore centers on unaccountable spend rather than total spend. Requirements 200.3, 200.5 and 200.8 address failure modes observed in practice: platforms that omit the system name, budgets scored against lifetime totals, and silent collection failures that look like falling usage.

8. Questions for respondents

  1. Are 80% and 100% the right default alert thresholds?
  2. Should owner-basis attribution be acceptable indefinitely, or only as a transition state?
  3. How should shared or platform-wide consumption be allocated?
Comment on AIGS 200All standards