How the Agent Economy Index is calculated
The Agent Economy Index (AEI) is an inferred estimate of the extent of agentic economic activity, informed by the sources and research published here. It tracks transaction delegation and emerging disruption signals by market and sector.
AEI · METHODOLOGY 1.0
Measure delegation, not AI enthusiasm.
The Agent Economy Index follows Rob Dixon’s founding article (15 September 2026). It weights the value of transactions by who searches, chooses and executes the deal. Current coverage is consumer commerce; it is not a whole-economy GDP measure.
| Tier | Who does the deal? | Weight |
|---|---|---|
| T1 | Human to human | 0% |
| T2 | Humans using deterministic software | 0% |
| T3 | Agent researches; human executes | 10% |
| T4 | Agent chooses and executes on one side | 50% |
| T5 | Agents choose and execute on both sides | 100% |
Economic activity and GDP
Current AEI readings use consumer transaction-value inferences, not GDP value-added data. They cannot be interpreted as the percentage of GDP executed by agents. A GDP measure would require value-added attribution that avoids counting the same supply-chain value multiple times. A strict Tier 5-only GDP measure would reach 100% only when all covered value added is generated through agent-to-agent execution, with no human in the execution loop. That is different from this dashboard’s normalised index.
What today’s percentage means
This is the paper’s normalised index, not the literal percentage of all spending delegated to agents. The denominator discounts T3 and T4 as well as the numerator. An economy entirely in T3 therefore scores 100%. Discovery alone can cross 1% AEI at approximately 9.17% T3 share, so the paper’s claim that 1% necessarily proves execution does not follow from its formula. This dashboard separately reports T4 + T5 and requires execution evidence for confirmed take-off.
How the starting figures are treated
The five geographic baselines are the low-confidence inferences from July, not independent published measurements. We preserve T1, T3, T4 and T5, and balance T2 as 100 minus the other tiers because the rounded table does not sum to 100%. We calculate the formula consistently rather than copying its rounded headline. The world baseline becomes approximately 0.17%, versus approximately 0.15% in the paper. No confidence interval is claimed because no statistical error model exists.
What can update the AEI
An eligible observation needs transaction value, a matched denominator, a defined market and sector, a reference period, and mutually exclusive tier classification. Count each deal once at its highest executed tier. Gross merchandise value, payment volume and revenue cannot be combined without reconciling scope, currencies, refunds and overlap. Human-confirmed checkout is not automatically tier 4. Surveys, visits, infrastructure launches and pilots remain supporting signals.
Countries, sectors and missing data
No regional estimate is copied into a country. No country estimate is copied into an industry. Global consumer surveys are not allocated geographically. Industry signals are not added together to calculate a country total. A valid aggregation requires non-overlapping transaction-value weights in a common period. Until those exist, cells remain unmeasured.
Weekly monitoring and provenance
A scheduled research task checks primary sources every Monday morning in Europe/London, appends a dated snapshot, validates the data and republishes this site. Source publication dates, observation periods and review dates stay separate. Unchanged estimates are carried forward with the original observation date. Failed checks do not imply no change. If more than nine days pass without a research update, the dashboard displays “Update overdue”. Publications may be monthly or quarterly even though research runs weekly.
Doubling and disruption
Observed doubling uses elapsed months × ln(2) / ln(latest AEI / initial AEI) on comparable measured observations. It needs three unique observations across at least 91 days and a minimum 0.05% starting AEI. Two doublings require a distinct intermediate observation at least twice the initial value and a final observation at least twice the intermediate value. The growth model is descriptive, and the warning thresholds are research hypotheses rather than empirically validated probabilities. Estimated phase labels are not confirmed disruptions.
Source access and reproducibility
All evidence cards link to the source publication, retain the reported metric and describe its limitations. Download the complete data or use Export history for the tier inputs. The published founding article is linked above; the original July draft remains archived with the dataset. Reference sources, readings, archived documents, and research logs are stored in a versioned database. Each published dataset is validated and retained, and downloads use the current database revision.