Measurement methodology · Updated September 22, 2026
How Weave measures AI cost and engineering output
The definitions, attribution methods, and limits behind an output-based view of AI spending.
What does Weave measure?
Weave connects AI usage and attributed spending with engineering output. Its output unit is an estimated expert hour: the amount of work an expert engineer would need to complete a pull request. Cost per output hour divides attributed AI spending by those output units for the same measured group and time period.
That ratio describes spending relative to observed work. It does not by itself establish incremental revenue, hours saved, or a causal return on investment. Keep adoption, attribution confidence, and code quality alongside the ratio when making an investment decision.
How is engineering output calculated?
Code Intelligence analyzes the source code changes in a pull request. Silk 1 considers intent, complexity, risk, and dependencies across files, and provides reasoning for its estimate. The scoring scale is expressed in expert hours: a score of 3.0 represents an estimated three hours of expert effort.
An output hour is not an engineer's recorded working time. It is not automatically an hour saved by AI, and it does not measure the business value of a feature. PR counts and lines changed can provide context, but they are not substitutes for this output unit. Review the underlying PR and the model's reasoning when a score appears inconsistent with the work.
What costs and attribution signals are available?
Token Intelligence combines configured subscription costs with tool billing and usage telemetry. The level of detail depends on the connected tool and the available data. Its published per-user cost coverage includes Amazon Q, Claude Code, Cline, Cursor, Firebender, Sourcegraph, and Windsurf. For tools without per-user tracking, costs are calculated at the organization level.
The product distinguishes three AI contribution signals:
Scroll horizontally if needed →
| Attribution method | Evidence | Interpretation |
|---|---|---|
| Verified native API | Data from a connected coding tool, such as Cursor or Copilot | Direct tool evidence; availability depends on the integration |
| Inferred commit metadata | Git co-authorship on commits | A signal of assistance, not a complete record of every tool used |
| Estimated time-series inference | Activity windows | Modeled attribution that should remain labeled as estimated |
These are contribution attribution methods. They do not imply that every prompt can be traced to a specific merged change, or that every tool's bill can be allocated to an individual. Shared branches, several tools on the same task, and missing telemetry limit what a comparison can establish.
How does the public Weave Index calculate cost per output?
The Weave Index publishes two related monthly ratios. Both use only engineers with attributed AI spend and merged code output in the same month. Ignored PRs are excluded. Spend that cannot be attributed to a person, such as organization-level overages, is excluded from these ratios.
Scroll horizontally if needed →
| Metric | Calculation | Unit |
|---|---|---|
| AI cost per output hour | Total attributed AI spend ÷ total output for the matched cohort | Dollars per estimated expert hour |
| Return on token spend | Total output ÷ total attributed AI spend for the matched cohort | Estimated expert hours per dollar |
These are ratios of totals, not averages of each engineer's ratio. The second is the inverse of the first when calculated from the same unrounded inputs. The Index's name “return on token spend” refers to output per dollar, not a financial ROI percentage.
An engineer with no observed spend may have disconnected billing data. An engineer with no observed output may work in repositories outside the connected set. Treat missing data as missing, not as zero. The Index's cost-per-employee and output-per-employee series use a broader population, so dividing those displayed averages does not recreate the matched-cohort ratio.
The public matched cohort contains 2,799 engineers in January 2026 and 7,020 in August 2026. It is a changing sample of Weave customer telemetry, not a randomized industry sample. Use the date, cohort definition, exclusions, and available sample information when citing a result. A total company AI budget can be higher than the attributed spend included in this series.
Worked example: cost per output hour
Hypothetical example, not a customer result or benchmark. Suppose a measured group has $1,200 in attributed AI costs and 300 expert hours of merged output during one month. The same group and period must supply both totals.
- Cost per output hour: $1,200 ÷ 300 = $4.00 per expert hour.
- Output per dollar: 300 ÷ $1,200 = 0.25 expert hours per dollar.
- If 60 PRs merged, cost per merged PR is $1,200 ÷ 60 = $20 per PR. This is a different denominator that does not account for differences in PR scope.
If another $200 of organization-level charges cannot be attributed to people, report that exclusion next to the matched-cohort result. A separate full-budget calculation would use $1,400 and yield $4.67 per output hour; label it separately so it is not mistaken for the Index methodology. Neither figure proves that AI created all 300 output hours.
Which quality signals belong beside the cost ratio?
Engineering Intelligence reports review activity, review cycles, reverts, code turnover, and delivery measures. Use relevant signals to check whether output is creating avoidable downstream work. A lower cost per output hour can still be a poor result if changes need more repair or introduce production failures.
Keep quality measures explicit. The Index cost-per-output formula does not apply a hidden quality multiplier. For a team evaluation, define the review or post-merge observation window and show reverts, defects, and rework beside the output total. Recent PRs may not yet have enough history to assess maintainability or production impact.
How should a team evaluate AI ROI?
- Define the decision, such as renewing a tool or changing model use, and choose a consistent time period, repository set, and team.
- Reconcile subscription and usage charges with billing records. Avoid charging the same included usage twice. Record allocations and costs that cannot be attributed.
- Check identity mapping and tool coverage. Show how much spending and output is observed, and which contribution labels are direct, inferred, or estimated.
- Compare similar work before and after adoption, or use a comparable group. Account for staffing, project scope, PR size, and workflow changes.
- Inspect representative PRs and their quality outcomes. Separate measured output changes from assumptions about time saved or monetary value.
If a financial ROI percentage is needed, use (estimated incremental benefit − total incremental cost) ÷ total incremental cost and disclose how the benefit was valued. That requires a defensible baseline and economic assumptions beyond the Index's observed cost/output ratio. Do not equate expert-hour estimates with payroll savings.
Tool-specific measurement guides
- Claude Code: adoption, costs and engineering outcomes
- Cursor: usage, cost and code attribution
- Codex: usage, spending and output measurement
Sources and citation
This methodology describes the public Weave Index, Token Intelligence, Code Intelligence, and Engineering Intelligence definitions reviewed on September 22, 2026. Individual integrations and account configurations can expose different data.
When citing this page, use: Weave, “AI ROI measurement methodology,” updated September 22, 2026, and link to this URL. For a numeric benchmark, cite its individual Index metric page and observation period as well.