Who can see it?
Anyone with the Team lead, Manager, or Admin role can access the AI report. Its data follows your report visibility settings, so by default a team lead sees the teams they lead and those teams’ members. The Compute tab only appears for roles with Organization metrics access, which managers and admins have by default.Sections
The report is split into tabs: Overview, ROI, Observability, Code review, Models, Tools, Skill library, and Compute. Each tab has a filter bar for the date range and, depending on the tab, team, person, AI tool, and repository. Repository and project filters don’t narrow AI spend.Overview
A summary of AI spend and AI-attributed output for the selected period.- AI output & spend: AI output on the left axis and usage-based token spend on the right, for each period. Seat-based plan costs aren’t included; Cost by fidelity on the ROI tab is the view with the full total. See AI output attribution for how output is attributed to AI.
- Cost per output over time: AI tool cost, including prorated seat costs, divided by AI code output, for all tools and for each tool. The Pre-AI cost per output line shows the price of one unit of output before AI: one hour of expert engineering work at your hourly rate. Use the title dropdown to switch to AI cost over time, which shows spend by tool or by model.
- Cost per person: each person’s AI usage cost for the period, with their cost per output and how much of it is billed versus estimated. Click View details to see the usage rows behind a person’s cost.
- Cost by tool: AI cost by tool shows each tool’s total cost, and Cost per output by tool compares tools on cost per unit of AI output, where lower is better. Flat-rate tools without an included allowance show list-price consumption rather than plan cost.
ROI
The ROI (return on investment) tab shows where your total AI spend goes. Cost by fidelity splits AI spend, including seat cost, by how each dollar is sourced; Team spend flow shows how spend flows from your organization to its top teams; and Projected annual AI spend annualizes AI tool cost at the current spend rate. Admins can also upload a Copilot spend sheet or a ChatGPT Enterprise workspace report from the Upload menu. See AI ROI.Observability
Prompt quality, agent effectiveness, friction, and shipped work for AI agent sessions.- Prompt analytics: overall conversation quality on a 0 to 100 scale, with tabs that break conversations down by complexity, work type, domain, topics, effectiveness, and outcomes, and Conversation prompt quality for each engineer. See conversation prompt quality.
- Developer AI effectiveness: one row per engineer with sessions, autonomy, prompt quality, first-shot success, completion rate, friction rate, an overall effectiveness score, token efficiency, and cost efficiency.
- Session charts: Goals, Outcomes, Friction types & what helped, Languages, Time of day, Tool errors & shipped work, Parallel sessions, Response time, Steering lag, and Work type output.
Code review
How AI tools review code, resolve comments, and introduce bugs. The tab covers review-only tools such as CodeRabbit, Greptile, and Cursor Bugbot, plus the review activity of Copilot, Codex, and Gemini Code Assist.- Summary tiles: Recorded review-tool spend, Suggestions shipped, Cycles replaced, Review output, and $ per cycle replaced.
- Charts by tool: Code review output (switchable to output per dollar), Human review cycles replaced, AI comments addressed by authors, AI suggestions addressed (%), Review cycles by AI tool, and Bugs per 1K AI LOC by tool, which counts bugs attributed to AI-associated code per 1,000 AI lines of code (LOC).
Models
Model spend flow shows how total AI spend breaks down across model providers and their top models. The Model comparison table compares each tool and model on sessions, average cost per session, total tokens, autonomy, first-shot success, token efficiency, average AI output per session, and AI code turnover.Tools
A table of the AI tools that write code, such as Claude Code, Cursor, Copilot, Codex, and Devin, with each tool’s Output, Output / engineer, % of AI output, AI LOC, Cost, Bug ratio, and Feature ratio for the period. Review-only tools appear on the Code review tab instead. Click a tool to open its own report, which adds AI code vs. output over time, AI adoption over time, Agent autonomy, First-prompt success rate, and lines of code modified with AI per engineer.Skill library
Every skill, command, and built-in tool Weave detects across your organization, with prompt quality and invocation counts. Skill session stats shows how the sessions that used each skill performed, as a correlation rather than a measure of causal ROI. See skill library. For how effectively engineers work with AI, see AI skills.Compute
Modal compute spend shows organization-wide metered Modal usage, which you can filter by workspace, app or object, environment, and Modal tag, along with coverage and freshness, a metered breakdown, and monthly Modal billing. It needs a connected Modal workspace, and the tab only appears for roles with Organization metrics access.Provider-specific notes
GitHub Copilot
GitHub Copilot
Copilot usage comes through your GitHub integration. If GitHub reports that your organization has disabled Copilot Metrics API access, enable it by following GitHub’s
Copilot policy documentation.When GitHub returns per-user Copilot usage metrics for your organization, Weave attributes Copilot lines of code to each engineer whose GitHub account is linked. On days without per-user data, Weave falls back to totals for each team synced from GitHub, or else to an organization-wide total. Neither fallback is broken down by person.Copilot lines removed aren’t counted. In the team and organization totals, GitHub reports Copilot Chat code insertions as a count rather than as lines, so Weave counts each insertion as 10 lines added.Once Copilot billing is connected, Copilot cost can be broken down per engineer from GitHub’s billing usage report. See GitHub Copilot for how it’s collected.
Sourcegraph
Sourcegraph
The Sourcegraph usage CSV reports accepted characters rather than lines added or removed, so Weave counts Sourcegraph lines as additions only and converts characters to lines at 50 characters per line.