Completed work
Completed work in AI measurement is a unit of business work that has reached a predefined acceptance state after any required AI generation, human review, correction, approval, and system update. It is the output an organization can count when measuring throughput, cost, quality, or value.
An AI response, draft, prediction, or agent action is not automatically completed work. Completion depends on the workflow's rules. A support case may count only after resolution and closure. A contract review may count after an approved disposition is recorded. A software change may count after review, testing, and merge.
How should completed work be defined?
Start with a clear unit of work. Name the source record, the completion event, the acceptance criteria, and the treatment of reopened, rejected, duplicate, or partially completed items.
The definition should be observable in a system of record. If a team must reconstruct completion from memory or survey responses, the metric will be difficult to audit. Record who owns the definition and when it changed, because a revised acceptance rule can break comparisons across periods.
The Scrum Guide offers a useful general principle: work becomes part of an increment when it meets the Definition of Done, which provides a shared understanding of what was completed (Scrum Guide). AI-enabled workflows need the same clarity even when they do not use Scrum.
Why does completed work matter for AI measurement?
Completed work gives operational metrics a defensible denominator. Teams can calculate accepted throughput, cycle time, cost per unit, rework rate, or acceptance rate without counting every model output as value.
This distinction matters because AI can increase activity while failing to increase usable output. More prompts, generated drafts, automated steps, or active users may describe adoption. They do not prove that more work reached an accepted business state.
AWS advises organizations measuring AI ROI to choose a business-value metric that cannot be easily manipulated. Its example favors bugs remediated over lines of code because output volume can be inflated without improving the business result (AWS guidance on AI ROI).
How is completed work measured?
Count unique units that satisfy the acceptance rule during the measurement window:
Completed work = count of unique accepted units completed in the period
For AI impact analysis, compare completed work across a credible baseline and AI-enabled period, or between comparable treatment and control groups. Segment materially different work by complexity, channel, team, or case type. A higher count may reflect easier work, higher demand, or changed staffing rather than AI.
Track quality and exceptions with the count. Reopened work, reversals, defects, escalations, and delayed corrections can reveal that apparent completion was premature. Microsoft recommends using canonical metric names, plain definitions, source systems, and a small set of measures tied to the business outcome (Microsoft agent metrics reference).
How does completed work connect to business value?
Completed work is an operational result, not financial value by itself. It supports AI unit economics when the organization attaches attributable cost or benefit to each accepted unit.
A higher completion rate may improve AI time to value if accepted outcomes appear sooner. It may support cost avoidance if the organization can show that the added capacity prevented a specific future expense. Neither conclusion should be assumed from the completed-work count alone.
Common completed-work mistakes
Common failures include counting generated outputs, mixing work types, excluding review, changing the definition mid-period, counting duplicate records, and ignoring reopened work. Another mistake is treating completed work as proof of ROI without connecting it to attributable cost or benefit.
A completed-work metric becomes a vanity metric when the acceptance rule is weak or the number cannot change a decision. The useful version shows what finished, at what quality, at what cost, and under which operating conditions.