Unit of work
A unit of work is the consistent business item used to count and compare completed work. In AI measurement, it connects model or tool activity to an operational result, such as a resolved support case, reviewed contract, reconciled invoice, approved campaign asset, or merged code change.
The unit matters because prompts, tokens, sessions, and generated outputs describe AI activity. They do not show how much usable work the organization completed.
How do you choose a unit of work?
Choose the smallest item that represents a meaningful completion event for the workflow owner. It should have a clear start and finish, an acceptance rule, an accountable system of record, and enough volume for comparison.
The same AI tool may require different units in different workflows. A support team may count cases resolved to its quality standard. Finance may count reconciliations completed without material correction. Legal may count reviewed agreements by complexity band. The label should follow the work, not the software.
Write the definition in a metric dictionary. Specify what qualifies, what does not, how reopened or duplicate work is handled, and which source records prove completion. Data lineage should make those rules traceable from the reported metric back to the workflow system.
Why is a unit of work useful for AI value measurement?
A stable denominator supports comparisons across cost, speed, quality, and capacity. Teams can calculate:
Cost per completed unit = attributable workflow cost / accepted completed units
Cycle time per unit = completion time - start time
They can also compare correction, escalation, or defect rates per unit. These measures expose tradeoffs that a usage dashboard misses. A workflow may generate more outputs while completing fewer accepted units because review and rework increased.
The FinOps Foundation defines unit economics as a way to connect technology spending with the value it creates. Its guidance notes that generative AI measures can progress from cost per token toward outcome-oriented units such as cost per assist, agent action, or case deflected (FinOps Unit Economics). Its AI cost forecasting guidance also uses cost per unit of work as an outcome measure (FinOps AI cost forecasting).
How is a unit of work used in an AI pilot?
An experiment design should name the unit before the pilot starts. The baseline and pilot must use the same inclusion rules, quality threshold, and complexity treatment. Otherwise, the comparison may reward the AI workflow for receiving easier work.
Complex workflows may need a normalized unit or separate cohorts. A short routine case should not carry the same weight as a complex escalation unless that choice reflects the business decision. Keep the raw count alongside the normalized view so stakeholders can inspect the adjustment.
The measurement window must also capture enough units to represent ordinary variation. Report the count alongside the average. A large apparent change based on very few completed units should remain provisional.
Unit of work versus AI output
An AI output is something the system produces. A unit of work is something the organization accepts as completed under a defined rule. One completed unit may require several model calls, tools, human reviews, or retries. One model response may also support only part of a larger unit.
This distinction lets teams allocate total cost of ownership and evaluate AI cost optimization against accepted work. Lower token cost is useful, but it does not prove lower cost per completed unit.
Common mistakes
Weak definitions count every output, change the unit during the pilot, ignore reopened work, combine incomparable tasks, or rely on a record that teams do not update consistently. Another mistake is choosing a unit because it is easy to count rather than because it supports an investment decision.
A useful unit of work gives finance, technology, and the workflow owner one shared denominator. If they cannot agree on what counts as completed, any ROI calculation built above it will remain fragile.