Measure AI training ROI by comparing verified improvements in completed work with the full cost of the program over the same period. Include review and correction time. Keep employee capacity separate from cash savings, and state what the evidence can establish about training's contribution.
At Ajaia, we look at employees' comfort, the quality of their work and their amount of usage. These measures answer different questions. Someone can feel more confident while producing an unreliable answer. Another employee can produce excellent work in a lab but struggle to repeat the task afterward. A useful readout shows those differences so the sponsor knows what to do next.
Start with a task you can observe before and after training
Choose a recurring task, its eligible users and its review owner before the pilot. Define what makes the completed output acceptable. Then collect a comparable baseline with the team's current method.
For an HR policy-answering workflow, the criteria might require every policy claim to match the approved source, eligibility conditions to remain intact and unanswered questions to be escalated. Record the time needed to find the source, draft the answer, check it and make corrections. If the AI-supported version produces a faster draft but takes longer to verify, the baseline should expose that.
Use several comparable examples where practical and record the sample size. A single unusually easy task is a weak basis for projecting the team's future performance. Note differences in task difficulty, policy versions, tool access and review requirements.
The editable training scorecard provides a place to record those assumptions. Assign someone to collect the follow-up evidence and agree on the observation period before the first session.
Measure comfort with specific questions
Comfort affects whether someone will attempt the task and ask for help when they get stuck. A confidence survey is self-report; it cannot establish task competence on its own.
Ask the same task-specific questions before and after training. For example: "How comfortable are you drafting a policy explanation in the approved tool?" Follow with: "How comfortable are you checking that answer against its source?" Use a consistent response scale and report how many eligible people responded.
Also ask what would make the next attempt easier. The answer might be a clearer data rule, an example for their role or a private place to ask questions. Those responses help shape follow-up support.
Do not convert a confidence increase directly into financial value. A learner who feels ready still needs to demonstrate the task. Pair the survey with an exercise so the readout distinguishes perceived ability from observed performance.
Test quality on a fresh example
Give learners equivalent before-and-after tasks using the same review criteria. Include a source gap or an ambiguous detail so the test requires judgment as well as drafting.
Record whether each output meets the threshold and which errors appear. For example, an answer that invents a payout rule fails the policy task even if its tone is excellent. Have the reviewer assess source accuracy and handling of unknowns before judging speed.
Use a fresh input at follow-up. Repeating the facilitator's demonstration can show recall; a new example shows more about whether the learner can apply the method. Record how much assistance was needed and whether corrections were completed before the output was used.
Keep quality visible in the executive readout. A speed improvement that comes with more unsupported claims is a reason to revise the workflow, not a benefit to celebrate.
Use analytics to understand repeated use
Enterprise tools can provide useful adoption evidence. The available metrics depend on the product, plan, permissions and reporting setup, so confirm what your organization can access.
For example, ChatGPT Enterprise and Edu workspace analytics reports adoption and usage signals such as active users and message counts. Its impact survey provides self-reported outcomes; OpenAI explicitly distinguishes those signals from causal ROI measurement. Microsoft also provides a Copilot adoption report with its own prerequisites and definitions.
Define the denominator. Active users as a share of people with usable access answers a different question from active users as a share of all employees. Use the same period and eligibility rules when comparing cohorts. Record reporting lag and missing data where they affect interpretation.
Usage alone cannot show that the selected workflow improved. Pair it with a small sample of reviewed work and a conversation about what employees are using the tool for. A higher prompt count can reflect repeated attempts to fix a poor answer.
Explain the purpose of measurement to participants and use the organization's approved reporting and privacy practices. A team-level view may be enough to decide which support to provide; collecting everyone's conversation text is not a prerequisite for assessing the training.
Calculate value from the whole workflow
For a task that meets the agreed quality threshold, estimate released capacity as:
Eligible completed tasks × minutes saved per task ÷ 60 = hours released.
Count the tasks where the method was actually used successfully. Do not multiply every eligible employee by every possible task while assuming universal adoption. Include setup, review and correction in the completion time.
To calculate ROI percentage, use: (recognized benefit minus total program cost) ÷ total program cost × 100. If cost is zero or unknown, report ROI as unavailable. Use benefits and costs from the same measurement period.
Program cost can include trainer fees, participant time, preparation, incremental licenses, management, review and follow-up support. Avoid counting the same cost or saved hour in multiple workflows. Agree with finance on which benefits can be recognized and which remain scenarios.
Fictional example: capacity value versus cash savings
This example is fictional and contains no client results. Suppose 20 employees each complete 30 eligible tasks per month. Comparable tasks take 20 minutes before training and 14 minutes afterward, including review and correction. Both samples meet the same quality threshold, and all 600 monthly tasks in this scenario use the tested method successfully.
The estimated monthly capacity released is 20 × 30 × 6 ÷ 60 = 60 hours. At an illustrative loaded labor rate of $50 per hour, that represents $3,000 of capacity value per month, or $9,000 over three months if usage and results persist. If total three-month program cost is $12,000, the capacity-value return is ($9,000 − $12,000) ÷ $12,000 = −25%.
Change the assumptions and the result changes. If review and correction bring the new completion time to 18 minutes, the monthly capacity released falls to 20 hours. At the same rate, three-month capacity value is $3,000 and the capacity-value return is −75%. If only half the tasks use the method, the original 60-hour estimate falls to 30 hours.
Neither calculation establishes cash savings. Payroll may be unchanged. Finance needs evidence such as avoided overtime or reduced external spending before recognizing a cash benefit. To count released capacity as business value, the workflow owner should explain what that time enabled.
State how much the comparison can attribute to training
A before-and-after improvement may also reflect a new model, an easier task mix, a changed process or different access. Record those changes alongside the result.
Use a comparable group or staggered rollout where practical. If that is not feasible, label the result as a before-and-after estimate with limited attribution. Avoid describing every improvement observed after training as caused by training.
Risk and quality improvements can matter even when a cash benefit is not yet visible. Report source-checking behavior, escalation and errors directly. An avoided hypothetical incident should not become a made-up dollar saving in the ROI calculation.
Give the sponsor an expansion decision
The readout should name the task, period, eligible population, participation and evidence limits. Show comfort responses, quality results, repeated use, completion time and full costs. Include unresolved access or policy problems and the support provided.
Then recommend a next action based on that evidence. Stronger quality with low usage may call for office hours and manager support. High usage with poor quality may call for a narrower task and more review practice. A technically blocked workflow may need implementation work before a larger training cohort.
Use the AI ROI calculator for scenarios and keep those forecasts separate from measured results. The AI education program workbook includes a measurement sheet for planning the evidence before launch.