How to measure the return on AI training

Why do satisfaction scores mislead?

Because they measure the wrong thing. A participant can rate a day highly and change nothing about how they work on Monday. A sponsor who reports a high average rating to the executive committee has reported on the facilitator, not on the investment.

The gap between use and value is wide. Nearly nine in ten organisations now use AI regularly, yet only 37% attribute any earnings (EBIT) impact to it (McKinsey, The state of AI in 2026; Indicative). Access alone does not close it: a study of 25,000 workers found that staff who were simply given AI chatbots saved, on average, 2.8% of their work hours (Humlum and Vestergaard, 2025). Tools alone save little.

When AI is applied deliberately to defined tasks, the change is measurable and large. In a randomised experiment with 453 professionals, those using AI on writing tasks took 40% less time and produced work rated 18% higher in quality (Noy and Zhang, Science, 2023). That study measured exactly what a sponsor should: time and quality, on real tasks, against a comparison.

Measurement is also part of what separates results. Organisations getting most from AI are twice as likely to have committed senior leaders and a way to measure impact (McKinsey, 2026; Indicative).

What should an AI training impact report contain?

Four parts. At Cader Group we call the result the Impact Report, and every program ends with one.

1. Baseline. What the work looked like before: time taken and quality on named tasks, how AI is currently used, and which decisions are pending. Taken before the first session, or drawn from a readiness assessment.
2. Outputs. What participants produced: decisions, workflows, agents, each with an owner.
3. Measured change. The same tasks, measured the same way, after the program.
4. Next step. Scale, fix or stop, with a named owner and a date.

If any part is missing, the report is a description, not a measurement.

What should you measure for each audience?

Different audiences produce different results, so they need different measures.

Three rules keep the measures honest. Name the tasks before training, so they cannot be chosen afterwards to flatter the result. Measure quality as well as time: a faster output that needs rework has not saved anything. Where you can, have quality judged by a reviewer who does not know which version came before training and which came after.

How long before you see results?

Set expectations by stage, and report each one when it is due.

• At the end of the program: outputs exist, and the first before-and-after on named tasks is available.
• At 90 days: the real test of adoption. Workflows still in use, or not. This is where most of the drop-off shows.
• Over a pilot of six to eight weeks: results against a baseline agreed in week one, and a decision to scale or stop.
• Over several quarters: effects on budgets and headcount plans, which depend on decisions well beyond the training.

Claiming financial savings at the end of a training program overstates what training can show. Report time released in hours, and let finance decide how to value it once the time has actually been redeployed.

How do you report AI training ROI to leadership?

On one page, in this order:

• The baseline and the change, on the named tasks, in time and quality.
• What is in use: decisions funded, workflows running, agents in pilot, each with an owner.
• What did not work, and why. A workflow that was dropped is useful evidence.
• The next step, with the decision you need from leadership.

Keep adoption numbers (logins, prompts sent) in an appendix. They show activity, not value. For what a program should leave behind in the first place, see AI training that builds capability, not awareness; for setting the baseline, see how to assess AI readiness.

Measure where your own team stands: take the free Readiness Snapshot →