Kirkpatrick Four-Level Evaluation Model
Kirkpatrick Model
What is the Kirkpatrick four-level evaluation?
- Four levels of evaluation: reaction, learning, behavior, results
- Evaluate training by results, not just satisfaction
- Higher levels are more valuable to measure
What the Kirkpatrick model really changes is when you evaluate. Most training managers worry about what to ask after a course ends. Used properly, this model has you write down, at the planning stage, 'What behavior should we see on the job six months after this training for it to count as a success?' That sentence is what changes the curriculum, creates assignments, and gives team leaders a role. The model pays off when you use it as a design tool, not an evaluation tool.
The four levels move in opposite directions on measurement cost and persuasiveness. Level 1 (reaction) takes one survey but carries no weight with executives. Level 4 (results) uses numbers like sales, defect rates, and turnover, but factors outside training get mixed in, making causation hard to claim. That is why many teams in practice make Level 3 (behavior) their main focus. Asking supervisors and peers 4 to 12 weeks after training, 'What has changed about this person?', gives the best credibility for the cost. For reference, Jack Phillips added ROI as a fifth level, and the 'New World model' put forward by Kirkpatrick’s successors treats the on-the-job environment that supports Level 3 (manager support, rewards, systems) separately as 'required drivers'.
The most common misconception is that a high Level 1 score means Level 4 will follow. It is not unusual for a course with a 4.8 instructor satisfaction score to show zero behavior change three months later, because enjoyment and transfer are different variables. Another pitfall comes from the training team trying to prove Level 4 alone. Report that sales rose 12% after sales training, and if an executive asks, 'Isn’t that just the new product launch?', your case falls apart on the spot. Claim results without a control group, or at least an 'right after training vs. non-attendees' comparison, and you lose credibility at next year’s training budget review.
Hwi KimCEO of HRDIST & Talent Development Systems Consultant
John KimJohn Kim, author of B2B Marketing Design and CEO of Nanum ConsultingA four-level evaluation model (reaction, learning, behavior, results) that Donald Kirkpatrick first published in a series of articles in 1959, drawing on his 1954 doctoral dissertation. It became widely known through his 1993 book Evaluating Training Programs.
Take Company A, a building-materials manufacturer with 900 employees. It had 14 safety accidents the previous year and got a warning from its division head. The next year it ran two rounds of a 4-hour 'hazard prediction training' for all production staff, yet accidents did not decline. The training manager was asked, 'Satisfaction is 4.6, so why is nothing changing on the shop floor?' From the third round on, they built Kirkpatrick's four levels into the design stage.
- The training manager changes the first line of the plan. Instead of 'Deliver hazard prediction training,' they first write a Level 3 goal: 'In the 5-minute pre-work TBM, the foreman names at least two hazards on their own', and then work backward from it to trim the curriculum.
- On the Level 1 survey, drop the item 'The instructor was enthusiastic' and replace it with an open question: 'List three items you could use on your line tomorrow.' The share of blank responses falls from 38% to 11%.
- At Level 2, instead of a quiz, participants get 12 real photos and must find the hazards. The average rises from 4.2 to 8.1 hazards found, but the production manager says 'spotting a hazard isn’t the same as speaking up about it' and asks for a Level 3 check.
- At Level 3, six weeks after training, a safety manager observes 27 randomly chosen TBM (toolbox meeting) sessions. Foremen named hazards unprompted in only 33% of them, revealing that the problem is not the training but an environment where foremen are too busy.
- So the TBM checklist is cut down to a single page, and on-site observation is added to the foremen's KPIs. In the re-measurement, the share of foremen speaking up rises to 81%, near-miss reports (a Level 4 indicator) grow from 6 to 23 per month, and actual accidents fall from 14 to 5 per year.