The Kirkpatrick four-level model of training evaluation: reaction, learning, behavior, results. Definitions, examples, sample questions per level, plus the New World Kirkpatrick upgrade for 2026.
The Kirkpatrick model is a four-level framework for evaluating training programs, developed by Donald Kirkpatrick in 1959. The four levels measure progressively deeper effects of training: Level 1 Reaction (how participants felt), Level 2 Learning (what they learned), Level 3 Behavior (whether they apply it on the job), and Level 4 Results (whether the organization benefited). It remains the most widely used training evaluation model worldwide in 2026.
Donald Kirkpatrick developed the four-level model as part of his 1954 doctoral dissertation at the University of Wisconsin, then published it as a series of four articles in 1959 in the Training and Development Journal. His 1994 book Evaluating Training Programs: The Four Levels codified the framework and made it the de facto standard for corporate L&D evaluation. His son Jim Kirkpatrick and daughter-in-law Wendy Kayser Kirkpatrick continue the work today through Kirkpatrick Partners, including the 2016 New World Kirkpatrick Model update covered later on this page.
The reason the model has lasted seven decades is structural. Each level answers a question a different stakeholder actually asks. Participants and instructors care about Level 1 Reaction. L&D teams care about Level 2 Learning. Managers and HR business partners care about Level 3 Behavior. The CFO and the board care about Level 4 Results. A single evaluation that satisfies all four levels means the same dataset can answer four different audiences.
The challenge is that most enterprise training programs stop at Level 1. The end-of-session smile sheet captures Reaction and goes nowhere near Behavior or Results. Capturing Levels 3 and 4 traditionally required expensive 360 surveys 3 to 6 months post-training, which most programs cannot afford. The result is a generation of L&D programs that satisfy Kirkpatrick Level 1 on paper while having no defensible answer to "did the training change behavior" or "did the business benefit."
The AI-native upgrade to Kirkpatrick does not replace the four levels. It changes the data collection underneath them. Persistent participant IDs mean every Pre, Mid, Post, peer rating, audio reflection, and LMS event lands on the same row automatically. AI extraction on collection means open-ended responses become Level 1 sentiment, Level 2 evidence, and Level 3 indicators in real time. Cross-system joins with the LMS and feedback platforms make Level 4 attribution defensible. The rest of this page works through one cohort end to end, with each section mapped explicitly to the Kirkpatrick levels it satisfies.
12 weeks, 24 participants, one persistent learner ID each. Open-ended responses captured alongside scaled metrics. Mid-cycle coaching interviews ingested as structured evidence. AI narrative summaries written for every participant.
The four Kirkpatrick levels measure progressively deeper effects of training: Reaction (L1), Learning (L2), Behavior (L3), and Results (L4). Each level answers a question a different stakeholder actually asks. The data collection design determines which levels you can actually evaluate. Most programs stop at L1 because L2 to L4 traditionally required expensive separate studies. The AI-native approach captures all four levels from one persistent record.
Every measurement point in the cohort above maps to one or more Kirkpatrick levels. The Pre assessment captures L1 baseline reaction plus L2 starting score. The Mid interview captures L2 mid-cycle plus an L3 early-application signal. The Post score plus peer 360 captures L2 final, L3 sustained behavior, and inputs to the L4 distribution analysis. The same persistent record feeds all four levels. The four cards below cover each level in detail.
What it measures. Engagement, perceived relevance to the job, confidence in the instructor or content, and willingness to recommend the program. The most superficial level, but a leading indicator: a participant who flags low engagement at Pre is unlikely to show L2 learning at Post unless the program intervenes.
Traditional method. The end-of-session smile sheet survey. Captures Reaction once, at the worst possible moment (right after a long workshop, when participants want to leave). Produces inflated positivity that does not predict L2 to L4.
AI-native upgrade. Sentiment captured continuously from one open-ended question at Pre, Mid, and Post. AI extracts engagement, perceived relevance, and risk flags from each response. A risk flag at Pre week 1 is addressed in week 2, not noticed at Post week 12. The Spring 2026 Communication Skills cohort cleared 4 of 4 risk flags by Post.
Example Level 1 question. "What worries you most about applying these skills at work? Be specific about a recent situation if you can." AI extracts sentiment polarity, theme cluster, and risk flag from a single paragraph.
What it measures. Knowledge, skills, attitudes, confidence, or commitment acquired during the training. The classic Level 2 instrument is a Pre to Post score delta on a target competency. For Communication Skills, the Spring 2026 cohort moved an average of 24 confidence points (52 to 76 on a 0-100 scale). The distribution shift is more informative than the average: the cohort moved from 100% Low confidence at Pre to 70% High at Post.
Traditional method. Pre-test and post-test on a knowledge assessment. Captures recall but rarely captures application or attitude shifts. Often graded by the instructor, introducing scoring bias.
AI-native upgrade. Same scaled self-rating Pre and Post for clean delta. Six skill dimensions on a radar chart with Pre and Mid overlaid so the program manager sees which competencies still need work in week 7. AI extracts evidence of concept mastery from open-ended responses at each measurement point. Marcus Thompson moved Voice 4 to 9, Structure 3 to 8, Pushback 2 to 7 over 12 weeks.
Example Level 2 question. "On a 0 to 100 scale, how confident are you speaking up in cross-functional meetings?" Same question, asked at Pre, Mid (in the interview), and Post. The delta is the L2 signal.
What it measures. Whether the learned skill or behavior shows up in the participant's daily work. The hardest Kirkpatrick level to capture defensibly. The Spring 2026 cohort's peer-rated effectiveness moved from 6.4 to 7.6 of 10 (+1.2 points), and Marcus Thompson logged 9 speaking events (meetings led, presentations given) during the 12 weeks.
Traditional method. 360 feedback survey 3 to 6 months post-training. Expensive (often requires HR to coordinate), slow (results arrive after the cohort has moved on), and most programs skip it entirely.
AI-native upgrade. Three signals captured during the program, all on one record. First, peer-rated effectiveness from 6 cohort members at Post. Second, real-world application count (speaking events for Communication Skills, customer calls for Sales Enablement, customer presentations for Customer Success). Third, LMS application activity joined via persistent ID. The Spring 2026 cohort's L3 signal was visible at week 12, not 3 months later.
Example Level 3 prompt. "Walk me through the hardest moment so far where you applied a skill from this program. What did you try? What worked?" Asked during the Mid-cycle interview. The story becomes Level 3 evidence captured as structured data, not anecdote.
What it measures. The targeted business or operational outcome the training was designed to produce. For Communication Skills, the L4 result was peer-rated effectiveness lifting by 1.0+ points and 4 of 4 risk flags cleared. The cohort delivered +1.2 peer effectiveness and 4 of 4 risk flags cleared, outperforming Toastmasters P75 (+18) by 6 points and corporate L&D average (+9) by 15 points.
Traditional method. Business KPI attribution. Methodologically difficult because the training is one of many factors influencing the KPI. Most attempts produce numbers that the CFO discounts as unfalsifiable.
AI-native upgrade. Multivariate regression with standardized beta coefficients ranking program drivers. The Spring 2026 model returned: Mentor session minutes β=0.42, Peer pair sessions β=0.31, Speaking events β=0.24, AI narrative engagement β=0.18, LMS module completion β=0.09 (not significant). The model explains 68% of the variance. This is the defensible attribution that satisfies a finance team. The cross-system AI agent in Component 3 above lets a program manager interrogate the L4 result further with plain-English questions.
Example Level 4 question. "Compared to the Toastmasters P75 benchmark of +18 confidence points, did our cohort beat or miss?" The benchmark comparison surfaces in Report 2 of Component 2 above. The board uses this number, not the raw delta.
| Level | What it measures | Traditional method | AI-native upgrade | Cohort signal |
|---|---|---|---|---|
| L1 · Reaction | How participants felt | End-of-session smile sheet | Continuous sentiment from open-ended responses | 4 of 4 risk flags cleared |
| L2 · Learning | What they learned | Pre-test, Post-test | Skills radar with Pre/Mid overlay, AI evidence extraction | +24 confidence · 100% Low → 70% High |
| L3 · Behavior | If they apply it on the job | 360 survey 3-6 months post | Peer rating + application count + LMS activity, all on one record | Peer 6.4 → 7.6 · 9 events for Marcus |
| L4 · Results | If the organization benefited | Business KPI attribution | Multivariate regression with standardized β | +6 to +15 above 3 benchmarks |
| L5 · ROI Phillips extension |
Financial ratio of benefits to costs | Cost-benefit analysis on L4 result | Monetize L4 outcome, divide by program cost | (Optional · only when CFO requests) |
Walk one of your past cohorts through the L1 to L4 measurement design. 30 minutes with a Sopact specialist. See the reports your CFO would actually accept.
Book a demo →Same 24 participants. Same Pre, Mid, Post evidence. Different shape for different audience. Multilingual is a toggle, not a translation project.
Spring 2026 Communication Skills cohort · N=24 · Pearson correlation analysis
Pre to Post movement · cohort distribution · benchmark comparison · for board and exec audiences
Movimento Pré para Pós · distribuição da coorte · comparação com referências · para diretoria e executivos
Linear regression · 5 program variables predicting Pre-to-Post confidence delta · N=24
Three questions per measurement point usually outperforms a 40-question survey: one scaled self-rating, one yes/no behavioral check, and one open-ended question that earns its keep. The same three-question pattern maps to all four Kirkpatrick levels when designed correctly. The open-ended question is where Level 1 sentiment, Level 2 evidence, and Level 3 behavioral signal all come from in a single paragraph.
The questions below are the actual instrument used for the Spring 2026 Communication Skills cohort. Each question is tagged with the Kirkpatrick level it primarily satisfies. A few questions satisfy multiple levels, which is the point: well-designed instruments map a single response to multiple levels of analysis.
No SQL. No BI ticket. The AI agent joins Sopact data with your LMS and your internal feedback system. Click a prompt to watch the answer come back with the sources tagged.
The engagement paradox lives in two participants who completed everything in the LMS but barely moved on Post confidence.
Plotting LMS module completion against Post confidence for the Spring 2026 cohort surfaces a quadrant pattern. Most participants cluster around the diagonal: high LMS engagement tracks with high Post confidence (top-right). But two outliers break the pattern in opposite directions.
Aisha K. (P-1244) completed all 12 LMS modules with a 95 average quiz score, the highest in the cohort. Her Post confidence only rose +6 points (52 to 58), bottom quartile. Pattern matches participants who treat the LMS as a checklist exercise without internalizing the skill. Diego R. (P-1243) finished only 8 of 12 modules but his Post confidence jumped +22 points, driven by 14 attended peer-pair sessions and 9 volunteered speaking events.
What this means: LMS completion is not the change driver. Two participants saturated on async content and still showed the smallest growth. Three under-engaged on LMS but grew most. The human elements of the program carry the lift.
The human elements outrank every single LMS module. Mentor sessions correlate twice as strongly with confidence lift as your best async module.
I correlated each program element with the Pre-to-Post confidence delta across 24 participants. Higher r means the element more reliably predicts a participant's confidence growth. Two non-LMS elements (mentor sessions, peer pairs) are ranked alongside the 6 Cornerstone LMS modules to show the comparison.
What this means: The 22-minute video on handling pushback (Module 04) is the only async content with a meaningful signal. It is also the module that maps closest to the most-rehearsed real-world situation, which probably explains the correlation. The five other modules sit at or below r=0.42.
Action: for Summer 2026, recommend keeping Module 04, replacing Modules 01 and 03 with one extended mentor session, and tracking whether the freed time materially shifts the cohort's Post confidence distribution.
Five Spring 2026 graduates qualify as Summer 2026 mentors based on the three-system join.
Filter criteria applied across all three systems: Sopact · completed program with Post confidence above 75. Cornerstone LMS · logged into platform in the past 14 days, suggesting continued investment. Lattice · gave at least 4 pieces of peer feedback in the past month, indicating they are comfortable being a source of feedback for others. Five of 21 graduates meet all three criteria.
Note on Diego: his SOPACT score is the lowest of the five at 71, but the lift was outsized (+22) and his Lattice giving rate suggests he learned through peer practice rather than module completion. Could be the strongest peer-style mentor for Cluster B participants in Summer 2026.
The New World Kirkpatrick Model, published by Jim and Wendy Kirkpatrick in 2016, updates Donald's original four levels with backward planning from Level 4 and the concept of Required Drivers as the bridge between Level 2 Learning and Level 3 Behavior. Phillips ROI Methodology adds a fifth level for financial return. CIRO reorders around context and input. Brinkerhoff's Success Case Method uses narrative evidence. Kirkpatrick still anchors most programs because its four levels map cleanly to questions stakeholders actually ask.
The original Kirkpatrick four levels came out in 1959. A lot has changed since then, including the rise of cheap LMS data, AI extraction on collection, and the demand from CFOs for defensible financial attribution. The model has been updated and challenged repeatedly. The two most important variants are the New World Kirkpatrick Model (2016) and the Phillips ROI Methodology (1996). Two less common but historically relevant alternatives are CIRO and Brinkerhoff. The table below covers all four.
Published in 2016 by Jim Kirkpatrick (Donald's son) and Wendy Kayser Kirkpatrick through Kirkpatrick Partners. The four levels stay the same: Reaction, Learning, Behavior, Results. The big change is the planning sequence and the introduction of Required Drivers.
Backward planning. Start at Level 4 by defining the operational outcome with the business sponsor. Then design Level 3 Required Drivers (the systems, manager behaviors, and processes that reinforce on-the-job application). Then Level 2 Learning objectives. Then Level 1 Reaction design. This is the opposite of how most legacy programs work, which start at Level 1 and hope Level 4 will follow.
Required Drivers. The processes that reinforce L3 behavior change between training and 6 months post. Examples: a manager 1:1 in week 4 that explicitly asks about applying the new skill, a peer accountability pairing, a Slack channel where cohort members share weekly wins. The Kirkpatricks argue that the bridge between L2 and L3 is where most training programs fail, not in the training itself. The Spring 2026 Communication Skills cohort's mentor sessions (β=0.42 in the multivariate model) are an example of a Required Driver delivering measurable L3 lift.
| Model | Levels or stages | Key idea | Best for | Where it falls short |
|---|---|---|---|---|
| Kirkpatrick 1959 original · 2016 New World |
4 levels: Reaction, Learning, Behavior, Results | Progressively deeper effects of training, each level a different stakeholder question | Anchoring any training evaluation program; clearest stakeholder mapping | L3 and L4 are expensive without persistent IDs and cross-system data |
| Phillips ROI 1996 |
5 levels: Kirkpatrick's 4 plus Level 5 ROI | Monetize the Level 4 outcome, divide by program cost, report as a financial ratio | Justifying training spend to finance and the board; ROI calculations | Isolating training's contribution from other variables is methodologically difficult |
| CIRO Warr, Bird, Rackham 1970 |
4 stages: Context, Input, Reaction, Outcome | Needs analysis first (Context), then resource fit (Input), then reactions, then outcome | Program design choices before training begins, especially HRD contexts | Less commonly understood by external stakeholders; lacks ROI calculation |
| Brinkerhoff Success Case Method, 2003 |
Not levels; narrative cases of high and low performers | Identify the top 5% and bottom 5% of participants, interview both, look for the design and manager behaviors that drive the gap | Surfacing which design choices and manager behaviors drive outcomes | Qualitative-heavy; hard to aggregate across large cohorts without AI extraction |
The Spring 2026 Communication Skills cohort used Kirkpatrick as the spine with one Brinkerhoff overlay: Marcus Thompson (peer 7.8, +34 confidence lift) and Aisha K. (LMS 12/12 modules, +6 lift) were treated as the high-performer and low-mover cases. Their archetypes informed the Summer 2026 design changes. The cohort did not run Phillips L5 ROI because the business sponsor's question was operational, not financial. CIRO would have been overkill for a single-cohort program of 24 participants.
The Kirkpatrick model is a four-level framework for evaluating training programs, developed by Donald Kirkpatrick in 1959. The four levels measure progressively deeper effects of training: Level 1 Reaction (how participants felt), Level 2 Learning (what they learned), Level 3 Behavior (whether they apply it on the job), and Level 4 Results (whether the organization benefited). It remains the most widely used training evaluation model worldwide in 2026.
Level 1 Reaction measures how participants felt about the training. Level 2 Learning measures what knowledge, skills, or attitudes they acquired, typically via Pre to Post score deltas. Level 3 Behavior measures whether they apply the learning on the job, usually via 360 feedback or peer-rated effectiveness. Level 4 Results measures whether the training produced the targeted business or operational outcome. The Phillips ROI Methodology adds a fifth level for financial return.
Level 1 measures how participants felt about the training: engagement, perceived relevance to their job, confidence in the instructor or content, and willingness to recommend the program. Traditional Level 1 measurement is the end-of-session smile sheet. AI-native Level 1 captures sentiment continuously from one open-ended question per measurement point, flagging risk responses in real time so a participant can be re-engaged in week 7 rather than counted as a failure at Post.
Level 2 measures what participants learned during the training: knowledge, skills, attitudes, confidence, or commitment. The most common Level 2 instrument is a Pre to Post score delta on a target competency, expressed as a single number (such as +24 points on a 0 to 100 confidence scale). A skills radar chart with Pre and Mid overlaid reveals which sub-competencies have moved and which need additional work mid-cycle.
Level 3 measures whether participants apply the learning on the job. Traditional Level 3 instruments are 360 feedback surveys 3 to 6 months after training, which are expensive and slow, and most programs skip them entirely. AI-native Level 3 combines peer-rated effectiveness from cohort members at Post, real-world application counts logged during the program (such as speaking events or sales calls), and LMS application activity all on the same persistent participant record.
Level 4 measures whether the training produced the targeted business or operational outcome. Examples include cohort distribution shift from 100% Low confidence at Pre to 70% High at Post, 4 of 4 risk flags cleared, or peer-rated effectiveness lifting by more than 1.0 points. Isolating training's contribution from other variables is methodologically difficult, which is why multivariate regression with standardized beta coefficients is the AI-native approach to defensible Level 4 attribution.
The Phillips ROI Methodology adds a fifth level on top of Kirkpatrick's four: Level 5 Return on Investment, expressed as a financial ratio of training benefits to training costs. Kirkpatrick stops at Level 4 Results, which captures operational outcomes. Phillips's L5 requires monetizing the Level 4 result and dividing by total program cost. Most enterprise L&D programs use Kirkpatrick as the spine and add Phillips L5 only when a CFO asks for ROI specifically.
The New World Kirkpatrick Model, published by Jim and Wendy Kirkpatrick in 2016, is an update to Donald Kirkpatrick's original four levels. The major change is the emphasis on backward planning: start at Level 4 by defining the business result, then design Level 3 Required Drivers, then Level 2 Learning, then Level 1 Reaction. The model also introduces the concept of Required Drivers, which are the systems and processes that reinforce on-the-job behavior change as the bridge between Level 2 and Level 3.
Define the Level 4 business result first, with the business sponsor. Design Level 1, 2, and 3 measurement instruments that map to it. Use persistent participant IDs so all Pre, Mid, Post, peer, audio, and LMS data lands on one record. Capture open-ended responses at each measurement point for AI extraction of L1 sentiment and L2 evidence. Conduct a Mid-cycle structured interview for L2 plus L3 mid-signal. Run multivariate regression on the final dataset for defensible L4 attribution.
Donald Kirkpatrick created the four-level training evaluation model as part of his 1954 PhD dissertation at the University of Wisconsin. He published it as a series of four articles in 1959 in the Training and Development Journal, then expanded it in his 1994 book Evaluating Training Programs: The Four Levels. His son Jim Kirkpatrick and daughter-in-law Wendy Kayser Kirkpatrick continue the work today through Kirkpatrick Partners, including the 2016 New World Kirkpatrick Model update.
From backward planning at Level 4 to multivariate attribution. Frameworks, sample instruments, report templates, and how persistent IDs make all four levels measurable from one record.
Read the stakeholder intelligence guide →Kirkpatrick is one of several frameworks the Sopact architecture supports. Here is where else the same persistent record, open-ended evidence on collection, and cross-system AI shows up.
The complete training evaluation guide covering all four methodologies (Kirkpatrick, CIRO, Phillips, Brinkerhoff). Includes sample questions, four report types, and LMS integration.
Application, intermediate, and outcome data on one record per scholar. Rubric scoring with open-ended evidence.
Theory of change, Pre/Mid/Post measurement, distribution shift, and impact attribution. For social impact and corporate programs alike.
Application review, milestone tracking, and outcome reporting in one place. Built for funders managing portfolios of 20+ grantees.
Open-ended response extraction, multilingual analysis, AI-generated narrative summaries. The data-collection engine under every use case.
Persistent IDs across Pre, Mid, Post. Why this pattern outperforms cross-sectional surveys for measuring change.
The complete framework: how to design data collection, run analysis, and turn results into action across any stakeholder cohort.
Walk one of your past training cohorts through the L1 to L4 measurement design above. See the reports your CFO would actually accept. 30 minutes with a Sopact specialist.
