The Kirkpatrick model of training evaluation

Donald Kirkpatrick, an academic at the University of Wisconsin, set the four levels out in a series of articles for the journal of the American Society of Training Directors, later ASTD, in 1959, drawing on his doctoral research earlier that decade. He called them steps rather than a model; the branding came later, along with his 1994 book on evaluating training programmes. It survives because it asks the question sponsors actually care about: did anything change afterwards?

The four levels at a glance

Level 1: reaction deserves better questions

Level 1 measures the experience: engagement, relevance, whether the room stayed with the facilitator. It is not a measure of effectiveness and never was. It dominates because it is cheap, it happens while everyone is still in the room, and it produces a number for Friday's deck.

The fix is not to abandon it but to stop asking questions that only produce a warm glow. Ask what the person intends to do, and what will stop them.

• What is one thing you will do differently in the next two weeks, and with whom?
• What part of today felt furthest from your actual job?
• How confident are you that you could do this unsupported tomorrow?
• What will make this hard to apply when you get back to your desk?

High satisfaction with low intention to apply is the most useful result a level 1 survey can give you. The session worked as an event and failed as training.

Level 2: learning, measured before and after

Level 2 asks whether anything shifted: knowledge, skill, attitude, confidence, commitment. The critical design decision is a baseline. Without a before measure you cannot separate what the training added from what people already knew, and post-only assessments flatter the programme.

• For knowledge, run the same short assessment before and after, with the same items.
• For skill, watch a demonstration: have people run the conversation, the pitch, the difficult review, scored against criteria set in advance.
• For confidence, self-rate at the start and the end, and treat large jumps with suspicion. Confidence rises faster than competence in a supportive room.
• For commitment, ask for a written, specific, dated intention.

Post-only measurement does most damage on skill subjects. Someone can pass a quiz on active listening and still interrupt every second sentence in the next meeting. Watch the behaviour, do not test the vocabulary.

Level 3: behaviour, the level that decides everything

Level 3 asks whether the person does the job differently four, eight, twelve weeks later. This is where training either converts or evaporates, and it is the level most organisations skip, because it means going back to people after the invoice has been paid.

The uncomfortable truth of level 3 is that the training is rarely the binding constraint. People return from a good workshop to a manager who never mentions it and a system that still rewards the old behaviour. James and Wendy Kirkpatrick, who developed the New World Kirkpatrick Model from around 2009, named this well: required drivers, the coaching, accountability and reward that must exist at work for a behaviour to survive contact with it.

How to actually measure level 3

1. Define the critical behaviours before the training runs. Two or three, observable, phrased as actions: "opens one-to-ones by asking rather than updating", not "communicates better".
2. Ask the observer, not the participant. Self-report inflates. Managers, peers and direct reports see the change or its absence.
3. Measure twice. At six weeks, then at three months. The first check catches early collapse, the second tells you whether it stuck.
4. Record what got in the way. The obstacles list beats the score, because it is the part you can change next time.
5. Build the drivers into the design, not after it. Manager briefings beforehand, a structured follow-up conversation after, one practice commitment per person.

This is also where everyday feedback quality does more work than any evaluation instrument. If nobody ever comments on the new behaviour, it has no reason to persist.

Level 4: results, and the attribution trap

Level 4 connects behaviour to the outcome the training was bought for: retention in a team that was leaking people, fewer escalations, faster onboarding, a safety record. Four rules keep it honest.

• Agree the measure with the sponsor before the programme. A result chosen afterwards is a result chosen to flatter. If the sponsor cannot name the number they want moved, you have a briefing problem, not an evaluation problem.
• Use measures the organisation already collects. Evaluation that needs a new data pipeline never gets done twice.
• Track leading indicators too. Escalation volume moves months before attrition does, and early signals hold a sponsor's attention while the real number catches up.
• Be honest about attribution. The same quarter contained a reorganisation, a pricing change and a market shift. Claiming the whole delta destroys your credibility the first time the number moves the wrong way.

Design backwards, evaluate forwards

The most useful correction the Kirkpatricks made to their own model was the order of operations. Plan at level 4 and work down. Measure at level 1 and work up.

1. Start with the result the sponsor wants and its leading indicators.
2. Ask what people would have to do differently for that result to move. Those are the critical behaviours.
3. Ask what they would have to know, do or believe to behave that way. That is the learning objective.
4. Only then design the session, and only then decide what a good reaction would look like.

Programmes built the other way round, topic first and outcome invented afterwards, are why so much training evaluates well and changes nothing.

Where the model breaks down

Kirkpatrick is a checklist of questions, not a theory of how learning works. Four honest limitations:

1. The causal chain is assumed, not proven. The model implies a good reaction leads to learning, then behaviour, then results. Research reviews going back decades have repeatedly failed to find a reliable link between satisfaction scores and later performance. People enjoy sessions that teach them nothing and resent sessions that change them.
2. It ignores everything before the training. There is no level for needs analysis, participant selection, or whether this was ever a skills problem. Many programmes that fail at level 3 were solving a systems problem with a workshop.
3. Levels 3 and 4 cost money, so they do not happen. In practice the model is honoured almost entirely at level 1. Any plan assuming unlimited follow-up capacity will be quietly abandoned.
4. Attribution at level 4 is genuinely hard. Outside a controlled comparison you cannot separate the training's effect from everything else that quarter. Treat level 4 as evidence in an argument, not proof.

Used well, it still earns its place by forcing one conversation: what would have to be different in six months for this to have been worth doing? Ask that before you book the room and most of the evaluation design writes itself.

Kirkpatrick compared with the alternatives

Phillips inherits the attribution problem and adds a costing burden, so use it where outcome data is already reliable. Brinkerhoff is often the better second instrument: rather than averaging everyone, it studies the few for whom the training clearly worked and the few for whom it clearly did not. The gap between those groups is usually about the workplace, not the workshop.

A minimum viable evaluation plan

• Before: the sponsor names one business measure and two or three critical behaviours. Managers get a one-page brief.
• During: observe skill demonstrations against pre-set criteria. That is your level 2.
• Immediately after: four questions, including intention to apply and expected obstacles.
• Six weeks after: a fifteen-minute manager conversation per participant. That is your level 3, and it is worth more than everything above it.
• Next quarter: revisit the sponsor's measure and say plainly what you can and cannot attribute.

None of this replaces the thing evaluation is trying to detect. Behaviour change is visible long before a survey catches it, which is why experiential learning and evaluation fit together: put a team into a real challenge under real pressure and level 3 happens in front of you, unprompted. Who reverts to old habits when time gets short, who tries the new approach and drops it the moment it gets uncomfortable, who quietly coaches someone else through it. Read honestly, that tells you more about whether the training survives Monday than any post-course score. The four levels then give you the language to report it.

Tour De Force runs this as live, experiential training for teams worldwide — online and in person. Talk to us, or play Gamified learning appsThe Weekly Challenge to see the method in ten minutes.