Most organisations measure training by counting things: completions, hours, attendance, satisfaction scores. Those numbers describe activity, not impact. Measuring training ROI properly means connecting what training cost to what changed as a result — in behaviour, in output, and eventually in dollars. This guide sets out a practical, honest framework for doing that, written for L&D managers, COOs and HR leaders who suspect their training spend is unmeasured and want proof rather than assurances.
I run Learning Elements and the Training Intelligence Model (TIM). Everything below comes from practice: what actually gets measured in real organisations, what data actually exists, and where the honest gaps sit. Where I don’t have evidence, I’ll say so.
What training ROI actually is (and isn’t)
Training ROI is the ratio of the financial value of the changes training produced to the full cost of producing them. A workable definition to hold onto:
Training ROI is the measurable financial return of behaviour change that training caused — divided by everything it cost to cause it.
The standard formula, adapted from the Phillips ROI Methodology, is:
ROI % = (Programme Benefits − Programme Costs) ÷ Programme Costs × 100
That formula is simple. The hard part is neither term. “Costs” is easy to under-count (people’s time is the biggest one and the one most often ignored). “Benefits” is easy to over-claim, because training is rarely the only thing that changed. The discipline is in isolating the effect of training from everything else happening in the business at the same time — new systems, new managers, seasonal variation, a restructure. If you can’t isolate the effect, you don’t have ROI; you have a correlation and a hope.
The five levels of training measurement — and where most organisations stop
The Kirkpatrick model’s four levels (Reaction, Learning, Behaviour, Results) and Phillips’ fifth level (ROI) remain the most useful map, not because they’re fashionable but because they describe where measurement effort actually runs out:
| Level | What it measures | Typical evidence | Where it fails |
|---|---|---|---|
| 1. Reaction | Did participants find it useful? | End-of-course surveys | Correlates weakly with anything downstream |
| 2. Learning | Did knowledge or skill change? | Pre/post assessments | Doesn’t show whether learning is used |
| 3. Behaviour | Do people do the job differently? | Observation, work samples, manager checks | Requires data most organisations don’t collect |
| 4. Results | Did business outcomes move? | KPIs, quality, retention, error rates | Attribution: many other forces move the same KPIs |
| 5. ROI | Is the financial return positive? | Isolated benefits vs full costs | Requires levels 3–4 to be credible first |
Most organisations live comfortably at levels 1 and 2, because those levels produce data automatically. Levels 3–5 require deliberate design: deciding before the training what behaviour will change, what signal will show it, and how you’ll separate training’s effect from everything else. If your measurement plan starts when the course ends, you have already lost levels 3 to 5.
A practical framework: the seven questions to answer before spending a dollar
Before commissioning or building any training programme, answer these seven on one page. If you can’t, the programme’s ROI is unknowable by design — which is a finding in itself.
- 1. What business problem is this training solving? Not “communication skills” — a specific, currently-measurable gap (e.g. first-contact resolution dropped from 78% to 71% over two quarters).
- 2. What will people do differently afterwards? Name the observable behaviour. “Use the new diagnostic script on every call” is observable. “Communicate better” is not.
- 3. What data already captures that behaviour? QA scores, call recordings, CRM entries, error logs, supervisor sign-offs. If no system holds the signal, decide who will collect it and how — before launch.
- 4. What is the full cost? Design + delivery + platform + participant time (wage × hours × headcount) + manager time + travel + backfill. Participant time usually doubles the “official” cost.
- 5. What’s the comparison group? Ideally some teams or individuals train later than others. That staggered rollout gives you a natural control. If everyone trains at once, you’ll need trend baselines instead — weaker, but better than nothing.
- 6. What result should move, by how much, by when? Set the expected effect size and a review date. “First-contact resolution back to 78% within one quarter of completion.”
- 7. What will we do if it doesn’t move? Deciding the fail action in advance is what makes the measurement honest rather than performative.
Sizing the benefit: convert behaviour change to money
ROI requires dollars, and dollars require a chain from behaviour to money. Two worked examples of the method (illustrative arithmetic, not case-study claims):
Example 1 — contact centre error reduction. A QA programme lifts process-correct handling from 70% to 78% of 10,000 monthly contacts. Each failed interaction costs roughly 12 extra minutes of rework at $45/hour loaded cost → 800 additional failures × 0.2 hours × $45 ≈ $7,200/month recovered. Annualise, subtract the fully-loaded programme cost, apply the ROI formula.
Example 2 — compliance completion-to-audit defence. An audit finding costs a day of management preparation per affected team. If audit-ready evidence cuts that to two hours per team per audit, value = hours saved × loaded rate × audits per year. It is modest, defensible arithmetic — and far more credible than an invented “productivity uplift”.
Two rules keep this honest. First, use conservative estimates and show the arithmetic — anyone should be able to re-run your numbers. Second, apply a confidence adjustment when you can’t fully isolate the training effect: Phillips suggests crediting only the portion of improvement you can attribute (a subjective but explicit step; state your reasoning and keep it consistent).
That measurement discipline — deciding signals and baselines before training, isolating effects, and converting behaviour to defensible dollars — is the core of the Training Intelligence Model (TIM), which treats training measurement as an intelligence function rather than an administrative one. The earlier articles in this series cover the mechanics in more depth: how to measure training ROI with a practical framework and which training metrics actually predict ROI.
The mid-guide checkpoint: what does your current data say?
Here is the honest test that separates organisations with a measurement function from those with a completion-tracking habit. Take your most expensive training initiative from the last twelve months and answer, from data you already hold:
- What did it cost in full, including participant time?
- What behaviour was it supposed to change, and what data shows whether it did?
- Which business metric moved, and can you explain why training was the cause rather than a contributor?
If you can answer all three from existing records, you’re in the small minority. If you can’t, the gap isn’t effort — it’s that the organisation never decided what evidence it would need. If you’d rather talk that gap through than keep guessing, discuss it with us directly — the first conversation is a working diagnosis of your current data, not a sales call.
Common failure modes (and how to avoid each)
| Failure mode | What it looks like | The fix |
|---|---|---|
| Counting activity | Dashboards of completions and hours | Report behaviour and outcome signals alongside; treat completions as logistics, not evidence |
| Happy-sheet ROI | Survey scores presented as effectiveness | Never compute ROI from level-1 data; start at level 3 |
| Unpriced time | “Free” internal training | Always load participant and manager time into cost |
| Attribution theatre | Claiming a full-year KPI improvement for a 2-day course | Isolate with comparison groups or conservative confidence adjustments |
| Measuring after the fact | Deciding what to measure when the budget review arrives | Design measurement before launch (the seven questions above) |
What “good” looks like at each maturity stage
| Stage | Measurement reality | Reasonable next step |
|---|---|---|
| Stage 1: Tracking | Completions and attendance only | Add one behaviour signal for your biggest programme |
| Stage 2: Signal | Some behaviour/outcome data, no baselines | Establish pre-training baselines; stagger the next rollout |
| Stage 3: Evidence | Baselines + isolation for major programmes | Convert to dollars; report per-programme, not aggregated |
| Stage 4: Intelligence | Portfolio-level view: what works, for whom, at what cost | Predictive use — redirect spend before it’s wasted |
Don’t jump stages. A Stage 1 organisation that attempts full ROI produces numbers nobody trusts, which is worse than admitting you’re at Stage 1 — because once a credibility-destroying number circulates, every later honest number inherits its reputation.
There is no universal benchmark, and anyone quoting one (“training returns 4x!”) is selling something. A defensible ROI is one where the benefit chain is documented, the attribution logic is stated, and the arithmetic can be re-run by a sceptical CFO. A conservative, well-attributed 50% beats an unattributed 400% in every decision that matters.
The cleanest method is a staggered rollout: train some groups first and compare them to comparable groups that haven’t trained yet. Where that’s impossible, use long-run baselines (the same metric across several periods before training), and adjust for known confounders explicitly. Phillips’ confidence-adjustment method makes the remaining judgement visible rather than hidden.
Yes, but the chain is longer: behaviour change (observable in structured observation, work samples or stakeholder feedback), then a proxy outcome such as escalations or attrition in the trained team versus untrained, then dollars. The links are weaker and the confidence adjustment larger. State the assumptions; don’t pretend precision.
Four things: a full cost record (including time), a pre-training baseline for the target metric, a behaviour-level signal collected after training, and a documented attribution decision. If your LMS holds only completions — the common case — you’ll need to add data sources.
If you want an external, structured look at what your current training data could and couldn’t support, that’s the conversation TIM exists for. Start it here, or connect with Peter on LinkedIn.
