Should you train your maintenance team internally or externally?

Most maintenance teams should do both, in a specific order. Buy the training in, then reinforce it internally.

External delivery does not teach better, and the research on that is clear. What it buys you is acceptance, and acceptance is the part that turns knowledge into changed behaviour. Your own people then make it survive contact with the plant.

This article covers what the evidence actually says about your options, about who should deliver your training, where internal delivery genuinely wins, the two kinds of training that fail whoever runs them, and what decides whether any of it lasts more than a month.

One thing to declare up front. Reliability Academy sells external training, so a conclusion that favours buying it in is commercially convenient for us. Every figure below is sourced.

And a counterweight, because you should read the rest knowing it is there. This article names 4 situations where you should not buy external training at all, including one where the regulation makes it impossible.

You are probably arguing about the wrong thing

The argument usually runs the same way in every maintenance department.

Someone says the team needs proper training and we should bring in a provider. Someone else says our people know this plant better than any outsider ever will. Both sides think they're arguing about competence.

They aren't.

So what does the evidence say?

Lacerenza, Reyes, Marlow, Joseph and Salas (2017) ran the only study that has directly tested internal against external trainers, pooling the results of every piece of research they could find. 335 separate studies, 26,573 participants.

One symbol to explain first, because it turns up all through this article. δ is a measure of how big an improvement was. Roughly 0.2 is small, 0.5 is moderate, and 0.8 or more is large.

None of the gaps between internal and external trainers were big enough to count as real. On learning, the internal trainers actually came out ahead.

What was big enough to count was something else entirely. Both internal and external trainers comfortably beat training that nobody ran. Internal delivery beat do-it-yourself on learning, and external delivery beat it on transfer to the job.

The dividing line is not internal versus external. It is whether anybody is actually running the training.

Two caveats, because the study deserves them and you should know them before you use it. This was leadership training rather than technical maintenance content, and the internal-trainer samples were small, with only 6 studies and 173 people in the business-results row. So read it as no evidence of a difference, not as proof that the two are equivalent.

Which leaves you with an obvious question.

If neither one teaches better, why does almost everyone in this industry believe the outsider lands harder?

They are not imagining it.

So why does the outsider land better?

There's a 1951 experiment that answers this, and it answers it by splitting your question in two.

Hovland and Weiss gave Yale students identical articles on 4 controversial topics. The only thing that changed was who the article was attributed to. Half saw a highly credible source, half saw a source nobody trusted.

The last column uses p, which is how often a gap this big would show up by chance alone. Below .05 is the usual bar for treating a difference as real, so p < .01 clears it comfortably.

Same words on the page.

So what changed?

The source made no measurable difference to what people learned. It made a three and a half fold difference to what they accepted.

That resolves the contradiction. Lacerenza measured learning and transfer, where source does not matter. Hovland and Weiss measured acceptance, where it matters enormously.

And this is the part people get wrong

It would be easy to read that as "external training is mostly persuasion", and to treat persuasion as something separate from teaching.

That reading is wrong, and it's worth being direct about why.

Persuasion is part of teaching. How could it not be?

You can tell a room full of planners what the reliability-centred maintenance research found.

If you can't then get them to see what it means for their plant, and act on it, the knowledge yields nothing at all. The point of training is capability, not information.

If you think training is knowledge transfer, the acceptance gap looks like a nice-to-have. It's the part that makes the knowledge worth transferring.

The bit that nobody earned

Menon and Pfeffer (2003) ran an experiment in Management Science that is uncomfortable to read if you employ internal experts. The people surveyed rated 2 managers who had achieved identical business results. One had copied a practice from another site in the same company. The other had copied it from a competitor. The scores below are average ratings, so what matters is the gap between the two columns rather than the size of either number.

Same idea. Same outcome.

The only difference was where it came from.

So why would that happen?

The authors put it down to status. Taking a colleague's idea marks you as the follower. And internal knowledge is abundant, so it gets picked over, while external knowledge is scarce and therefore looks special.

That part of the outsider's advantage is unearned.

Your internal experts are on the wrong end of it, and there's nothing they can do about it.

But there is a second reason, and it is real

When I run a site assessment, I very rarely find something nobody knew.

The findings are already in your building. What tends to happen is that one specific finding sits with one specific person, in one part of the organisation, and has done for years. Someone in the store knows why the kits are wrong. A supervisor knows why the schedule never holds. A planner knows why the notifications are useless.

So what does an outsider actually add?

What an outsider does is put all of it on one page.

Instead of 1 person holding 1 finding, you get 30 findings in front of the leadership team at once, with the links between them drawn out. This is a problem, and because of it this happens, and then this, and then that. The individual points are usually known and understood. How they fit together, and what to do about it, usually isn't.

So there are 2 reasons the outsider lands harder, and they are different in kind. One is a status effect that nobody earned. The other is a genuine contribution, because the outsider has seen the whole system, and has seen it at 20 other sites.

And the uncomfortable one

Gino (2008) found that people used advice much more when they had paid for it, even though the advice itself was no better. Same instinct as sunk cost.

We proved that on ourselves, at our own cost.

Our scholarship program was free when we started it. We received applications by the hundred, and the vast majority of recipients did nothing with them.

So we changed it.

Recipients now contribute 25% of the cost.

That's the strongest argument in this article against doing it internally because it's free.

Free training gets treated as free. Why would your team value it any differently?

And why it wears off

Hovland and Weiss went back to the same people 4 weeks later.

Agreement with the credible source had fallen 10.7 points. Agreement with the untrusted source had risen 7.4 points. After 4 weeks the 2 were almost indistinguishable.

Why would it move in both directions?

Content is learned and forgotten at roughly the same rate whatever the source. What an untrusted source does is interfere with acceptance, and that interference fades faster than the content does.

Treat that as a tendency rather than a law. A later review by Kumkale and Albarracín (2004), pooling the studies on this, found the flip only shows up under fairly specific conditions.

The applied version is blunter. Saks (2002) surveyed training professionals across 150 organisations, who estimated that 62% of people apply the material immediately, 44% at 6 months, and 34% at one year. Those are practitioner estimates rather than measured behaviour, so treat them accordingly.

The credibility you paid for has a half-life of about a month. Two-thirds of the value is gone inside a year.

None of that is an argument against buying external training.

It's an argument against buying it and walking away.

"But we're different"

This is the main reason organisations decide to keep training in-house, so it is worth dealing with before the cases where they're right.

The claim arrives in a predictable set of forms, and you'll have used one or two yourself. Our site is remote. Our asset base is old. We're unionised. Our regulator is unusual. Our lead times are impossible. We carry no inventory. Therefore a generic external course cannot possibly fit us.

The belief is measurably expensive

Flyvbjerg, Budzier, Christodoulou and Zottoli (2024) at Oxford did something clever with 219 projects. They asked the project managers how unique they considered their project, on a scale of 1 to 10. Then they measured what actually happened.

• A one-point rise in perceived uniqueness predicted a 5 percentage point rise in cost overrun (p = .012)
• Projects rated 10 ran 45 percentage points worse than projects rated 1
• Perceived uniqueness had no relationship to actual project size, measured 4 different ways

That last point is the one that matters.

Perception is doing all of the work here. The projects that felt one of a kind were no bigger and no longer than the rest of them.

The authors explain it exactly as you'd expect: believing you are one of a kind stops managers learning, because they think there is nothing to learn from anyone else.

Four caveats you should know. It's a working paper rather than a peer-reviewed article. The authors think the projects they studied probably lean towards better-run organisations. The effect held for cost overrun but not for schedule or benefit overrun. And they were IT projects, not maintenance.

There's also a lovely finding from 1983 that you'll recognise instantly. Martin, Feldman, Hatch and Sitkin showed that organisations assert their uniqueness through 7 story types that turn up in near-identical form across completely different organisations. Everyone is special, and they all say so in the same 7 ways.

What it looks like on site

An Australian mine site told me before I arrived that they ran an effective remote planning and scheduling team using non-technical planners, and that it worked well. They were a bit different, they said.

They weren't, and it didn't.

So where does it break down?

The mechanism is the part worth writing down. A planner who has never overhauled a centrifugal compressor, a large pump or a transformer, and who may not be a tradesperson at all, does not know the steps. They don't know the parts, or how to find them in the CMMS. They don't know the specialist tools, the reinstatement requirements or the testing requirements.

From there, one of 2 things happens.

The root cause is almost always the same.

They confuse planning with scheduling, and nobody has ever explained work preparation to them. We plan jobs to eliminate waste during execution. Miss that, and the whole function looks like administration.

The distinction that makes this fair

Telling people they aren't special is neither accurate nor useful, because half of what they're claiming is true.

The fundamentals are the same everywhere. How you put them into practice genuinely depends on your size.

A team of 8 that tries to run a refinery's planning model will drown in process. A refinery that runs an 8-person model leaves a fortune on the table.

Both are applying the same fundamentals, and both should put them into practice completely differently.

So which layer are you actually talking about?

When people say they're different, they usually haven't separated the two. We split our worked ROI examples by site size for exactly this reason.

The technology version of the same mistake

The same instinct shows up as buying what's new while the basics are broken.

Here's the evidence that convinced me, and you won't find it published anywhere. Over the last 2 years, at least 3 separate providers of predictive maintenance, condition monitoring and AI technology have approached me privately with the same problem. Their customers are not getting value.

Their diagnosis was identical each time.

The customer only reacts to the worst symptoms, the things about to trip the plant next week. The early indications, which is the entire point of the technology, get buried in the reactive noise of the day and nobody ever acts on them.

That is not a training company arguing that fundamentals matter more than technology. That is the technology vendors saying their own product cannot deliver into a reactive organisation.

McKinsey reached the same conclusion in 2021, noting that few companies have managed to deploy predictive maintenance at scale, with missing skills one of the 6 barriers it names. That paper contains no survey and no percentages, so treat it as informed opinion rather than data.

Where internal training genuinely wins

So when should you keep it in-house?

Four cases, and the first one isn't a judgement call.

Where the law requires it

29 CFR 1910.147, the lockout/tagout standard, requires your energy control procedure to contain "specific procedural steps for shutting down, isolating, blocking and securing machines or equipment", and requires retraining "whenever there is a change in machines, equipment or processes."

29 CFR 1910.119, process safety management, is equally specific: "The employer shall train each employee involved in maintaining the on-going integrity of process equipment in an overview of that process and its hazards and in the procedures applicable to the employee's job tasks."

No external provider can deliver this, because the regulation ties the training to your written procedures for your equipment.

The law draws that line for you.

Anyone selling your site generic lockout training as a substitute is selling you something that doesn't satisfy the standard.

Where the numbers turn over

Building training yourself costs far more than most people assume, and there's a decent dataset on it.

Chapman Alliance (2010) surveyed 249 organisations covering 3,947 development professionals and nearly 20 million learners.

Run that out against a course you're considering building.

A 5-day internal course implies roughly 1,720 development hours and about $237,000 at those rates, before anybody teaches it once.

Those are 2010 dollars from a self-selected panel, and the study itself notes long courses are overstated by the model. Treat it as a rough scale rather than a precise figure. A competitor with every reason to say otherwise backs it up: Life Cycle Engineering publishes development ratios of 80:1 to 250:1 and says building the same capability in-house runs into hundreds of thousands of dollars.

Scale changes your answer, and 2 independent datasets a decade apart agree on the shape:

We see the other side of this in client conversations. Several organisations a year ask us to build them a customised, company-specific version of our training. I always take the conversation seriously, then walk them through what it costs us to build and what it will cost them to maintain.

They reach the same conclusion every time.

They don't want to own it, because they aren't in the business of developing and maintaining training.

Structured on-the-job training is real. Following Joe around is not.

IDCON publishes a simple sum that argues against relying on internal on-the-job training alone, and it sticks in the mind even though it carries no source. The trainee absorbs about 75% of what the veteran teaches, and the veteran teaches about 75% of what they actually know. Multiply those and you get roughly 56% of a decent training outcome.

There's also older research suggesting structured on-the-job training substantially outperforms the unstructured kind, though the original work is from 1992 and is a model of one plant rather than a measurement, so the direction is more useful than the numbers.

Internal is not one thing. A designed internal curriculum is a genuine option. Walking a new hire around behind a veteran for 6 months is not training, it's hope.

And one study that says internal wins outright

Jones, Woods and Guillaume (2016) pooled the research on workplace coaching and found internal coaches at δ = 1.40 against external coaches at δ = 0.20, with no overlap between the two ranges. The authors had predicted the opposite, and the result went against them. Their explanation is that internal coaches understand the organisation's culture and constraints.

The internal figure rests on 3 studies and 182 people, and the authors caveat it themselves. It also covers one-to-one executive coaching rather than technical instruction.

It goes in anyway, because you should see it.

A page that only reports the evidence supporting its own conclusion isn't worth your time.

Two kinds of training that fail whoever delivers them

One of these is usually delivered internally or by your software vendor. The other is usually bought in. Each fails for its own reason, and each has a different fix.

CMMS training that teaches transactions

Start with a finding almost nobody uses, and one you can check yourself.

The Plant Maintenance Resource Center surveyed CMMS implementations in 2000, with 87 responses, and again in 2004, with 105. Both surveys asked implementers what they should have spent more time and effort on.

"Effective training" came first both times, at 24.1% and then 19.0%.

The 2000 survey also asked people who were about to start an implementation what mattered most. Alexander Dunn's write-up puts it plainly:

"In comparison, those currently considering implementing a new CMMS place far greater importance on CMMS selection. Effective Training is not considered to be at all important."

Number one regret afterwards. Invisible beforehand.

Two independent samples, 4 years apart, same answer.

So what does that training usually consist of?

Now look at what most sites, including yours, actually buy. I saw this repeatedly at Shell early in my career, and I have seen it at client sites ever since. Technicians get trained in specific SAP transactions. How to raise a notification. How to look one up. How to check whether it's a duplicate.

Never the process flow.

Never why the step matters, and never what happens downstream if it's done badly.

The consequence is predictable.

People see it as bureaucracy, decide they're simply feeding a system, and get frustrated that they seem to exist to keep the CMMS fed rather than to do a job. And when someone doesn't understand why a step matters, they are less likely to do it, and far less likely to keep doing it once the pressure comes on.

Even a vendor will concede the point. Brightly Software's sales director told Automation World that "the focus is on getting users familiar with basic functionalities like work order generation. But this is just scratching the surface."

The fix is a sequence, and the order is the whole point:

1. What the process looks like, and why it looks like that
2. What good looks like
3. The steps, and why each one is there
4. What happens if you skip them
5. The person's specific role in that process
6. Last, the CMMS actions that support that role

Get that order right and your people understand how their work contributes to something.

Get it backwards and you've taught somebody to press buttons.

This is why our own training is deliberately not CMMS-specific. We teach the end-to-end work management process without the transactions, fields and templates, which means it works whether you run SAP, Maximo, JD Edwards or Infor, and whether you're in mining, oil and gas or heavy manufacturing. You still have to map it onto your own system, and that's a much smaller job than building the whole thing yourself.

One honest note. No survey anywhere separates system training from process training in maintenance, so the argument above rests on the implementation regret data and on experience rather than on a study that measured it directly.

Reliability-centred maintenance training that never leaves the classroom

This one usually arrives from outside, and it fails for 2 separate reasons.

Why does so much of it not land?

The first is that many of the people delivering it are not practitioners. They are working from the books, rehashing theory they have read rather than applied, and they cannot place any of it in a practical context.

The second is more interesting, because the content is genuinely counterintuitive. Everyone walks in believing that equipment wears out with age, and therefore needs more maintenance as it gets older. It is what experience seems to tell you.

Nowlan and Heap found in 1978 that 89% of items had no wearout zone at all. Only 6% showed clear signs of wearing out, and just 4% followed the classic bathtub curve.

That's a hard thing to accept the first time you hear it.

It's harder still to turn into a decision on Monday morning.

Put those 2 failures together and you get a very specific outcome.

People leave the room believing they understood it, quietly more confused than when they arrived, with no idea what to do with any of it. So they do nothing.

The irony is that the original 1978 report prescribed the opposite of the generic classroom you'd be sold today. Nowlan and Heap specified a one-week course on the actual equipment, plus "much more detailed training on the particular aspect of the equipment they are to analyze."

Dana Netherton, who chaired the SAE committee that wrote the RCM standard, described the buyer's problem exactly:

"if an organisation said that it wanted help in using or learning how to use RCM, it could not be sure what process would be offered."

Even NASA's own facilities guide says rigorous RCM is "labor intensive and often postpones the implementation of obvious PT&I tasks", and is "rarely needed for most facilities and collateral equipment items."

So here's our position, and it costs us work to say it.

We do not deliver true reliability-centred maintenance training, and we don't claim to. Very few providers genuinely do. What we teach is the theory, what the original research established, and how to apply it practically to build a PM program that works. Beyond that, our advice is consistent: run full RCM on a small percentage of your equipment, the most critical and the most painful, and when you do it, hire an experienced RCM facilitator. It is a specialist discipline and we have no intention of pretending otherwise.

What actually makes training stick

Blume, Ford, Baldwin and Huang (2010) is still the study everyone works from on whether training gets used once people are back at work. It pools 89 separate pieces of research, and nothing better has come along since.

The last column is a correlation, written ρ. It says how closely a factor tracks whether training gets used, on a scale from 0 to 1. Around .10 is negligible, .30 is a moderate link and .50 is a strong one. A minus sign means the factor points the other way, and "not significant" means the number cannot be told apart from zero.

Three things in that table should change what you measure.

Your feedback forms predict nothing. Whether people enjoyed the course scores .08 against whether they actually use it. Whether they believed it applied to their job scores .46, the highest number in the study. If you only measure satisfaction, you're measuring the wrong thing.

The action plan on the last afternoon does not work. Setting goals at the end of a course, and the exercises meant to stop people slipping back, show no measurable effect at all once the studies are pooled. Reinforcement has to be built into how the work is organised, rather than tacked on before everyone goes home.

And "our people are too old or too set in their ways" is not supported by anything. Age scores .04.

One warning about how those numbers were collected.

When the same person answers both the question about the cause and the question about the result, the numbers come out inflated. Work environment scores .54 when one person answers both, and .23 when the two answers come from different people.

Ford, Quiñones, Sego and Sorra (1992) measured this properly. They followed 180 airman and supervisor pairs 4 months after training, collecting the supervisor data separately. Same training, same standard, same trainees, and the opportunity to actually perform the trained tasks varied enormously. What drove it was the supervisor's attitude and the backing of the crew.

Hughes, Zajac, Woods and Salas (2020) found that support at work accounts for about a third of the difference in whether training gets used, and that backing from workmates is the single biggest part of it.

Which points at a practical conclusion for you.

Train crews, not scattered individuals.

The same programme, twice, with opposite results

We ran the same coaching programme in 2 regions of a global chemical manufacturing company. Same provider, same content, same material.

The only variable that differed was whether leadership showed up.

That is Blume's finding about the work environment and Ford's finding about the chance to practise, playing out in a plant, on the same budget.

Why we build courses by role rather than one size

Everyone involved in the process should be trained.

They shouldn't all be trained the same way.

The implementation team needs the full process plus how to implement and sustain it, which is largely change management. Planners, schedulers and supervisors need the end-to-end process in depth. Frontline technicians and operators need 2 or 3 hours covering what the process is, their part in it, and why it matters.

That last point is practical rather than philosophical.

Technicians complain about a 2-hour course. You'll never get them through a 30-hour one, and they don't need it.

The payoff is the thing Hughes measured. Everybody ends up using the same language, understanding the other roles, and agreeing on what good looks like.

The decay problem, and what the research really says

You will read everywhere that people forget 50% of training within an hour, 70% within a day and 90% within a week, always credited to Hermann Ebbinghaus.

So what did Ebbinghaus actually do in 1885?

He was his own single subject. He memorised nonsense syllables.

And he measured savings, meaning how much faster he could relearn a list, rather than how much he could recall.

His actual figures were 58% savings at 19 minutes, 33.7% at 24 hours, about 25% at 6 days, and roughly 21% still there after 31 days.

He never came close to 90% loss. Not after a week, and not after a month. The claim contradicts the researcher named in the same sentence.

His own words are better than anything I could write about it:

"Whether they possess a more general significance so that, under other circumstances or with other individuals, they might find expression in other constants I cannot at the present time say."

He explicitly declined to generalise.

The training industry has spent 140 years doing the thing he refused to do.

Thalheimer (2010) reviewed 69 forgetting conditions across 14 studies and more than 1,000 learners. Forgetting ranged from 0% to 94%. Within a single study it ranged from –41% to +63%, meaning some learners remembered more on the later test. His conclusion: rules of thumb showing people forgetting at a predefined rate are simply false.

The curve itself is real, and it was reproduced successfully in 2015.

It's a sound description of one man relearning nonsense syllables. It is not a measurement of what your technicians retain.

What the evidence does support

Two more symbols below. d and g measure the same thing as δ further up, so 0.2 is small, 0.5 is moderate and 0.8 or more is large.

Three rules come out of that, and they are specific enough to build on:

1. Space the reinforcement at roughly 10 to 20% of how long you need it to last. If you need it for a year, prompt at 1 to 2 months.
2. Err long, not short.
3. Ask people to produce the answer rather than re-read it, and prioritise whatever they got wrong confidently.

Two different problems, two different mechanisms

It would be easy to blur these together, so here's how they differ.

Our own coaching sprints run on a fortnightly rhythm across 12 months, and that interval is not derived from the spacing research. During live implementation people are applying the material continuously, so opportunity and support are what matter, not spaced retrieval. Claiming otherwise would be exactly the sort of after-the-fact justification this section is complaining about.

The spacing research is what we're using to design automated reinforcement into our new learning platform, which is a different problem entirely.

The honest gap

Vlasblom and colleagues (2020) reviewed how well people hold on to skills in safety-critical jobs. 39 of the 40 studies were medical. One was aviation.

Nobody has properly measured knowledge decay in an industrial maintenance workforce. The evidence is all medical, and everyone applying it to plant maintenance is stretching it past what was measured, ourselves included.

The language question

Language is regularly given as a reason to keep training internal.

Our people don't read English well enough for an external course, the argument goes. Does that hold up?

The concern is completely legitimate.

Someone who can hold a conversation in a second language often can't absorb the full content of a detailed technical course in it.

The conclusion drawn from it is usually wrong.

So what should you do instead?

We ran a program in Panama that started with 20 to 30 people. Once the whole thing was translated into Spanish, more than 200 went through it.

The answer is to require delivery in the local language from your provider, rather than falling back on internal training because you have given up. Ask any external provider what languages they deliver in, and treat "English only" as a real limitation rather than a detail. We cover why this matters more than most people assume separately.

So what should you actually do?

Which order should you do it in?

The sequence matters as much as the choice you make:

1. Buy the content and the credibility from outside.
2. Train the whole crew, at the depth each role needs, rather than scattered individuals.
3. Put your supervisors and your leaders in the room, because the evidence says they decide whether any of it survives.
4. Reinforce on a schedule, by retrieval, at roughly 10 to 20% of how long you need it to last.

Most organisations get step 1 right and stop there.

That's where your training money goes, and it has nothing whatsoever to do with who delivered it.

If you want to talk through what that would look like for your site, we're happy to have that conversation.