A training need is a claim about failure. Check the record
Every claimed gap implies a failure rate. Find that failure in a quality, rework or incident register, code it by type, and design training only for the part it can fix.
Rajendra Kumar · 16 November 2026 · 9 min read

Representational image.
The TNA says blown film operators need training on machine setup. Thirty-four operators, three lines, three days of offsite training, quoted at two and a half lakh. The plant head signs it because setup errors are a known nuisance, and the request came from the production manager, who has been saying so for two years.
Nobody asks the question that would have taken an afternoon: what is the error rate, and where is it written down?
A training need is not a subject. It is a claim about failure. Someone is asserting that a defined group of people, doing a defined task, fail at it often enough, or expensively enough, to justify the cost of fixing them. That claim can be true or false, and it can be tested against a record the plant already keeps for other reasons. If the failure is not in the record, the claim is a perception, and perceptions need a different intervention.
The reflex that produces a training request
Incident and nonconformity processes are built to close. An investigation has a deadline, a form, and a signature. When the cause section lands on a person, the fastest action that looks like a cause is retraining, because it is scheduled, documented and defensible in an audit.
Practitioners in safety have been pushing back on this for years. Writing in Professional Safety, Ryan Word puts it plainly: employee error is inevitable, and "retraining is not a corrective action." His argument is that a human cause investigation has to continue past the person and into the conditions that made the error likely, or the same incident returns on the next shift with a different name on the form.
The work of an L&D manager starts one step earlier than that. Before you accept a training request, find out whether the failure it describes is a failure at all, and if it is, what kind.
Three failure types, and only two that training fixes
The UK Health and Safety Executive's human factors guidance splits human failure into errors and violations, and splits errors again. The split matters because each type has a different control, and training is the right control for only part of the list.
| Failure type | What happened | Does training fix it | What actually works |
|---|---|---|---|
| Slip or lapse | The intent was right, the action went wrong. Skilled, familiar task | No, largely | Task and interface design, checklists, cross-checks, removing interruptions |
| Rule-based mistake | A good rule misapplied, or a bad rule followed correctly | Partly | Procedure rewrite, better scenario coverage, drills |
| Knowledge-based mistake | No rule existed for the situation, so the person improvised | Yes | Training on the underlying system, decision aids, supervised exposure |
| Violation, routine or situational | Deliberate shortcut, usually well meant | No | Removing the reason for the shortcut, supervision, rule redesign |
The HSE guidance is explicit on the first row: slips and lapses "cannot be eliminated by training, but improved design can reduce their likelihood." It is equally explicit that these errors often happen to experienced, well-trained, well-motivated staff, which is why they cluster in routine maintenance and changeover work rather than in the tasks people are bad at.
Read that table against a typical training request. If most setup errors are lapses on a familiar sequence, three days of classroom training will produce good feelings and no drop in the error rate. If most are knowledge-based mistakes on a machine nobody was trained on, the request is right and scoping is the only question.
The check, in the order the work happens
1. Write the claim as a rate
Rewrite the request as one sentence with numbers in it: "Operators on lines 2 and 3 make setup errors on 3 percent of changeovers." A claim in that shape can be inspected. "Operators need setup training" cannot.
2. Find the denominator before the numerator
The denominator is the unit of exposure: changeovers completed, batches run, calls handled, tickets closed. It is almost always available in a production system, a ticketing tool or a shift log, and it is the number that makes the failure rate interpretable.
This is where most validation attempts die. Quality registers count bad outcomes. They rarely count total attempts. You will need to reconstruct the denominator from a system the quality team does not own, so plan on two people and a week.
3. Pull the events, one line each
From the quality, rework, scrap, complaint or incident register, pull every event that matches the claimed task over a defined window. Twelve months is a reasonable default: long enough to catch seasonal work, short enough that the procedure and the machine set have not changed underneath it.
Build one line per event: date, line, shift, person or crew, what the record says happened, and cost where the record has it.
4. Code each event by failure type
Now read the event descriptions and assign the table above. This is a judgement call, so do it with the supervisor who knows the work, and write the reason in the row. Two people coding independently and disagreeing on more than a fifth of rows means the record cannot support the claim. That is itself a finding.
5. Compare the shape of the answer to the shape of the request
A request for setup training implies knowledge-based mistakes. If the coded events come back mostly as lapses, the request is aimed at the wrong failure. If they come back mixed, the training is one part of a package, and the size of each part is now visible.
6. Check the people with no events
This is the cheapest step and almost nobody runs it. Count who in the named group has zero events in the window, and how many exposures each had. If fifteen of thirty-four operators completed thirty-six changeovers each with nothing in the register, then for that subgroup the record is compatible with a rate well under 8 percent, and the claim "everyone sets up wrong" is dead. The statistical basis is the rule of three: seeing zero events in n exposures caps the long-run rate at roughly 3/n with 95 percent confidence, as Hanley and Lippman-Hand showed in JAMA in 1983. At 36 exposures the ceiling is about 8 percent. At 100 it is about 3 percent.
A worked example, with illustrative numbers
These numbers are invented to show the method. They are not a case study.
A packaging plant claims a setup training need for 34 operators across 3 lines. Over 12 months the lines ran 1,240 changeovers. The quality register holds 34 setup-related scrap events, costing 4,000 to 60,000 rupees each depending on how far the run went before anyone caught it.
The event rate is 34/1,240, or 2.7 percent, with a 95 percent confidence interval of about 2.0 to 3.8 percent. That is the number the claim has to be argued against. It is not zero, so there is a real problem. It is also not the picture the phrase "setup errors" summons, which is a workforce that broadly cannot do the task.
Coding the 34 events with the line supervisor gives:
| Coded cause | Events | What it implies |
|---|---|---|
| Slip or lapse during a familiar sequence | 21 | Design and checklist problem |
| Rule-based mistake, old procedure | 6 | Procedure and drill problem |
| Knowledge-based mistake, new die set | 4 | Training need, narrow |
| Deliberate shortcut to hit changeover target | 3 | Target and supervision problem |
Nineteen of the 34 operators appear at least once. Fifteen have no event in the window. Some of those fifteen are on line 1, which runs a single product and changes over rarely, so their exposure differs and the comparison has to be made within line.
The honest conclusion: four events, possibly six, point at a knowledge gap on one die set introduced in the middle of the year. That is a two-hour session for the operators who run that die, not three days for thirty-four people. The larger block of thirty events points at the changeover procedure, the interruption pattern during shift handover, and a changeover time target that quietly encourages shortcuts.
When the record proves nothing
Three outcomes are common and each has a different response.
Zero events. A clean record is not proof of no problem, and the rule of three tells you how clean. Zero events in 40 exposures leaves a rate under 7.5 percent entirely plausible. Widen the window, pool a second site, or accept that the record has not settled the question and look for a direct measure instead.
No record at all. Plenty of real competencies produce no incident trail: quality of a customer conversation, accuracy of a forecast, how a supervisor runs a shift review. The register will be empty because nobody logs it. Here the validation moves to structured observation or a sample of outputs, and the honest thing to write in the TNA is that the need rests on observation rather than on a failure record.
A record that blames people. If every event row reads "operator error, counselled," the register carries no cause information you can use. Coding it will produce agreement statistics that mean nothing. Fixing that is a quality systems task that will pay for itself well beyond this one request.
The WorkSafe guidance on human factors in investigations makes the point that human failure alone should not be capable of producing a serious event, because barriers exist to catch it. Applied at plant level: if a person's mistake reached the customer, the mistake is one cause and the missing check is another. A TNA that only names the person will keep buying training for a barrier problem.
The honest limit
This check only works for performance gaps. A developmental need, meaning a competency someone will need for a role they do not yet hold, has no failure record by definition, and no amount of register reading will produce one. Regulatory and statutory training also sits outside the test: the requirement is the justification, and the record is about coverage, not causation.
Reporting bias is the other limit. Registers undercount, and unevenly: near misses and caught errors are reported far less than customer-facing failures. Treat the rate as a floor, not a measurement, and say so in the TNA.
And the check costs time: two people for a week, against a budget that is already approved. It routinely removes more spend than it consumes, and it produces something a plant head can act on.
What to do with the next request
Take the most recent training request on your desk and write it as a rate in one sentence. Then answer three questions before you open a vendor conversation: what is the exposure count, where is the event list, and which failure type does the request assume. If you cannot answer the first two, the request is not yet a need. Send it back with those two questions attached, and ask the requester to bring the register extract to the next review. That request will come back smaller, better aimed, and much easier to defend when someone asks what the money bought. If you want a place to hold the coded events and the rates that come out of them, a structured assessment record does that job better than a spreadsheet, which is roughly what the platform is built around.
Two related reads: how to separate a stated want from a measured need in the first place, in the difference between a TNA and a TNI, and how to report the resulting rate alongside the rest of your numbers in a six-line L&D scorecard.
Next step
Run the same measurement on your own team
Bring a real team and a real competency. We set up the assessment, run one cycle, and walk you through the report so you can judge the output yourself.
Written by
Rajendra Kumar
Head of Enterprise Business, Conceptial
Runs enterprise work at Conceptial, the training and consulting firm behind BolderBrain. Two decades of building assessment and development programmes for Indian enterprises.