A task inventory you can finish in three weeks

The task inventory arrived in month two as a 61 page annexure. It listed 914 tasks across nine grades of maintenance technician, rated on seven scales by 1,120 people. Two hundred and six people had answered. The consultant presented the mean importance score for task 447 and the plant head asked the only question that mattered: which five tasks should we train first? Nobody in the room could answer it.

The inventory had done exactly what it was designed to do. It was complete. It was also unusable, because nobody had decided in advance how much evidence was enough to stop collecting it.

Why the census is the natural mistake

A task inventory sounds like it should be exhaustive. You are trying to describe a job, so you list everything the job contains. Every incumbent you ask adds two tasks the last one forgot. The list grows, the rating form lengthens, and the response rate falls.

That is not a guess. The team that built the O*NET questionnaires ran a mail survey of 15 job incumbents in each of 80 occupations and got what the project's own revision report calls a low return rate. The fix they applied was not better persuasion. It was cutting items, cutting the number of response scales per item, and simplifying instructions, explicitly to reduce respondent burden. A shorter form was the intervention.

The public O*NET programme does this at national scale, and it still does not attempt a census. Data come from incumbent workers at establishments selected by a random sample, roughly 200 occupations a year. If the richest occupational database in the world samples, your plant can sample.

The mechanism underneath is simple. The unit of analysis in a task inventory is the task, but the unit of work is the role. Two fitters in the same grade do not do the same job, and the difference is real, not noise. A list built to cover both is a list of everything anyone might do, which is not a training plan. A list built to cover the tasks that most incumbents in a role actually perform is a training plan.

Three ways to get a task list, and what each costs

There are only three families of method in common use. They differ in who you ask and how you stop.

The panel method is real and it is fast. The DACUM process uses six to eight high-performing incumbents over two days with a facilitator, and then explicitly validates the chart through peer review and management review. Notice the second step. The panel produces a draft, and validation is a separate activity, which is where most in-house panels stop and therefore where the method quietly becomes one room's opinion.

The two-stage sample is the version that survives a real HR calendar. Here is how to run it.

Stage one: sample roles, not people

Start from the roster, not from the org chart. Build the list of roles in the grade family, then stratify by three things: process or product line, shift, and whether the role is line or support. Pick one role from each cell, up to about twelve.

Twelve is not a statistical constant. It is a budget decision: twelve roles at two to three interviews each is 24 to 36 conversations, which two people can run in a fortnight while doing other work. If your grade family has 47 roles, you have sampled roughly a quarter of the role space, and you will say so in the output.

Write down the roles you did not pick and why. That list is not an apology. It is the part of the deliverable that lets the next cycle extend the inventory instead of rebuilding it. Getting the roster right matters more than it sounds. A roster that quietly excludes contract staff or a whole shift produces a tidy sample of the wrong people, which is why it is worth auditing the roster frame before you draw from it.

Stage two: two incumbents per role, three interviews maximum

Inside each sampled role, pick two incumbents from different shifts. If the role has fewer than four people, take whoever is available and record that the role is thin.

Interview one incumbent at a time, for 45 minutes, with one question: walk me through yesterday, in order. Do not ask what skills the job needs. Ask what happened, from gate to gate. Convert what you hear into task statements yourself, in the room, in the form verb plus object plus condition. "Replace the worn seal on the pump housing while the line is stopped" is a task. "Pump maintenance" is a heading.

Between interviews, merge into one running list per role. That list is the instrument, and it grows for the first two interviews and then usually stops growing.

The stop rule

This is the part that makes three weeks possible. Set the rule before you start and apply it mechanically.

Stop interviewing a role when both conditions hold:

1. The top five tasks, ranked by time spent or frequency, are the same five in each of the last three interviews.
2. No new task has been added to the role's list in the last two interviews.

Both conditions, not either. The first guards against a stable but incomplete list. The second guards against a list that is still growing while the top five look settled, which happens when the fifth task is genuinely contested.

Keep a one-page log so the rule is auditable:

The method borrows its logic from qualitative sampling, where saturation is the standard stop rule and a pre-set interview count is known to be unreliable. A 2024 secondary analysis of five interview studies found that near saturation arrived at 33 to 60 per cent of the planned interviews, and true saturation at 91 to 100 per cent. It also found that structured guides and deductive coding reach saturation sooner. You have both: a fixed question about yesterday, and a coding frame that is just verb plus object plus condition. That is the same logic as the cheaper three training needs analysis instruments, applied to a single grade family. That is the fastest legitimate route to a stable list.

It is worth being precise about what that evidence covers. Those were interview studies in health research, not task inventories in manufacturing. What transfers is the shape of the finding, not a guarantee. A stable list is a decision you are willing to defend, not a proven population parameter.

Rate less, and rate what you will act on

Once a role's list is stable, send it back to the incumbents you interviewed plus their supervisor, as a short form. Three columns only:

• Frequency: how often in a normal month.
• Criticality: what happens when it is done wrong.
• Failure point: where people actually get it wrong, chosen from four options (sequence, specification, tool, judgement).

The Interstate Renewable Energy Council's job task analysis guidance uses frequency, importance and criticality style ratings and then insists the task list be validated with a representative sample of incumbents. Three scales is enough to rank, and you can collect them inside the assessment platform you already run, which removes the separate spreadsheet step. Seven scales produces a spreadsheet nobody reads and a form people abandon halfway.

Cap the form at 40 tasks per role. If a role will not compress to 40, it is two roles, and you have found something more useful than a task list.

A worked example, with illustrative numbers

A mid-sized auto component plant, 1,180 employees, four grades of maintenance technician. The numbers below are illustrative, but the arithmetic is the point.

They sampled 11 of 47 roles. Twenty-six interviews over nine working days, run by two L&D staff and one supervisor each. The merged list came to 148 task statements, down from the 600-plus a full census would have generated. Sixty-one tasks cleared the top-five ranking in at least one role. The training plan that came out had nine modules and replaced a proposed 22-module catalogue.

The coverage page said: 11 of 47 roles sampled, 36 roles not sampled, of which 9 were flagged as high-variability and scheduled for the next cycle. It also said which shift was missing from two roles because the plant runs a three-shift pattern and only two were covered.

That page is why the plan survived review. The plant head could see the edge of the evidence.

Where this method is wrong

The two-stage sample assumes tasks are shared across incumbents in a role. Where that assumption breaks, the method under-describes the job.

Three cases break it. Field service and territory roles, where each person's route is effectively a different job. Research and development roles, where the task mix changes with each project. And any role in its first year after a process change, where the list is not yet stable enough to have a top five.

For those, sample the situations rather than the people: the last twenty service tickets, the last ten project charters, the last month of changeover records. The stop rule changes with it. You stop when the last five records have added no new task, not when three interviews agree.

The other honest limit is the one already stated. Saturation evidence comes from interview research, and it tells you when new themes stop appearing. It does not tell you that your task list is complete in any absolute sense. It tells you that more interviews in the same role are unlikely to change your top five. That is the decision you are actually making.

What to do on Monday

Pick one grade family. Pull the roster, list the roles, and mark the cells you will sample. Then run three interviews in one role, build the list, and apply the stop rule out loud in front of whoever owns the training budget. If the rule holds, you have a defensible sample and a plan in three weeks. If it does not, you have learned something specific about that role's variety, which is the finding you needed before you spent eleven weeks chasing a census.

The inventory is not the deliverable. The coverage statement is.