Competency gap analysis that ends in a one-page plan

Twelve branch managers, six competencies, one page. The framework was signed off in September. In October the L&D head sent me the gap analysis: a spreadsheet with seventy-two ratings, colour-coded red, amber and green, and a summary that said the biggest gaps were customer service and communication. The training plan for the next quarter was customer service and communication.

I asked one question. Which branch manager, on which behaviour, did you actually watch? Nobody had watched anything. The ratings came from a form the managers filled about themselves, moderated by a regional head who had met most of them twice.

The arithmetic was fine. The inputs were guesses.

Two judgements, one subtraction

A competency gap is not a measurement. It is a subtraction of two judgements: the level the role requires, and the level the person currently shows. Both are estimates. Both carry error. When you subtract one estimate from another, the errors add rather than cancel, so the number at the end is the least reliable thing on the page.

That is not a reason to abandon gap analysis. It is a reason to stop treating the number as the finding and start treating it as a way of ordering a conversation.

The best evidence on self-assessment says the error is not random. Kruger and Dunning tested people on humour, grammar and logic across four studies. Participants in the bottom quartile scored at the 12th percentile and estimated themselves at the 62nd. The people with the least skill were the furthest off, because the same missing skill that produced the bad answer also prevented them from recognising it. A self-rated gap analysis therefore compresses at the top: the weakest performers look closer to the standard than they are, and the strongest look closer to the ceiling than they are. You get a page of tidy numbers that describes almost nothing.

There is a second failure, quieter than the first. Even with good ratings, most teams rank gaps by size alone. The largest raw gap wins. But a large gap on something the business does not currently pay for is a hobby, and a small gap on something that decides whether a branch hits target is the whole job. Size and importance are different axes, and only one of them is on the spreadsheet.

Step one: cut the page to six rows

You cannot run this on a full framework. Twelve competencies at four levels gives you forty-eight cells per person before anyone has a conversation. Six rows is the working limit for a page a manager will read on a Monday.

Choose the six by asking what the function must do differently in the next two quarters. Not what the role involves. What must change. If the answer is "sell more of the new product", the rows are the behaviours that move that number. If the answer is "survive an audit", the rows are the audit behaviours. Everything else stays in the framework and stays out of this page.

Then write the required level as a sentence, not a number, and only then convert it to the scale. "Level 3" means nothing in a calibration meeting. "Can run a pipeline review that ends with two named actions and a date, without the regional head in the room" is something two managers can disagree about, which is the point. If you have not built the framework yet, the build order in how to build a competency framework managers will actually use matters here, because a framework with fuzzy anchors cannot produce a sharp required level.

Use a four-point scale and define each point once, for all six rows:

Four points, not five. A middle option invites the rater to avoid deciding, and the whole exercise depends on people deciding.

Step two: rate against evidence, not memory

Every rating on the page gets one line of evidence. Three kinds are acceptable.

• Work artifact. A pipeline review sheet, a credit note, an escalation log, a staff rota the manager built. The document exists and someone can read it.
• Observed episode. A meeting, a customer call, a branch walk-through where the rater watched the behaviour happen. Write the date.
• Standardised task or simulation. A scored exercise, an in-basket, a role play with a rubric.

Anything else is opinion, and opinion is allowed only when it is labelled as such. A page where nine of seventy-two cells have evidence is a page that tells you where to look next. A page where all seventy-two cells are opinions is a page that will produce a training plan for communication.

Two practical rules save the most time. First, rate only the people whose gap will change a decision. Six people across six rows is enough to plan with. You are not running an appraisal. Second, have the manager and the person rate separately before they talk, then reconcile in a fifteen minute meeting. The reconciliation is where the useful information lives, because the disagreements point at either a vague anchor or a genuine blind spot. If you want a structured way to handle the second case, the questions in six 360 feedback questions that actually surface a blind spot transfer directly to that conversation.

Step three: weight before you rank

Assign each of the six rows a weight of 1, 2 or 3 against the two-quarter objective. The rule for the weight has to be stated in one sentence in the page header, because an unexplained weight is the easiest thing for a sceptical operations head to attack.

A weight that survives scrutiny looks like this: "Weight 3 means the business objective fails without this behaviour in at least half the branches. Weight 2 means it slows the objective. Weight 1 means it affects quality but not the objective this quarter."

Then compute the weighted gap: (required minus current) times weight. That single multiplication changes the ranking, and it changes it in the direction the business cares about.

Figures are illustrative, built to show the method rather than a real client.

Read the table the way the branch head will. The largest raw gap is credit appraisal at 1.4, and it stays near the top because the audit is real. The CRM gap of 1.0 is the third largest raw gap and it drops to fifth after weighting, because the objective this quarter is sales, not reporting. When we looked at the actual CRM problem, nine of the twelve managers were fighting a report that required three exports and a manual merge. That is a process defect wearing a training costume, and no course fixes it.

The coaching gap is not the largest. It is the most expensive one to leave alone, because every other row on the page depends on a branch manager who can hold a useful conversation with a teller who is missing target.

Step four: route each gap to the right fix

Training is one of five routes. Choosing it by default is how L&D budgets get spent on problems that were never learning problems. CIPD makes the same point in its learning needs analysis guidance: consider all the appropriate ways to meet the need rather than defaulting to a course.

The "cluster low across everyone" signal is the one teams miss. If eleven of twelve managers cannot do something, the first hypothesis is the system, not the cohort. Twelve identical skill deficits are rare. One broken process creating twelve identical symptoms is ordinary.

The transfer condition nobody puts on the plan

A funded gap still fails if the branch does not support the new behaviour. Blume, Ford, Baldwin and Huang pooled 89 empirical studies on transfer of training. Corrected correlations with transfer were .31 for supervisor support, .27 for transfer climate and .14 for peer support. Compare that with .37 for cognitive ability, which is the thing most selection and training budgets are built around.

The practical reading: who the manager's boss is, and whether the branch rewards the new behaviour, matters about as much as how smart or how able the manager is. So the last column on the page is not "course name". It is "what changes in the branch so this sticks". For the coaching row, that was a standing item in the regional head's monthly review and the regional head attending two coaching conversations per branch, not a two-day workshop with a workbook. If the answer to that column is blank, do not fund the row.

If you are buying or building the assessment side of this, the workflow in our features page is built around evidence capture at the point of work rather than a self-rating form, which is the difference that decides whether the page is worth reading.

Where this stops working

The ratings are ordinal. Treating 1 to 4 as if the distance between 1 and 2 equals the distance between 3 and 4 is a convenience, not a fact. Averaging twelve of those ratings produces a number with a decimal point that implies precision the scale does not have. Do not report a weighted gap to one decimal place to a CFO and expect to defend it. Round to whole numbers when it leaves the room.

Twelve managers is not a sample. Nothing here reaches statistical significance, and a mean gap of 0.2 on a four-point scale is indistinguishable from rater noise. The page is a prioritisation device for a known group of people, not an estimate of a population.

And the page does not tell you whether the training worked. It tells you where to spend and what to re-observe. Proving that the number moved, and that the movement was yours, is a separate job, covered in your training worked, you just cannot prove it. Without that step you will be running this same analysis next quarter with the same rows and a new colour scheme.

Do this before you book anything

Take the framework you already have. Pick six rows against next quarter's objective and write the required level for each as a sentence a colleague could disagree with. Book ninety minutes with two branch heads, rate three managers each using one work artifact per row, and reconcile the disagreements out loud.

If the page still says communication, you have a framework problem, and the fix costs nothing. If it points at a process defect and one expensive coaching gap, you have a training plan, and it will be about a page long.