A competency framework for 40 people: how much is enough

A 40-person company I worked with had a competency framework with 22 competencies and four levels each. It came from a template the HR manager bought, and it was well written. The problem showed up in the first appraisal cycle: every manager rated everyone at level 2. With one person in most roles there was nobody to compare against, and the gap between level 2 and level 3 was a paragraph nobody could point to in a normal week. Six weeks went into the document. It changed no decision.

The instinct is to fix the writing. Cut it to twelve, rewrite the anchors, run a calibration session. That is right for a 400-person company and wrong for this one. At 40 people the binding constraint is not how well the framework is written. It is that there are not enough people, and not enough managers, to calibrate anything. So the answer to "how much framework is enough" is smaller than any template will tell you.

This post sizes a framework from the decisions you actually make, and gives a rule for when to stop. It is for the 30 to 60-person company where one person holds most roles. If you have 400 employees, the full build method for a larger company is the better read.

The constraint that changes everything at 40 people

A framework is a comparison device. Its job is to help two managers reach the same verdict about two people. In a large company you buy that agreement with machinery: several incumbents per role, a pool of raters, a retranslation sort, a norm group, a criterion study. Each piece needs volume.

At 40 people, none of it is available. Most roles have one incumbent, so there is no within-role comparison. There may be four managers in the whole company, and you cannot spare them for a sorting exercise without closing the business for a day. Hiring runs at a handful a year, so you cannot build a local norm group. A percentile against a global sample answers a question you did not ask.

What survives is one question: does the framework change a people decision, and can two people observe the difference it names? CIPD's competency framework factsheet says to balance detail with flexibility and avoid an overly prescriptive approach, advice that sharpens as the company shrinks. One vendor guide puts the point where informal judgement stops working at about thirty people and recommends a short framework for one or two role families. The response should still be small.

Size the framework from the decisions, not the catalogue

Frameworks grow because they are built by listing competencies. A workshop generates a long list from job descriptions, each function asks for its own, consolidation adds more, and nobody is ever asked to remove. The count is set by the catalogue, not the work.

Start somewhere else. Write down the people decisions you actually made in the last twelve months: hire, promote, move someone onto a bigger client, put someone on a performance plan, let someone go, set a pay band. This list is short: usually four to six decisions.

For each decision, write the tie it had to break. A decision needs a competency only when you were choosing between two people, or between a person and their own earlier performance, and would otherwise have treated them the same. "We had two finalists for the operations lead role and could not say why one was better" is a tie. "We should measure strategic thinking" is not.

Then work backwards to the smallest set of competencies that breaks those ties. Each must name a behaviour someone can watch, and at least two people must be able to see it in a normal month.

Five decisions, four competencies, because the third and fourth rows turn on the same behaviour. That is the count, not a target to grow toward.

Compare that with published anchors. The IAEA framework, built for a large international scientific organisation, restricts itself to four core and eleven functional competencies, and assigns an average of three to five functional competencies to any given job. The OECD uses six core competencies and three mastery levels aligned to job role and grade. Vendor guides typically recommend eight to twelve competencies per role. All three are ceilings for organisations with more roles and raters than a 40-person company has. The transferable part of the IAEA number is not the fifteen total, but holding each job against three to five competencies rather than the whole catalogue.

The stop rule: a competency has to change a decision

Adding to the count needs a test, because the pressure to add never stops.

For each candidate competency, ask which decision changes if you include it. If the answer is none, or "it gives a fuller picture", drop it. "Fuller picture" produces the nineteenth competency. One vendor guide puts it plainly: a framework that feeds no decision is a document, not a standard.

Then apply two more tests. First, can two managers watch the same month of work and disagree about this competency in a way a third person could adjudicate? If not, it is an adjective. "Client focus" fails. "Returns to the client with a revised date before the client asks" passes. Second, does the competency split a decision you already make, or imply a new one? Competencies imply processes. A succession competency implies a succession review, and a four-level leadership competency implies a calibration meeting with several people at each level. If you will not run that process this year, the competency is a promise you will not keep.

Levels are capped by the people you can compare

Four levels look like precision. They are a claim that you can tell the boundary between adjacent points, and that needs several people near each boundary. At 40 people with one incumbent per role, you have no boundary.

Use two levels, not four, named by scope rather than quality: "does it on their own work" and "does it through others". "Developing, proficient, advanced, expert" is a ladder nobody can place anyone on without an argument. The OECD aligns its three mastery levels to job role and grade, and the IAEA binds competencies to occupational roles rather than quality adjectives. Both anchor to something observable, and a 40-person company has roles too. Do not write a level 3 "across the function" if the function is one person plus a manager. An anchor nobody can see becomes a rating nobody can defend.

One evidence source per competency, and keep it cheap

A competency with no evidence is an opinion with a definition attached. Name one primary source of evidence for each competency, and one secondary, then stop.

At 40 people, prefer artefacts that already exist over instruments you have to buy. A scored assessment that returns a percentile needs a norm group of comparable people, and with one incumbent per role you do not have one. The percentile guide covers why that number misleads when the comparison group is wrong. Work samples, status threads, client emails and post-mortems are cheaper and closer to the behaviour.

No new forms: the four managers who would fill them are already doing the work. If you map a scored result onto a level, keep it to one result per level. The one-result-one-level mapping guide covers the drift when the same score is counted twice.

A worked example, with illustrative numbers

These numbers are illustrative, not a benchmark.

A 42-person B2B services firm, four managers, three role families: delivery, client and operations. They started with a bought template: 22 competencies, four levels, 88 anchors. Over twelve months they had made 7 hires, 2 promotions, 1 performance plan, quarterly assignment of 6 client accounts, and 1 exit. Five distinct decisions.

They ran the count in one afternoon with the four managers. Five decisions, four competencies, because two turned on the same behaviour. They set two levels named by scope, attached one existing artefact to each competency, created no new forms, and deleted 18 competencies, two levels and 80 of the 88 anchors.

The next two promotions were argued in terms of the four behaviours, and the HR head could say why one person got the role without a personality adjective. Six months later, two competencies were rewritten because managers kept disagreeing about one. That is a working framework: small enough that disagreement is visible and can be fixed. A 22-competency template hides disagreement in the volume.

The honest limits

This gives you a framework you can use and defend. It does not establish that your competencies predict performance. That needs a criterion study: measure the competency, measure an independent outcome such as error rate or client retention, and look at the correlation. At 40 people you cannot run one, so do not claim validity. The honest claim is that your managers agree on these behaviours and can point to evidence.

It will not survive growth unchanged. Past 60 to 80 people, or as soon as a role gains a second incumbent, levels become calibratable and a norm group starts to exist. Do not build for that future now. Set a review date instead.

It does not fix incentives. If ratings are tied to a fixed pay budget, managers will rate to the budget. And four competencies are a floor for decisions, not a portrait of a person. Someone can be excellent at work the framework does not name.

What to do this week

Write down every people decision you actually made in the last twelve months. Five lines, no more. For each, write the tie it had to break in one sentence. If you cannot name a tie, that decision did not need the framework.

Circle the competencies those ties need and count them. That number is your framework, and everything else in the template goes. Take the list to your managers for one afternoon and ask each to name a real example from the last month. Any competency nobody can point to comes off the list.

Write two levels named by scope. Put the whole thing on one page. Set a review date twelve months out, or for when headcount crosses 60.

Count before you cut. The number is probably four, and almost certainly smaller than the template you were sold. If you are rebuilding the assessment stack around those four competencies, start with what the platform actually measures rather than the catalogue.