the job description is the wrong unit of analysis
job titles average away the variation that matters, and the decisions enterprises now face need a smaller unit than the job.
Competency frameworks follow a schedule that is close to universal. Months one and two go to design workshops. Months three to five go to drafting profiles with subject matter experts. Month six is launch, with a communications plan and a manager briefing. Month seven is when the first line leader says the framework no longer describes the job, and the framework begins its slow death by exception.
Nobody kills it. It simply stops being used for decisions. Two years later a new HR leader commissions a replacement, and the cycle starts again with a fresh set of workshops.
how fast role requirements move
These frameworks are not badly built. They are built against a target moving faster than the build.
Lightcast's research on the speed of skill change in the United States found that the average job saw around one third of its required skills change between 2021 and 2024. For the fastest-moving quarter of occupations, roughly 75% of skills changed over the same three years. Lightcast's earlier Shifting Skills work found that 37% of the top 20 skills requested for the average job had changed since 2016, and around one in five was entirely new.
The forward-looking numbers agree. The World Economic Forum's Future of Jobs Report 2025 found employers expect 39% of key skills to change by 2030. LinkedIn's Work Change Report, published in January 2025, puts it higher still, projecting that 70% of the skills used in most jobs will change by 2030.
Take the conservative end of that range and assume role requirements drift at roughly 10% per year. A framework rewritten every three years carries an average error of about 15%, and a peak error of 30% in the months before the rewrite lands. No governance process survives being wrong about a third of its own content.
the mismatch between documents and drift
The structural problem is that a framework is built as a document and maintained as a project. Documents have versions. Workforces have rates of change. The two do not reconcile, and the version always loses.
Deloitte's research on skills-based organizations captures the resulting gap. In its 2023 Global Human Capital Trends work, 93% of respondents said moving away from a job-centric focus was important or very important, while only 20% said their organization was very ready to do it. Intent is nearly universal. The operating model is not.
three layers, three clocks
The fix is architectural rather than editorial. Separate the parts of the framework that change at different speeds and give each one its own refresh clock.
- Layer one, the capability spine. Twelve to twenty capabilities describing what the organization has to be good at. Credit risk judgment. Customer solution design. These change when the business model changes, roughly every three to five years.
- Layer two, role profiles. For each role, the skills required and the proficiency expected. These change annually.
- Layer three, evidence and tooling. The specific platforms, methods, certifications, and artifacts that demonstrate a skill. These change quarterly, sometimes faster.
Most frameworks rot because layer three content ends up in layer one. A capability named after a vendor product is obsolete the moment procurement changes suppliers. A capability named after a decision the business has to get right will outlive three tool migrations.
Run the test on your own framework this week. Read each capability and ask whether it would survive replacing every system you own. Anything that would not survive belongs a layer down.
a maintenance procedure you can run monthly
Once set up, the following takes one analyst about two days a month. It replaces the rewrite entirely.
set a drift budget
Decide the maximum acceptable staleness for each layer and publish it. A reasonable starting point is 36 months for the spine, 12 months for role profiles, and 3 months for evidence and tooling. The budget becomes the standard you will be held to, so pick numbers you can meet with the people you have.
instrument three signals
You need evidence of drift that does not depend on someone raising their hand. Three sources are usually available already. Internal requisitions show what hiring managers ask for when they are spending money. External postings for the same role families show where the market has moved. Assessment and project data shows which skills people are being asked to demonstrate in the work itself.
run the diff
Once a month, for each role family, produce two lists. First, skills appearing in requisitions and work evidence but absent from the profile. Second, skills sitting in the profile that have appeared in zero requisitions for twelve months. The first list is what the framework is missing. The second is what it is carrying for no reason.
apply promotion and demotion rules
Rules stop the monthly diff from becoming a monthly debate. A workable pair: a skill appearing in more than 30% of requisitions for a role family across two consecutive quarters is promoted into the profile. A skill absent from all three signals for four consecutive quarters is demoted to an archive rather than deleted, so the history stays auditable.
Publish the rules before the first diff runs. Rules agreed in advance are policy. Rules agreed afterwards are negotiation, and negotiation is what killed the last framework.
batch changes and publish on a fixed date
Continuous change to a framework is as unusable as no change at all. Batch every accepted change into a monthly release on a fixed calendar date, with a changelog. Line leaders own their role profiles. A central team owns the spine, the rules, and the release. Everything downstream, including learning systems and pay structures, then has a stable object to point at.
report framework health like uptime
Track one metric and put it in the HR operations review: the percentage of role profiles updated within their drift budget. Report it monthly alongside everything else that gets measured. A framework at 92% is healthy. A framework at 40% is a document, and everyone acting on it is acting on stale information without knowing it.
the benchmark worth copying
The United States Department of Labor sponsors O*NET, which maintains around 900 occupational profiles covering more than 55,000 jobs. Portions of every profile are refreshed annually, and faster-moving elements such as job titles and software skills are refreshed more often than that. The program treats occupational description as continuous data collection, drawing on job incumbents, occupational experts, employer postings, and analyst review.
A national labour agency has concluded that occupational descriptions cannot be maintained as periodic projects. Enterprises with far fewer roles, and far better direct evidence about them, still run the three-year rewrite.
what to stop doing
Three habits guarantee a month seven death, and dropping them costs nothing.
Stop writing forty-skill profiles. Long profiles are impossible to keep current and impossible to assess against, so they get ignored twice over. Eight to twelve skills per role is enough to support a decision.
Stop using five-level proficiency scales whose levels are not behaviorally distinct. If two experienced assessors cannot independently place the same person at the same level, the scale is decorative, and every downstream calculation inherits the noise.
Stop treating the rewrite as the solution. The triennial rewrite is what maintenance looks like in the absence of a maintenance process. It is expensive, it spends the goodwill of the same subject matter experts every time, and it produces a framework that begins decaying on launch day.
None of this needs new technology, though continuous drift detection is far easier when skill signals are extracted from work evidence rather than gathered in interviews. That is the practical argument for treating competency framework creation as an ongoing data problem rather than a consulting engagement with a delivery date.
A framework that is 85% accurate and updated monthly beats a framework that was 100% accurate in March two years ago. The first one gets used, and use is the only measure that matters.