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.
Ask most enterprises for their workforce plan and you will be sent a spreadsheet. It has a row per team, a column per quarter, and numbers that total to a budget. It is a useful document. It is not a plan, because it records an outcome without recording any of the decisions that produced it.
A workforce plan is a set of supply decisions. For every capability the business needs, someone has to decide where that capability will come from, and say what would change the answer. The spreadsheet hides that decision inside a headcount number, which is why plans survive review and then fail in execution.
The confidence data reflects this. Deloitte's 2025 analysis of workforce planning cites Gartner research finding that only 29% of chief human resources officers were confident in their ability to deliver on strategic workforce planning goals. McKinsey's 2020 survey on reskilling found that around nine in ten executives and managers said their organizations either had skill gaps already or expected them within five years, with 43% reporting existing gaps and 44% expecting them to appear.
four ways to supply a capability
There are four routes. Build, buy, borrow, and bot. Every capability in the plan gets assigned to one of them, on the record, with a stated reason. The value is not in the four words. It is in the requirement to choose, and to name the evidence that would flip the choice.
build
Build means developing the capability in people you already employ. Build is right when three conditions hold together. The capability is durable, meaning you will still need it in three years. Adjacency exists, meaning somebody is within one or two skills of it. And the learning curve is shorter than the hiring cycle for the same skill, which in tight markets it often is.
Build fails when it is chosen from sentiment rather than evidence. If nobody is within reach, build is not a plan. It is a hope with a training budget attached to it.
buy
Buy means external hiring. Buy is right when the capability is new to the organization, no internal adjacency exists, and the skill is durable enough to justify a permanent seat. Buy is also right when you need a teacher. Hiring two or three people who can build a capability in others is usually cheaper than hiring twenty who cannot.
Buy is over-used because it is the default. In the McKinsey survey, hiring was the most common response to skill gaps, cited by around two-thirds of respondents, with skill building second at 56%.
borrow
Borrow means contractors, agencies, partners, and internal secondments. Borrow is right when demand is spiky or time-boxed, when a skill is scarce and you need it before you can build it, or when you want to test whether a capability is worth owning permanently.
Borrowing a capability you intend to keep is a decision to pay a premium indefinitely. Borrowing to buy time while you build is a sound sequence. The two look identical in the budget and differ entirely in intent, so write the intent down.
bot
Bot means automation, including but not limited to large language models. This leg needs evidence rather than enthusiasm, and the evidence is now reasonably good.
Eloundou and colleagues, publishing in Science in 2024, estimated that around 80% of the United States workforce could have at least 10% of their work tasks affected by large language models, and that about 19% of workers could see at least half their tasks affected. The International Labour Organization's 2025 refined global index of occupational exposure found that roughly one in four workers globally sits in an occupation with some generative AI exposure, while only 3.3% of global employment falls into the highest exposure gradient.
That second study matters for anyone planning in Southeast Asia, the Middle East, or India. Exposure varies sharply with country income. Around 34% of employment in high-income countries falls into one of the four exposure gradients, against 11% in low-income countries. Automation assumptions imported from a North American headquarters will overstate the near-term opportunity in most of these markets, and a plan that budgets for the imported number will under-hire.
Bot is right when four things are true. The work is high volume. The task is rule-bound or text-bound. The output is checkable by someone who would notice an error. And the cost of an undetected error is bounded. Where error cost is unbounded, automation is a review-capacity decision, not a headcount reduction.
the four questions that pick the route
Score each capability on four variables before assigning a route. With the right people in the room this takes about twenty minutes per capability.
- Durability. How many years will we still need this capability? Score it under one year, one to three years, or over three years.
- Adjacency. How many people are within one development cycle of it? Score zero, one to five, or more than five. If you cannot answer, that inability is itself the finding.
- Time to value. Compare your hiring cycle and your learning cycle for this capability in weeks, using your own historical data rather than a published benchmark.
- Task structure. What share of the work is routine, checkable, and expressible in text or rules? Estimate it from a task inventory, never from a job description.
Then apply the rule. High durability with high adjacency is build. High durability with zero adjacency is buy. Low durability is borrow, whatever else is true. A high routine share with bounded error cost is bot first, and whatever work remains goes back through the other three questions, because what is left after automation is a different job with a different skill profile.
For each decision, write the single fact that would reverse it. That sentence is what makes the plan reviewable three months later.
modeling demand without last year's headcount
Demand modeling fails when it starts from current headcount and applies a growth percentage. That method encodes last year's staffing errors into next year's plan and calls the result a forecast.
Start from business drivers instead. Pick three to five drivers that leaders already forecast: revenue by segment, transaction volume, branch count, claims processed, regulatory filings. Express each capability's demand as a ratio to a driver. Twelve underwriters per hundred thousand policies. One data engineer per four analytics products in production.
The ratios will be rough. Rough and explicit beats precise and invented, because an explicit ratio can be argued with. Then interrogate each one. If a ratio has held steady for three years while the work underneath it changed, it is measuring something other than the work.
modeling supply without a single attrition rate
Supply modeling fails when one company attrition rate is applied to every group. Attrition is not uniform. It concentrates in particular role families, tenure bands, and locations, and the average conceals exactly the places where the plan will break.
Build supply in four moves. Start with current headcount by capability rather than by cost centre. Subtract expected exits using attrition rates segmented by role family and tenure band. Add and subtract expected internal moves, which requires knowing which roles are adjacent to which. Add confirmed pipeline only, meaning offers out and requisitions with live candidates.
Most plans skip the third move entirely. That is why internal mobility and workforce planning are usually run by two teams whose numbers have never been reconciled.
the gap is a shape, not a number
The output of the model is not one gap figure. It is two curves, demand and supply, plotted by quarter across eight quarters. The shape tells you the route.
A gap that opens in quarter two and closes by quarter five is a borrow problem. A gap that opens in quarter six and widens from there is a build problem, and you have time to use. A gap that is open now and structural is a buy problem, and you are already late. The same numeric gap points to three different decisions depending only on when it arrives.
The World Economic Forum's Future of Jobs Report 2025, based on responses from more than 1,000 employers across 55 economies, gives a useful sense of scale. Employers expect 39% of key skills to change by 2030. Of every hundred workers, 59 will need training, 29 can be upskilled inside their current role, 19 will need reskilling and redeployment elsewhere in the business, and 11 are unlikely to receive the training they need. That last group is the one a plan either accounts for or abandons by omission.
what makes a plan wrong
Every plan is wrong on the day it is published. The useful question is what makes it wrong faster than the review cycle can catch.
Three triggers should force a re-decision rather than a routine refresh. A business driver moving more than 15% from forecast. An automation result landing materially above or below the assumed task share. An attrition surprise in any segment carrying a build decision, because build depends on the people you assumed would stay.
Quarterly review, annual rebuild, and a named owner per capability is enough governance for most enterprises. Tooling helps once the capability inventory outgrows a spreadsheet, and connected workforce planning systems earn their place mainly by keeping supply data current rather than by producing better forecasts. The forecast was never the hard part. The decision was, and the decision is what the spreadsheet leaves out.