Strategic Workforce Planning Belongs to a Different Era

For most of the last decade, strategic workforce planning was a solution looking for a buyer. Consultants had been selling multi-year modelling exercises since the mid-2000s. They looked rigorous on paper and mostly gathered dust. By the middle of the 2010s the money had moved to skills platforms, talent marketplaces and workforce analytics, on the promise that if you could see your workforce clearly enough today, the future would take care of itself. That promise underdelivered too, but it was a newer kind of disappointment, and technology was (and still is) easier to fund.

Now SWP is coming back. We see it on conference agendas, in priority surveys and in the questions clients ask, and AI is the reason. Executives who cannot say what their workforce will look like in three years want a model, and the model on offer was built for mines and infrastructure projects, where roles held still for years. I understand the instinct. But anxiety creates demand for process. It does not validate the process being reached for.

So the question is whether SWP deserves to come back, and what should replace it. My short answer is that planning has to stop forecasting headcount and start bounding change. The output worth having is a ceiling on how far AI could credibly go in a given workflow, which I call the maximum envelope, together with an honest account of what would have to be true for the organisation to get there.

Where planning worked

I stumbled into my first major workforce planning engagement in 2006. Like all good origin stories, it began underground and in the dark: multi-year expansion modelling for large precious-metals mines, before moving above ground to state-owned infrastructure programmes.

The conditions suited the method almost perfectly. Roles were stable, certification paths were understood and the labour market could be observed. A mistake was costly and easy to trace. If we got the number of commissioning engineers wrong, the project stopped, contracts were breached and safety risk went up. The cost led so directly back to the planning failure that no serious operator treated the work as optional.

The uncertainty had a shape, too. We couldn’t predict commodity prices, demand curves, regulatory approvals or the sequencing of capital projects, but we knew roughly how far each could swing (and financial markets provided much in the way of established models). Scenario models could flex workforce demand against those futures, and they earned their keep even when the forecast turned out wrong, because they forced decisions. Which capabilities were non-negotiable? Which could be bought and which had to be grown? How much lead time did each kind of labour need?

The method earned its confidence in a particular kind of place: long horizons, stable roles, bounded uncertainty and consequences severe enough to justify sustained investment. The knowledge economy had almost none of those.

Where it broke

Around 2010 the method looked transferable. The baby-boomer retirement wave was accelerating, and the 2008 crisis had shown how little organisations knew about their own workforce costs and dependencies. If planning could spot an engineering shortage five years out for an energy major, surely it could do the same for a bank’s derivatives desk, or for a government department about to lose a large share of its senior civil servants.

It rarely transferred cleanly, and each of the three conditions broke down in turn. Role stability went first. Job descriptions in knowledge businesses changed faster than a multi-year model could be refreshed, and the outputs became spuriously precise: a skills-gap analysis for a consulting firm has nothing like the grounding of a headcount model for a port expansion, yet it was presented with the same confidence. Then the horizon shrank. A product pivot or an acquisition could make a workforce strategy stale between planning rounds, and the three-to-five-year horizon quietly became one to three years, which often felt optimistic. Finally the consequences blurred. In mining, a workforce gap stops the project. In a knowledge business it shows up as slower releases, higher attrition, expensive external hiring and thinner leadership pipelines. Real costs, but hard to pin on a planning failure.

Budget followed. Skills platforms and analytics answered different questions faster, and their return was easier to explain. Most practitioners lost faith in SWP because its predictions failed to come true, long before anyone diagnosed why.

What survived was modest and useful: a way to stress-test possible futures, force assumptions into the open and give leaders a shared language for workforce risk. We still use it for exactly that.

AI makes the old weaknesses worse

The weaknesses in traditional planning were there long before AI. AI makes them bigger. The World Economic Forum expects 39% of workers’ existing skill sets to be transformed or outdated between 2025 and 2030.¹ Roles are shifting while people are still in them, and organisations are trying to understand how people and machines perform together using models built for a world in which people did nearly all the work. Those models will still produce numbers for the largest workforce change most organisations have attempted. Leaders will trust the numbers more than they should.

The question most organisations still ask is how many roles AI can replace, and how fast. The evidence suggests leaders are acting well ahead of results. In McKinsey’s 2025 survey, 32% of respondents expected AI to reduce headcount at their organisations. A year later 14% reported reductions, two-thirds reported little or no AI-related change in employment, and 39% now expect cuts in the year ahead.² The Return on AI Institute surveyed more than 1,000 executives and found that 60% had already cut headcount in anticipation of AI gains, and 21% had made large cuts on that basis. Only 2% had made large cuts tied to AI actually in production.³ Challenger, Gray & Christmas attributed 54,836 of the 1,206,374 announced US job cuts in 2025 to AI, about 4.5%.⁴ None of these tells you how fast substitution will happen (two are executive surveys and the third counts announcements). Read together, they suggest companies are cutting on expectation well before AI delivers.

In conversation, senior HR and technology leaders have mostly moved past simple substitution. They talk about augmentation, human-AI collaboration and redesigning work. The planning models underneath haven’t moved. Most still assume that AI automates tasks, labour demand falls and headcount follows. Augmentation changes how fast and how well work gets done, and a headcount model has no field for that, so in most organisations augmentation remains a philosophy sitting on top of a framework that can’t hold it.

I’d rather start by asking what would need to be true for AI to deliver value at all, and what that means for the workforce. That leads quickly to modelling more than one thing.

Four dimensions

AI changes the economics of work in at least four ways, and they can move in different directions at the same time.

• Substitution is AI reducing the human effort needed for existing work. It is the most intuitive dimension, and it dominates planning discussions because it produces the cleanest numbers and the easiest business case. It also tends to be overstated at the planning stage.
• Expansion is AI creating demand the organisation could not serve before, products it could not build and expertise it could not scale. A company that models only substitution can end up planning for a smaller version of itself when it needs to plan for a different one. The early evidence leans this way. PwC’s 2026 Jobs Barometer found that headcount at the most AI-exposed large companies grew 52% between 2018 and 2024/25, against 36% at the least exposed, alongside stronger productivity growth.5 That is an association across larger firms grouped by sector, and it proves nothing about cause, but a pure substitution model would predict the reverse.
• Machine-work economics is the cost unit shifting from headcount to compute and control. Capability matters, and so does price. When machine work becomes cheaper than labour for a given workflow at acceptable quality, substitution accelerates. When it doesn’t, automation stays technically impressive and economically marginal. A workforce plan that models labour costs and ignores the AI cost curve is missing half the economics.
• Unintended consequences (or Secondary Effects) covers what AI does to the work around it. Some of the apparent gain leaks away into quality problems, exception handling, oversight, validation and rework, and some of it turns into new work nobody planned for. In 2024 Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer-service chats and the equivalent of 700 full-time agents.⁶ In 2025 its chief executive told Bloomberg that the cost-cutting had gone too far and that the company would recruit so customers could always reach a person.⁷ Volume handled is not value realised. I group these effects as rebound, ripple, residual and reshape, and they need an article of their own. In practice, work bounces back when the AI underperforms, checking and escalation pile up, and roles change a lot even where headcount barely moves. A plan that models only substitution is really an automation forecast.

A ceiling instead of a forecast

Traditional SWP produces point estimates: AI will reduce this function by 23%. Those carry false precision and get challenged on their assumptions the moment they reach a leadership team. Wide ranges are no better. Telling an executive that AI could affect between 10% and 60% of a function doesn’t help them decide anything. Point estimates set expectations nobody can meet, and wide ranges leave leaders with nothing to act on.

The more useful output is a defensible ceiling. The maximum envelope is the largest share of a role’s, workflow’s or function’s existing work that AI could credibly do in place of a person, judged on deployed evidence. The economics then decide how much of that is worth doing. Together they tell you how far substitution could plausibly go if the technology, the costs and the implementation all cooperated. It is not where you expect to end up, and it is not a headcount target.

That makes it, by construction, a lens on the first and third dimensions. Expansion may create more work, unintended consequences may absorb some of the gain, and roles may change more than headcount does. Treat the ceiling as the workforce plan and you have repeated the substitution mistake with better arithmetic.

In practice the ceiling changes the leadership discussion. Without one, leadership teams either dismiss AI impact estimates as too aggressive or accept them as inevitable, and both lead to poor decisions. A ceiling replaces that deadlock with a question that can actually be answered: given this is the most AI could affect, what would have to be true inside this organisation for that to become a problem, and what would have to be true for it to become an opportunity? A forecast starts a debate. That question starts a plan.

Building the ceiling

Start with work, not roles, and use the task as the unit of analysis. Break each workflow into tasks and decisions, then estimate the share of labour hours AI can take over at production quality. Weight by hours rather than by task or contact count, because the tasks AI handles best are usually the quickest ones.

“Demonstrated capability” has to mean something specific: tasks where documented deployments (not vendor claims) show the system resolving the work reliably, confirmed on a sample of your own cases. Then compare the cost of the work done by people with the cost of the AI-first route, counting inference, software, implementation, review and the human time spent on attempts the system gets wrong, as well as the model price. The crossover happens task by task, not function by function.

Take a contact centre with 100 full-time agents handling a million contacts a year, about nine minutes of handling each. At £40,000 a year per agent, fully loaded with supervision and facilities, the centre costs £4.0m, or £4.00 a contact. Gartner’s 2024 medians are $13.50 for an assisted contact and $1.84 for self-service,⁸ so this is a lean operation, which makes the case for automation harder rather than easier. The remaining figures are assumptions chosen for the example, not client data.

Illustrative contact-centre calculation

At those figures the AI-first route is about a third cheaper, so the economic ceiling and the technical one coincide. The margin is thinner than the capability suggests, though, and rests on two core assumptions. If failed attempts take twice as long as a normal contact rather than 50% longer, and platform and compute cost £1.20 a contact rather than £0.60, the AI-first route costs about £2.65 and the saving is gone. The technical ceiling would still be 30% on paper. The economic one would be close to zero. Those are the two numbers to test first, on your own contacts and your own contract.

Most substitution models stop at this point and convert 25% into headcount, but it doesn’t convert neatly. Staffing has to cover peaks, and the contacts left for agents are the harder ones, so a quarter of the handling hours frees fewer than a quarter of the agents. There is also more to ask. Where does resolution quality break down, and what exception work does that create for the agents who remain? Do customers and colleagues trust the system enough to use it, or do they route around it? And if cheaper, faster service brings in more contacts, how many of those hours come back? Answer those and you can end up with a very different workforce plan from the same technology. Realised substitution will land somewhere between zero and the ceiling, and the net effect on the workforce could land somewhere else entirely.

The ceiling is also not a number you set once, because two things keep moving. The capability boundary shifts as models improve and prices fall, which moves the ceiling. The organisational response shifts as work is redesigned, demand changes and roles are rebuilt, which changes how much of the ceiling lands. Re-estimate it every quarter rather than every planning round. If it hasn’t moved in a year, check that you are measuring the right tasks.

The envelope sets the outer boundary on substitution, and substitution is one side of the workforce equation. Most of the planning happens inside that boundary, once demand, new work and the organisation’s own behaviour start to move.

Productivity has to land somewhere

Value does not appear because a tool has been deployed. It appears when work is redesigned around the tool, people trust it enough to use it, and the organisation changes enough to capture the result. In McKinsey’s 2026 survey, about three in four AI high performers said they had redesigned workflows because of their AI use, against about one in four other respondents.² BCG found that only 5% of companies are “future-built” for AI, while 60% report minimal revenue and cost gains.⁹

Every AI workforce engagement I have been close to that went wrong had roughly the same shape. The technology worked well enough and the plan was technically coherent. What was missing was a serious reckoning with whether the organisation was ready to make it work, and three conditions decided that, more or less in this order.

Trust comes first. In practice it means people use the output without re-checking all of it, because re-checking is where the saving goes. Skills come next. In the most AI-exposed jobs they are changing about twice as fast as in the least exposed,⁵ so the learning cycle has to run faster than the work cycle. Then the operating model. When a team finishes its work in half the time, the spare capacity can disappear, move to other work, or get consumed by more review, more meetings and more output nobody asked for. Someone has to own the gain, and that is usually where it stalls. Are managers rewarded for capturing productivity or for protecting their structures? Do budgets and targets change when the work does? Can labour move between teams, or does it sit where it was hired? Can savings actually be taken out?

I have watched productivity gains appear and then vanish because they landed somewhere nobody was measuring. That is also why substitution models stay popular: their numbers are clean, even when they are wrong. Trust, skills and operating model determine how much of the envelope ever materialises, so they belong inside the workforce model as planning variables. They cannot be scheduled as change management once the modelling is done.

What to do on Monday

The envelope needs four owners, and today nobody owns the number that joins them. Technology estimates the ceiling, finance prices the crossover, HR builds the redeployment routes and the business-unit leader is accountable for delivering the gain. Before the next planning round, make three decisions.

Decisions before the next planning round

Size alone is the wrong filter. A 1,000-person workflow where AI capability is weak tells you less than a 300-person one sitting right on the frontier.

Then decide, workflow by workflow, whether to redesign, redeploy, reskill, automate, outsource or wait. Waiting is a legitimate answer where the crossover hasn’t arrived or trust isn’t there yet, as long as someone has actually decided it. Planning can start before ownership is settled, but no saving gets banked until a business leader owns it.

What replaces it

What the AI era needs is a different kind of planning, with a different output. It starts with work and value before it touches roles. It models substitution, expansion, machine-work economics and unintended consequences together, with the envelope anchoring the first. It sets a defensible ceiling instead of pretending to know where the organisation will land. And it treats trust, skills and operating model as variables with a direct effect on the result.

Twenty years ago, workforce planning worked best for me when the system was stable enough to model and the cost of error was clear. AI presents almost the opposite conditions: technology moving quickly, the economics shifting with it, and the work itself being redesigned while we try to measure it. I don’t know how much of any given ceiling will land, and neither does anyone selling a three-year headcount forecast. Planning still has a job in that setting. It should be judged less on whether it predicts headcount correctly and more on whether it shows leaders how far change could plausibly go, what would move the organisation within those limits, and which choices to make before assumptions harden into workforce decisions.

SWP is coming back - it shouldn’t come back unchanged.

Notes

1. World Economic Forum, The Future of Jobs Report 2025, January 2025. Draws on more than 1,000 employers representing more than 14 million workers. The 39% is employers’ expectation of the share of workers’ existing skill sets that will be transformed or outdated over 2025 to 2030.
2. McKinsey & Company, The state of AI in 2026: On the road to ROI, 25 August 2026 (1,719 respondents across 97 nations; about three in four high performers say they have redesigned workflows because of their AI use, against about one in four other respondents). Headcount figures: McKinsey, “AI job losses fall short of forecasts”, 3 September 2026, drawing on the McKinsey Global Surveys on the state of AI. In the 2025 survey (fielded 25 June to 29 July 2025), 32% of respondents anticipated AI-related headcount reductions; a year later 14% reported reductions, two-thirds reported little or no AI-related change in total employment, and 39% expect AI to reduce headcount over the next year. McKinsey does not publish the exact question wording, so the two years may not be strictly comparable.
3. Return on AI Institute (Thomas H. Davenport and Laks Srinivasan), Economic Maturity for Artificial Intelligence, survey of 1,006 C-suite executives across 11 countries and 32 industries, released 17 March 2026; Laks Srinivasan, “30x More Organizations Have Cut Headcount for What AI Might Do Than for What AI Has Done”, 24 April 2026. The 60% adds the post’s 39% low-to-moderate and 21% large reductions made in anticipation of AI; the 2% refers to large reductions tied to actual AI implementation, so the like-for-like comparison is 21% against 2%. The post does not define “low-to-moderate” or “large”.
4. Challenger, Gray & Christmas, Job Cut Announcement Report, December 2025, released 8 January 2026: 1,206,374 announced US job cuts in 2025, of which 54,836 were attributed to AI. The percentage is a calculation from those two figures.
5. PwC, 2026 Global AI Jobs Barometer: Two Futures for Jobs in an AI Era, 2026. Headcount growth is measured from a 2018 baseline to 2024/25 for companies grouped by the AI exposure of their sector. The sample covers larger formal firms (turnover of at least $50m) and excludes firms that exited the market, so it is not economy-wide employment data.
6. Klarna, “Klarna AI assistant handles two-thirds of customer service chats in its first month”, February 2024; OpenAI, “Klarna’s AI assistant does the work of 700 full-time agents”.
7. Charles Daly, Bloomberg News, “Klarna Slows AI-Driven Job Cuts With Call for Real People”, 8 May 2025.
8. Gartner, customer service cost per contact insight (document 5164231), 1 February 2024: median cost per contact of $13.50 for assisted channels and $1.84 for self-service; the page does not state the currency. The £40,000 loaded agent cost, the million contacts, the intent split, the 75% resolution rate, the handling time on failed attempts, and the platform, compute and review costs are assumptions for illustration.
9. Boston Consulting Group, The Widening AI Value Gap: Build for the Future 2025, September 2025.