Cut jobs for AI, then hire them back?

The cheapest way to show an AI saving is to cut a role, and Gartner says that is the expensive way to do it. Its forecast is that by 2029, 30% of the employees laid off because AI replaced them will need to be rehired, often at a significantly higher cost, and that by 2027, 75% of organisations which treat AI gains as cost savings will be eclipsed by competitors that reinvest them. A second prediction makes the same point from the other side: agentic AI built by a vendor's engineers tends to be abandoned once the costs arrive. The Conference Board puts the sequence round the other way — redesign the work first, then decide where the gains go. Singapore's own numbers are quieter but consistent: only 6.2% of AI-using firms cut headcount. Our view is that the firms which change the work keep the gain, and the firms which only remove people will hire again.

We would not plan a project around a 2029 forecast that exists to start a conversation. What is worth taking seriously is the mechanism behind it. Cutting staff removes institutional knowledge and drains the pipeline a firm hires from later, and where the working population is flat or shrinking, the competition for the people you let go will be sharper than the market you cut them for. That is why Gartner expects 30% of those laid off for AI to be rehired by 2029, often at a significantly higher cost. The second half of the same release says where the money should have gone instead: by 2027, 75% of organisations that treat AI gains as cost savings will be eclipsed by competitors that reinvest those gains in innovation, modernisation and upskilling. The numbers are forecasts, and Gartner is in the business of making them, but the mechanism is worth understanding before a budget decision rather than after it.

The second Gartner release, on 29 September 2026, is about what happens after the tools arrive. Gartner predicts that most enterprises will abandon agentic AI built by a vendor's forward-deployed engineering team, trapped by soaring costs and unable to evolve the system on their own. The phrase worth keeping is the plainest one: success is measured not by whether implementation finished, but by whether the organisation can manage, optimise and scale the thing after the consultants leave. That is the same failure as the headcount cut, seen from the other direction. Money is spent to make a capability appear, and nobody funds the part where the work is redesigned around it.

We would add that the sequence is the whole argument. The Conference Board's framework for agentic AI and work redesign is written for HR leaders, and its value is the order it insists on. Start from the business outcome and use agents only where they serve it. Redesign the work task by task before deploying anything, deciding what belongs to AI alone, to people with AI, and to people alone. Build the permissions, data and quality checks the agents need. Measure what people and agents produce together rather than headcount and cost per employee. Then decide where the gains go — costs, customers, capacity or employees — and revisit that decision when the results come in. The framework was built from months of research sessions with senior HR and AI leaders, and its central point is the one we would underline: the technology cannot decide who receives the dividend. Leaders can, and mostly have not.

Singapore's evidence is the part we trust most, because it is about what firms actually did rather than what they say. Among the establishments using AI, most reported better worker productivity and more redesigned job roles than cut ones, and only 6.2% reported reduced headcount. A separate firm-level study adds a longer view: AI use was associated with higher revenue and employment, but firms saw no significant gain in productivity or profit within four years of first adopting it. Two other findings are worth saying plainly. Most firms had not begun adopting AI at all, and only a small fraction had integrated it into core processes; the most common barriers were the cost of implementation and a lack of in-house expertise. Local evidence, in other words, is mostly about redesign and rework, not replacement.

We would not plan a project around a 2029 forecast, and we would not treat a productivity gain as a saving. What is worth taking seriously is the mechanism, because Singapore's numbers already show it: the firms that benefit from AI are the ones that changed the work, and the firms that simply removed people from the payroll are the ones that will have to hire again.

Our own rule is simple and unglamorous. Buy tools that take a task out of a week, not a person out of a team; put the hours saved back into the work that was never finished; and when a client asks what AI changed, answer with what the work is now, not with the money it used to cost. The gain is real. It is just not the same thing as a saving.