Somewhere in the order of a third of the customer-facing workforce, gone — replaced by AI agents in a matter of weeks. Within six months, the company was quietly hiring the same roles back.
The board pack made a clean case. Replace a large slice of the customer-facing workforce with AI agents, cut the largest line item on the P&L, and post the kind of headcount-cost chart that gets applauded in a results briefing. The exercise ran over a matter of weeks. The first set of numbers afterward looked exactly like the pitch promised — cost per interaction down, headcount cost down, agents running around the clock with no overtime, no leave, no attrition.
Within a couple of quarters, four different problems started compounding at once, and the company had built a plan sturdy enough to survive maybe one of them.
The financial problem: cost that scales the wrong way.
The AI cost wasn't the fixed, predictable saving it had been modelled as. It scaled with volume, the same way a call centre's cost scales with volume — except nobody had built that assumption into the business case. A small technical team was quietly consuming enormous hours rebuilding workflows every time an edge case broke the agents' logic, which was often, because customers don't behave like the clean test scripts the system was demoed on. Escalations that used to be resolved by an experienced staff member in one conversation were now bouncing between the AI, a queue, and whichever remaining human happened to be free. None of this showed up as one alarming red number. It showed up as a dozen small ones scattered across different budget lines, easy to explain away individually, until finance added them up and the total cost of running the AI-first model was higher than the headcount it had replaced.
The human capital problem: you can't rehire what you didn't document.
The instinct, reasonably, was to rebuild the team. This is where the real cost showed up. The staff who'd been retrenched had moved on — new jobs, in some cases new industries — and rehiring at the same experience level took months, not weeks, in a labour market small enough that word travels fast about a company that let go of a third of its team to bet on AI and is now quietly hiring again. The institutional knowledge that used to live in those roles — how to actually handle the edge cases the AI kept failing on — had left with them, and had never been documented, because nobody had planned for a scenario where it would need to be rebuilt. Industry estimates on the true cost of replacing a departed employee run anywhere from half to twice their annual salary once you count recruitment, onboarding, lost productivity during ramp-up, and the output gap while the seat sits empty — a number that gets far worse when it's a large share of one function, all at once, in a tight window, under public scrutiny. The staff who stayed through the retrenchment weren't unaffected either. Survivor guilt and quiet resentment are real productivity drags, and the fastest churn in the six months after a retrenchment often comes from the people who didn't leave the first time, but decided they didn't want to be there for the second act.
The legal problem: retrenching fast doesn't mean retrenching safely.
Singapore's Tripartite Advisory on Managing Excess Manpower sets out what a responsible retrenchment looks like — objective, non-discriminatory selection criteria, proper notice, a fair retrenchment benefit generally benchmarked at two weeks to one month's salary per year of service, and mandatory notification to the Ministry of Manpower within five working days when ten or more employees are affected. None of this carries criminal penalties today. That changes materially once the Workplace Fairness Act takes effect, expected around 2027 — at that point, an employee who believes the selection criteria for who got cut and who didn't was applied unevenly has a real claim path through the Employment Claims Tribunal, with damages caps rising as high as $250,000. A retrenchment exercise built quickly, under pressure, to hit a cost target, is exactly the kind of exercise least likely to have documented, defensible, objective selection criteria sitting in a file somewhere in case anyone ever asks. Nobody in this case got sued. But the company was one disgruntled ex-employee and one plausible discrimination claim away from finding out how thin that margin actually was — and the same risk resets the moment you start a second, faster, messier hiring wave to undo the first one.
The governance problem: nobody owned the decision after the decision.
This was never really a technology failure. The technology mostly did what it was capable of doing. It was a board that approved a headcount number and let the AI strategy write itself underneath it, without asking the questions it would automatically ask about any other major structural decision. What is the total cost of ownership over eighteen months, not the first quarter's demo numbers. What is the reversibility plan if the projected savings don't hold. Who, specifically, owns daily oversight of what the agents are doing, and what's the exposure the day one of them gets something wrong in a process that matters.
The expansion problem nobody saw coming.
Here's the part that turned this from an awkward correction into a genuine crisis: the retrenchment had been sold internally, in part, as the funding mechanism for the following year's regional expansion. The savings were earmarked to open new markets. Instead, the company spent the following two quarters firefighting its existing operations, rebuilding a workforce it had just dismantled, and quietly shelving the expansion timeline it had promised investors — because you cannot scale into new markets with a core operating team that's simultaneously being rebuilt from scratch. The AI bet hadn't just cost money. It had cost the company its own growth story for the year, at exactly the moment its board was expecting to hear about progress on it.
What actually got fixed.
Brought in as the business worked through the second half of the rebuild, the mandate wasn't to reverse the AI decision — that ship had sailed, and reversing it fully would have wasted the genuine value the agents did provide for routine, low-complexity interactions. The mandate was to build the governance layer that should have existed from the start.
That meant three concrete things. First, a hybrid operating model with an explicit, written division of labour — which interaction types stayed with AI, which escalated to a human, and a documented threshold for when the system routes to a person rather than continuing to try. Second, a rehiring process built around the same objective, documented selection criteria the Tripartite framework expects, partly to fix the compliance exposure and partly because a defensible process is simply a better process — it forces you to say, on paper, what you actually need this time, rather than rehiring the last org chart by memory. Third, and most consequentially, a knowledge-capture requirement built into the AI workflows themselves: every edge case the system escalated got logged, categorised, and reviewed monthly by a small oversight group with a named owner reporting to the COO — the single point of failure the first retrenchment had created, closed by design rather than by luck.
Six months after the rebuild began, the metric that mattered most wasn't headcount cost. It was escalation resolution time, which dropped by roughly a third once the hybrid model and documentation were in place, and staff retention in the rebuilt team, which held through the first full year — the real test, given how thin the goodwill was going in.
The technology was never the risk the board failed to govern. The absence of anyone accountable for it was — and the fix was never a better AI model. It was a governance structure the first version of the plan never had.