HomeThere is a number that lives in almost every Singapore company's annual budget, and it is almost always wrong. It sits in the workforce planning spreadsheet under a column labelled headcount — a count of human beings approved, requested, or frozen as the primary unit by which capacity, cost, and capability are measured. The CFO reads it. The board approves it. Hiring managers fight over it. And for most of the last century, it was roughly the right number to watch.
It is not the right number anymore.
The problem is not that headcount is a lie — it is that headcount is the wrong unit of analysis for a workforce that now contains both humans and machines. When a team of twelve can deploy AI agents that absorb the equivalent of three or four roles' worth of routine work, what does it mean to say that team has twelve headcount? It means very little. It tells you how many humans are sitting at desks. It tells you almost nothing about how much work the team can actually do, at what quality, and at what cost. The unit of analysis is broken, and virtually every company in Singapore is still using it.
What the most forward-thinking operators are quietly building in its place is something we might call work-count — a budget and a management discipline organised around the actual units of work: tasks, outputs, decisions, cases handled, documents reviewed, customers served. Work-count asks not "how many people do we have?" but "how much work needs to get done, who or what is best placed to do each piece of it, and what does that actually cost?" It is a harder question. It is also, in the era of agentic AI, the only question that gives you an honest answer.
The stakes are not academic. The World Economic Forum's Future of Jobs 2025 report projects, on reported figures, approximately 170 million new roles and around 92 million displaced globally by 2030 — a net positive at the aggregate level that disguises enormous compositional churn beneath it. Approximately 86% of employers surveyed expected AI-driven transformation to reshape their workforce within the decade. Meanwhile, Microsoft's 2026 Work Trend Index describes what it calls a "redesign gap": productivity gains from AI tools are already outpacing the organisational redesign required to sustain them. Companies are getting more output from their existing people — and then not changing anything about how they plan, budget, hire, or manage. The productivity lands in the quarter; the organisational reckoning lands later, and it is messier.
Singapore stands at an interesting angle to all of this. We have world-class institutions, a tripartite labour model that other economies quietly envy, a government that funds reskilling at scale, and a regulated financial sector with explicit AI-governance expectations. We also have, in most of our boardrooms and finance teams, a headcount spreadsheet that has not been interrogated for what the world now demands of it. That gap — between our institutional readiness and our management practice — is the redesign that needs to happen. This piece is about how to close it.
You will find more context on related shifts across our Insights hub — including the specific story of what it means when Singapore CEOs become managers of agents rather than managers of people, and the question of whether your job description is already obsolete.
The World-Class Move: Shifting the Unit of Production
To understand why headcount became the default unit of workforce management, you have to understand what it was solving for. In a world where every unit of work required a proportionate unit of human attention — where the relationship between output and labour was roughly linear — headcount was a perfectly serviceable proxy. You needed to process ten thousand invoices a month? You needed a team of a certain size. You needed to serve five hundred customer queries a day? You needed enough agents to rotate shifts. Scale up the revenue target, scale up the headcount. The maths was close enough to true that nobody needed to interrogate it.
AI breaks the linearity. It does not break it cleanly or completely — not yet, not in every function — but it breaks it in enough places, with enough force, that the assumption underneath the headcount spreadsheet is no longer safe. A finance function deploying modern AI agents for invoice processing, reconciliation, exception flagging and routine reporting does not need to scale its headcount proportionally when transaction volumes grow. A customer-service operation using AI to handle tier-one queries — routine, repetitive, answerable at high confidence — does not need a new hire for every additional thousand monthly contacts. The relationship between work volume and human labour has become conditional: it depends which tasks are in the mix, how much of the volume is routine, and how well the AI has been trained and governed.
This is not hypothetical. It is visible in the operating metrics of the companies that moved early. Shopify's leadership famously told teams to treat AI as the first resource before requesting headcount. The Shopify AI-first approach raises direct questions for Singapore operators about whether that discipline can be imported and at what cost. Klarna reported using AI to handle a volume of customer interactions that would previously have required a staff count of considerable size — though such claims warrant scrutiny and context about service quality and edge-case handling. DBS, in Singapore, has publicly described AI reshaping roles across thousands of positions, creating new capability in some places while displacing routine work in others.
The pattern across all of these is the same. It is not the headline number that matters. It is the structural shift underneath: the unit of production is quietly changing from the person to the task, and the organisations that build their planning around the new unit will have a structural advantage over those still counting bodies.
A split-view of a modern Singapore corporate floor at golden hour — on one side, traditional rows of identical workstations; on the other, an open and fluid collaboration space where humans work alongside glowing AI interfaces, shallow depth of field, cinematic, deep navy and charcoal tones with warm amber accents
What work-count actually looks like in practice
Work-count is not a piece of software. It is a management practice — a way of building the budget, the resourcing plan, and the talent strategy around a more granular decomposition of the work to be done.
In practice, it means that before you fill a requisition, you ask a prior question: what tasks does this role actually need to perform, and which of those tasks are genuinely human tasks versus tasks that could be routed to an AI agent? If the answer is that 60% of the role's weekly hours consist of volume-driven, rules-based, or information-extraction work that an agent can carry, the right decision is not to refuse the hire — it may be to hire a more senior, more judgment-capable person at a higher salary for 40% of the original headcount cost. The work gets done. The quality improves. The cost is lower. But none of that shows up if you are reasoning from job title to headcount rather than from task to resourcing.
Work-count thinking also changes how you measure capacity. Instead of asking "are we fully staffed?" the relevant questions become: "what is the total work volume coming into this function, what fraction is the AI carrying, what fraction requires human judgment, and is the human capacity appropriately sized for the human fraction?" These are harder questions to answer than "do we have ten people or twelve?" But they are the ones that actually tell you whether your organisation is well-resourced or not.
The discipline extends to budgeting. A work-count budget has a line for human labour and a line for AI operational cost — model API fees, platform licences, the engineering and governance overhead of running agents well. These are not equivalent lines; the cost curves behave differently, and the trade-offs between them change as agents improve. But they belong in the same budget, on equal terms, as alternative forms of capacity. The companies still treating AI operational costs as an IT overhead and human labour as the only strategic resourcing decision are misreading their own cost structure.
The timing question: when does the shift become urgent?
The honest answer is that urgency is heterogeneous. For a legal services firm running high volumes of contract review, the shift is already urgent — agents are demonstrably capable of the initial pass, and firms not using them are paying a human premium for something machines handle well. For a strategic advisory firm where the product is senior judgment, the shift is slower — agents can accelerate research and structure output, but the irreducible human component is still dominant. The mistake is applying the timing question uniformly rather than function by function. The right question is not "is AI ready for our industry?" It is "which tasks in which functions in our organisation are ready for AI right now?" That question can only be answered at the task level. Not the industry level.
For Singapore firms operating in regulated environments — banking, insurance, healthcare, legal — there is also a governance timing question. The answer to "can we automate this?" is never purely technical; it is partly a question of what MAS expects, what data is involved, and whether the accountability chain is robust enough to satisfy a regulator's scrutiny. Getting the governance right adds time, but it also adds durability — the changes that stick are the ones a regulator can stand behind.
The Misread: AI Does Not Replace People, It Replaces Tasks
Before any operator can do the work-count redesign, they have to escape the most seductive misread in the current AI discourse. The misread is this: that AI is primarily a replacement technology — that capable AI is a headcount reduction waiting to happen, and that the strategic question is simply how many and how fast.
This is not just wrong. It is wrong in a way that is expensive to discover, because it produces a series of predictable failures that look right on the spreadsheet for two or three quarters before the bill arrives.
AI doesn't replace people — it replaces tasks. The winners redesign the work; they don't just cut the headcount.
The distinction is not semantic. A person is a bundle of tasks, and those tasks are not uniform. They include the repetitive, the rules-based, the information-extracting, and the document-formatting work that machines handle better than tired humans with competing priorities. They also include the ambiguous, the relationship-dependent, the judgment-heavy, and the emotionally-loaded work that machines handle badly or not at all. Replace the person and you replace both bundles. Automate the routine tasks and you free the person to do more of the bundle that actually justifies their existence in the first place. Those are very different decisions with very different outcomes.
When leaders confuse the two, they run into three failure modes with recognisable symptoms.
The first failure is cutting the wrong slab. Headcount-first reduction removes whole people to hit a target number. The problem is that most people contain both automatable and non-automatable tasks, and removing them removes both. The institutional memory, the relationship, the judgment-under-ambiguity that sat quietly in the background of their role — these are not captured in the job description and they are not on the cost dashboard, so they do not appear to have value until they are gone. Companies that cut this way discover the missing value in the form of a churned client, a compliance gap, or a process that stalls inexplicably because the only person who understood the edge case is no longer there. The savings were real. The cost was real too. The difference is timing.
The second failure is under-investing in the redesign. If the story is "AI replaces people," there is nothing to design — you subtract and move on. If the story is, correctly, "AI replaces tasks," there is substantial design work to do: map the tasks, route them appropriately, rebuild the human role around what is left, retrain the person, and re-engineer the process so the handoffs between agent and human are clean and trusted. That work is where the durable gains live. Replacement thinking skips it — and leaves the gains on the table while paying the morale cost of the cut.
The third failure is the trust collapse. Nothing degrades discretionary effort faster than the belief among your workforce that they are being measured for the chopping block. The brutal irony is that the very qualities you need to run AI well — the humans who catch the agent's errors, who own the relationship cases, who exercise judgment when the model is uncertain — are the ones who leave first when the environment feels threatening. Replacement framing drives away exactly the capability that the redesign requires.
The time-axis confusion
There is also a misread about tempo that compounds the problem. Replacement thinking treats AI adoption as an event: a switch flips, capability arrives, headcount falls, savings accrue. The reality is a curve. Agents improve gradually. The workflow around them needs to be built. The humans who work alongside them need to learn to supervise, edit, govern, and catch mistakes rather than simply execute. The organisations that hire the right people and build the workflow well get compounding returns. The organisations that fire first, on the assumption the switch has already flipped, discover they removed the capacity they needed to run the transition — and often end up rehiring, at a premium, people they paid to leave.
The leaders who get this right describe a trajectory. The ones who get it wrong describe an event. Andy Jassy's framing — that AI would reduce Amazon's corporate workforce over time — signals the former. The executives who announce headline cuts and then backtrack eighteen months later, often quietly and without press releases, illustrate the latter. Singapore operators have time to watch those experiments and learn from them. The lesson is: redesign before you reduce, redesign deliberately, and treat the redesign as the substantive work — not the prelude to it.
Redesign Before You Reduce: The Three-Bucket Model
If replacement is the wrong frame, what is the right one? The most useful tool for operators we work with is a simple decomposition that applies to any role, in any function, at any scale. Take the work, list its tasks, and assign each task to one of three buckets.
Bucket One — Machines Are Simply Better
Some tasks belong to AI, fully and without apology. These are the high-volume, rules-based, fatigue-sensitive, consistency-dependent tasks that machines do faster, cheaper, and more reliably than any human performing the same task at scale: reconciling large transaction sets, monitoring system logs around the clock, extracting structured data from documents, drafting the routine status report, processing standard applications, answering the standard question at any hour in any language. Humans are not just slower here; in many cases they are worse, because attention degrades, errors compound with volume, and the cognitive cost of switching to high-judgment work after hours of data extraction is real.
The redesign move in bucket one is decisive: route these tasks to the agent completely, without nostalgia. Every hour a skilled person spends on a bucket-one task is an hour removed from the bucket-two and bucket-three work that actually justifies their salary and their continued development. The liberation of routine is not a threat to the workforce — it is the condition for a better one.
Bucket Two — Humans Are Still Clearly Better
Other tasks remain stubbornly, irreducibly human. The distressed customer whose situation does not fit any script. The negotiation where trust and body language are the product. The judgment call that has no precedent and will be scrutinised in retrospect. The accountability moment where someone must be answerable, not just responsive. The creative leap that recombines things in a way no training set has seen. These tasks are not beyond AI's ambition — researchers are working on harder and harder versions of all of them — but they are, today and for the foreseeable planning horizon of most operators, beyond AI's reliability. The asset a company has in its experienced people is precisely their bucket-two capability: the accumulated judgment, the relational trust, the comfort with ambiguity that agents lack.
The redesign move in bucket two is to protect and concentrate. Counterintuitively, good automation should push people further into their humanity at work — higher-stakes conversations, harder decisions, more complex cases — not further into the low-value repetition that machines should have taken years ago. If AI is not making your people demonstrably more human in their working hours, you have not redesigned; you have just added a tool.
Bucket Three — Better Together
The richest bucket, and the most neglected, is the one where human and machine together outperform either alone. The agent drafts; the human edits with judgment the model lacks. The AI surfaces the anomaly in a thousand records; the analyst interprets it with business context the model never had. The model proposes options; the manager chooses with organisational and political awareness that lives nowhere in the training data. The agent monitors the queue overnight; the human reviews the escalations in the morning with fresh judgment. This is augmentation — and it is where the great majority of productivity gain from the next decade will actually live. Not in replacement. Not in pure automation. In the carefully designed interface between agent capability and human judgment.
Bucket three requires design, not just deployment. It is not enough to give people a tool and tell them to use it. The question is: at what point in the workflow does the handoff happen, in which direction, under what conditions, with what confidence threshold, and with what human-in-the-loop checkpoint before the output goes further? Designing these handoffs is the actual work of hybrid-workforce transformation. It is unglamorous. It requires iteration. It produces the compound returns.
Redesign before you reduce. The companies that map tasks first — and let the redesign tell them what the human footprint should be — end up with smaller, sharper, better-paid workforces doing demonstrably higher-value work.
The three-bucket model is not a one-time exercise. As agents improve, some tasks migrate from bucket three to bucket one. As organisations discover edge cases, some tasks migrate back toward bucket two. The firms that build the mapping as a living practice — reviewed quarterly, updated as capabilities shift — are building an organisational capability, not running a one-time cost exercise. The firms that do it once, declare victory, and move on find themselves re-running the exercise in eighteen months without the institutional memory of how the last one went.
A senior professional in a Singapore financial district office reviewing layered data dashboards alongside AI-generated analysis, warm directional light through floor-to-ceiling glass, shallow focus on hands and screen, cinematic composition, charcoal and navy tones with warm gold accents
What This Means for Singapore: The Local Stakes
Singapore has a specific relationship to the hybrid-workforce question that makes generic international analysis insufficient. It is not that the global dynamics do not apply here. They do, and with force. It is that Singapore's particular combination of market characteristics, institutional architecture, and competitive positioning changes both the risk profile and the opportunity profile in ways that operators need to understand precisely.
The small-market trust premium
Singapore is a dense, relational, reputation-sensitive market. In the United States, a company that degrades service quality through over-automation faces friction distributed across a vast population. In Singapore, word travels in days. Relationships that took years to build are protected with corresponding seriousness. This does not mean Singapore firms cannot automate — it means the bucket-two tasks carry a higher premium here than the headline automation-capability would suggest. The distressed client handled with judgment and warmth, the complex case resolved with accountability, the high-stakes decision made by a person who can be held responsible — these interactions are worth more in a small market because the alternative (an automated failure visible to the same tight network) is worth correspondingly more to avoid.
This has a direct implication for how Singapore operators should balance the three buckets. Bucket-one automation should be pursued aggressively — the efficiency case is the same here as anywhere. Bucket-two protection should be more vigilant than the international headlines suggest, not less. And bucket-three design should pay particular attention to the transitions where an automated process hands off to a human — because the quality of that handoff is visible to a Singapore customer in a way it might not be in a larger, less connected market.
The demographic compression
Singapore's workforce faces a demographic reality that makes the hybrid-workforce transition not just a competitive option but something closer to a structural necessity. With an ageing workforce, constrained foreign labour inflows, and a tight labour market in skilled roles, the country cannot grow output through headcount addition at historical rates. The alternative is productivity growth — extracting more value from the people and the capacity you have, which is precisely what a well-executed work-count model enables. This makes the case for hybrid-workforce investment in Singapore stronger than simple cost calculus suggests. The question is not just "can we save money by using AI?" but "can we sustain our service quality and output growth as the workforce tightens, without AI?" For many Singapore firms, the honest answer to the second question is no.
The regulatory architecture
For firms in MAS-regulated sectors — banking, insurance, payments, wealth management — the hybrid-workforce question has a governance dimension that cannot be separated from the operational one. MAS's FEAT principles — Fairness, Ethics, Accountability, Transparency — establish explicit expectations for how AI is deployed in consequential financial decisions. The accountability requirement is particularly salient: there must be a human in the loop for decisions where accountability matters, and the accountability chain must be traceable. In practice, this is a regulatory mandate for something very close to the three-bucket model. Bucket-two tasks — credit decisions, dispute resolution, advice on complex financial products — require human accountability that a pure-automation path cannot provide. Bucket-three augmentation requires governance documentation that an ad-hoc AI deployment cannot produce.
The firms that treat the FEAT framework as a constraint to navigate are going to do this more expensively and less durably than the firms that treat it as a design spec. The regulator has, in effect, told you what the governance-layer of a good hybrid-workforce model looks like. Building to that spec from the start is cheaper than retrofitting it under regulatory scrutiny. For sectors where MAS applies similar expectations around responsible AI deployment, the same logic holds.
The competitive landscape: who moves first
Singapore's position as a regional hub for financial services, professional services, logistics, and tech creates an interesting competitive dynamic around hybrid-workforce adoption. The international firms operating here — banks, consultancies, tech companies — are running hybrid-workforce programmes out of global centres and deploying them into Singapore operations. If local firms wait for the international players to prove the model before attempting it, they may find the gap in operating efficiency too wide to close in a competitive market. Being a fast follower works when the innovation is product. It is less reliable when the innovation is operating model — because operating-model advantages compound over time in ways that are hard to replicate quickly.
The banks and large enterprises are moving. DBS's publicly discussed plans around AI and workforce reshaping, OCBC and UOB's deployment of AI across customer service and risk functions, the broader adoption of AI tools across Singapore's professional services sector — these are not experiments. They are early-stage implementations of a hybrid model that will reshape what it means to staff a competitive financial-services function in this city. The question for Singapore's mid-market operators is whether they treat this as a large-company story or as advance notice of where competition is heading.
The Singapore Enablers: An Institutional Stack Others Envy
One of the underappreciated facts about Singapore's position in the hybrid-workforce transition is that the country has already built, over many years, a set of institutional mechanisms that other economies are now improvising. The challenge is not to build these mechanisms — they exist. The challenge is to actually use them.
Workforce Singapore and the job-redesign mandate
Workforce Singapore has been running job-redesign support long before the current AI wave made it urgent. WSG's framing — that job redesign is not a euphemism for cuts but a discipline of taking a role, analysing its tasks, and rebuilding the human job around higher-value work — predates the agentic-AI moment by years. The vocabulary is different; the logic is almost exactly the three-bucket model. WSG's job-redesign consultancy support co-funds exactly the kind of task-level mapping that is the critical first step in any serious hybrid-workforce transition. Operators who skip this step because they think they know what the redesigned role should look like almost always get it wrong — and usually err in the direction of keeping people on tasks the machine should hold, or cutting people who held tasks the organisation needed.
The practical implication: WSG's job-redesign support is not just a subsidy. It is access to a structured methodology that has been refined across many Singapore industries and that embeds the kind of disciplined task-level analysis most management teams do not have a natural process for. Use it. The co-funding is the inducement; the methodology is the value.
Career Conversion Programmes and e2i
When tasks migrate from bucket-two to bucket-one — when a task that once required a person is absorbed by an agent — the person still exists. The question is what they do next. In many international contexts, the answer is: they do not do it at a company that needs them. In Singapore, the designed answer is different. Career Conversion Programmes (CCPs), operated under WSG, support employers to reskill existing employees into new or redesigned roles, with co-funding that defrays salary and training costs during the transition period. The logic is sound: the employee who understands the legacy process, the customer relationships, and the organisational context is a more valuable input to the redesigned role than a fresh hire who has the new skill but lacks everything else. The CCP model makes the economics of keeping and retraining competitive with the economics of cutting and rehiring — and in most cases, accounting for hidden costs, the retraining path wins.
e2i — the Employment and Employability Institute, NTUC's workforce-development arm — adds ground-level support for the human side of these transitions: placement, training coordination, worker counselling, and the kind of trust-building between employer and employee that makes a redesign feel like a commitment rather than a prelude to another round of cuts. This is not a small thing. The redesign-before-you-reduce philosophy only works if workers believe it. e2i's involvement provides a third-party anchor for that belief that internal management communication alone rarely achieves.
SkillsFuture as the reskilling base layer
Underneath the specific programmes sits SkillsFuture, Singapore's national commitment to continuous reskilling and lifelong learning. The credits, the subsidised courses, the mid-career support pathways, the enterprise grants — these collectively lower the cost of the most important single input to a successful hybrid-workforce transition: people who are competent to work alongside agents. Supervising an AI — knowing when to trust it, when to override it, when to escalate, and how to catch its failures — is a skill set that does not come pre-installed in most of the current workforce. SkillsFuture's infrastructure is, among other things, a national investment in making that skill set accessible at scale.
The aggregate picture — WSG for redesign methodology, CCPs and e2i for transitions, SkillsFuture for reskilling — is a more coherent and better-funded support stack than most Singapore operators realise. The tragedy is that most companies engage with one piece in isolation, use it suboptimally, and miss the compounding effect of using them together as a designed programme. The governance and grant-advisory layer — structuring these engagements correctly, sequencing the applications, ensuring the redesign qualifies for the support it is entitled to — is exactly the kind of work FMC Collective specialises in. The schemes exist. Getting the full value from them requires someone who knows how they fit together.
Tripartism as competitive infrastructure
The deepest enabler is the hardest to replicate and the easiest to take for granted: Singapore's tripartite model of labour relations. The collaboration between government, employers, and unions — institutionalised through the NTUC framework and decades of precedent — means that major workforce transitions are not zero-sum battles between capital and labour. They are negotiated transitions with shared rules, shared institutions, and a shared interest in an outcome that does not destroy the social compact that makes Singapore work.
This matters for hybrid-workforce redesign in a specific way. The companies that attempt a major AI-driven restructuring without the social license — without workers who believe the transition is being run in good faith, without institutional partners who can certify that the employer is operating by the rules — face a friction cost that does not show up in any spreadsheet and is very hard to recover from once accumulated. The companies that engage the tripartite machinery honestly — working with e2i, plugging into the CCPs, communicating transparently about which roles are changing and how — run the same technical transition with a fraction of the social friction. That trust differential is a competitive moat. It is also why the tripartite model is not a constraint on Singapore employers — it is infrastructure that makes the harder, better path cheaper and more durable.
The machinery exists. The question is whether your organisation treats it as paperwork or as strategy.
The Operator's Playbook: Five Moves to Run Now
Strategy that stays at altitude is not strategy — it is commentary. If you run a Singapore business and the shift to hybrid-workforce planning is on your agenda, here are five concrete moves, in the order they should happen. The sequence matters as much as the moves themselves.
Move 1 — Audit the work before you touch the org chart
This is the foundational move and the one most reliably skipped. Pick one function — customer service, finance operations, HR, procurement, any function where you sense the AI opportunity is real — and decompose every role into its constituent tasks. Not titles. Not levels. Tasks. What does this person actually do each week, in what proportions, and what is the nature of each task? Then run each task through the three-bucket logic: is this machine-better, human-better, or better-together?
The output of this exercise — an honest task map, with hours and bucket assignments — is the only legitimate basis for any subsequent resourcing decision. Companies that skip to headcount decisions without this map are not making AI-driven decisions; they are making gut-feel decisions with AI as the justification. The map also doubles as the input to your WSG job-redesign support application, which means the analytical work and the grant application are the same work. The efficiency is real.
Plan for this to take one to two weeks of serious internal effort, or a shorter engagement with outside support. Do not try to do the whole company at once; pick the highest-opportunity function and prove the method before scaling it. A structured AI-readiness assessment — the kind of rapid task mapping across functions that Freemansland runs with operators at the start of a redesign engagement — compresses this timeline significantly and imports a methodology that has been refined across many implementations. The goal is a defensible map, not a perfect one. Defensible is enough to make the next decision well.
Move 2 — Automate bucket one without sentiment, protect bucket two without apology
Once the map exists, act on it. Move bucket-one tasks to agents decisively. Do not hedge, do not run pilots that never reach a decision, do not let the fact that a human used to do something create an invisible barrier against automating it. Humans are not better at invoice reconciliation than machines. They are not better at answering the standard query on the seventeenth iteration of the day. Protecting those tasks to protect the headcount is protecting the wrong thing.
At exactly the same time, draw a hard line around bucket two. Name the tasks where a human must remain accountable — not because the agent cannot produce an output, but because the accountability, the trust, and the quality under scrutiny require a person in the loop. Resource those tasks properly. In a world of AI cost pressure, there is a temptation to thin even the bucket-two headcount because it looks expensive relative to the automated alternative. That temptation is exactly the failure mode to avoid. The judgment-heavy, relationship-dependent, accountability-bearing work is the core of what your organisation actually sells. Do not understaff it to fund automation. Fund automation to free it.
The discipline here is to be operationally ruthless about routine and strategically generous about judgment in the same breath — simultaneously, not in sequence. The companies that automate first and then discover they have also hollowed out the human layer have confused the sequence. The moves happen together.
Move 3 — Design the bucket-three handoffs explicitly
Bucket three is where the real leverage lives and where the real design work is concentrated. Do not treat augmentation as a default — design it as a product. For each task you have assigned to bucket three, answer four explicit questions: at what point in the workflow does the agent produce an output and hand off to a human? Under what confidence or quality conditions does the agent pass without human review versus flag for review? Who specifically is the human in the loop, and what is the minimum they need to look at to make a sound decision? And what does escalation look like when the agent is uncertain and the human is unavailable?
These are not abstract governance questions. They are the practical engineering of the workflow. A bucket-three task without designed handoffs is not augmentation — it is a bottleneck in waiting, where the output sits in a queue because nobody agreed in advance how it should move. The handoff design is as important as the automation deployment, and it takes longer than most operators expect. Run it with the actual people who will use the workflow, not just the IT and management teams who designed it. The users will find the gaps the designers missed. Finding them in testing is far less expensive than finding them in production with a client involved.
Move 4 — Redeploy before you release, and use the co-funding
For every employee whose tasks shift materially — whether because bucket-one automation removed a large fraction of their current role, or because a redesigned bucket-three workflow requires different skills — make an explicit decision before acting. Redeploy, retrain, or release — but make the decision with full information about costs and benefits on each path, not with an assumption about which is cheaper.
The information most companies are missing when they make this decision is the true cost of the release-and-rehire path: the institutional knowledge walking out the door (often invisible until it is gone), the rehiring cost (typically 50-100% of annual salary when you add recruitment, onboarding, and productivity ramp), and the reputational signal to the remaining workforce (which raises future retention costs). Against that, the retraining path — accelerated by SkillsFuture credits, CCP co-funding, and e2i support — is often materially cheaper and demonstrably lower risk.
The cases where release is clearly right are narrower than they appear. They are the cases where the task migration is so complete that no redesigned role retains the person's core contribution, where the retraining cost would exceed the retention value, and where the process is managed with enough transparency and fairness that the remaining workforce trusts the judgement. These cases exist. They are not the majority. Make sure you are actually in one before choosing that path.
Move 5 — Govern it, measure it, and tell the truth about it
The last move is not an afterthought — it is what makes the others stick. Wrap the whole redesign in governance documentation that can survive scrutiny. For MAS-regulated firms, this means accountability mapping, model-oversight records, and FEAT-aligned governance that demonstrates human accountability for consequential decisions. For every firm, it means a clear audit trail of what decisions were made, on what basis, with what checks.
Alongside governance, measure the right things. Track output per labour-dollar, not just headcount. Track quality metrics — customer satisfaction, error rates, complaint volumes — alongside the efficiency metrics. Track the tasks migrated, the skills reskilled, the roles redesigned. The measurement set you choose sends a signal about what you actually value. If you measure only headcount and cost, your organisation will optimise for headcount and cost. If you measure output, quality, and the productivity of the human-AI pairing, you will get an organisation that optimises for that.
And tell the truth to the people inside. The redesign-before-you-reduce philosophy only has the morale advantages we described if the workforce believes it — and they will only believe it if the communication is specific, honest, and followed by action that matches the words. Tell your people which tasks are moving, what the redesigned roles look like, what support is available for the transition, and what the timeline is. Generic reassurances that "AI will create new opportunities" do not count. Specific commitments — here is the role, here is the training, here is the timeline — do.
The governance and grant-structuring side of this — ensuring the redesign is documented to qualify for co-funding, that the workforce-transition plan is robust enough to withstand MOM and WSG scrutiny, that the accountability chain satisfies the regulator's expectations — is exactly the advisory territory that FMC Collective covers for operators who want the redesign to be not only effective but defensible. Effective and defensible together is how Singapore's best operators will be remembered for getting this right, rather than running a cut that looks like a transformation in the press release and a liability in the litigation two years later.
The Investor Close: Where Operating Leverage Shows Up
Strip the discourse to its financial skeleton, because that is what ultimately matters for the firms building this and the capital assessing them. The hybrid-workforce question, at its most fundamental, is a question about operating leverage — about whether a business can grow output faster than it grows cost.
For most of business history, service companies were structurally limited in their ability to generate operating leverage. Each additional unit of output required a roughly proportional additional unit of human labour, which meant margins were capped by the linear relationship between revenue and headcount cost. You could be more efficient at the margin, but the structure was sticky. The business that generated more revenue than its competitor was usually the one with more people — and more cost. Agentic AI is the first broadly applicable tool to break this structure. It does not break it everywhere or immediately, but it breaks it in enough functions, across enough cost-structure, that the operating model of a well-redesigned company looks genuinely different from its unredesigned competitor.
The financial fingerprint shows up in revenue per employee — the simplest single metric for operating leverage in a service business. Watch it over time, relative to industry peers, and note what it correlates with. A company executing a genuine hybrid-workforce redesign should show revenue per employee rising, and it should rise with stable or improving quality metrics, not at their expense. That latter condition is the crucial qualifier. A rising revenue-per-employee line bought by degrading service quality is not operating leverage — it is a quality tax deferred, with interest accruing. Investors who cannot distinguish the two will eventually discover the difference on their balance sheet.
The valuation implication is asymmetric. A company that simply reduces headcount — bucket-one automation used to justify a cut, with no redesign underneath — gets a one-time step down in cost and a modest improvement in one period's margins. The benefit is real but one-shot: the structure reverts to linear at the new, lower headcount base. A company that genuinely redesigns — bucket-one automated, bucket-two protected, bucket-three engineered, humans redeployed into higher-value work — builds a repeating capability to keep flattening the cost curve as agents improve. Each generation of better agents creates another opportunity to absorb routine work without adding headcount. The cost structure becomes structurally more favourable over time. This is a multiple-expansion story, not just a cost story. And it compounds.
A Singapore executive in a glass-walled boardroom reviewing a rising revenue-per-employee chart on a large screen, warm afternoon light, cinematic lens compression, deep navy suit against charcoal and ivory tones, shallow depth of field on the graph
The distinction matters for how capital should evaluate Singapore businesses in the AI era. The question is not "did they announce an AI programme?" — every company will. The question is "did they redesign the work, or did they just cut the headcount and call it transformation?" The former shows up in sustained operating leverage. The latter shows up in a better quarter followed by quality problems and talent attrition that are harder to attribute but real in their impact. Analysts who can tell the difference — by looking at output quality metrics, employee retention in the redesigned functions, and the trajectory of revenue per employee rather than just the headline headcount number — will consistently out-select the field.
The shift from headcount to work-count is, ultimately, a shift in how we account for what a business is actually doing. The old accounting — so many people, at so much cost, producing so much output — was always a simplification, but it was close enough to true for a world where people were the primary production unit. The new accounting has to be more sophisticated: so much human judgment, at this cost; so many agent-hours, at that cost; so much output of this quality, at what margin. The firms that build the internal accounting first will have a structural information advantage over those still running the old model — and investors who build the external version of the same accounting will have a comparable edge in reading the field.
Singapore has the talent, the institutions, the regulatory infrastructure, and the co-funding to run this transition better than almost any economy its size. The tripartite model, the WSG job-redesign support, the CCPs, e2i, SkillsFuture — these were not built for the AI era specifically, but they fit it with unusual precision. They collectively make the redesign-before-you-reduce path not just the ethically preferable option but the economically better one: cheaper than the cut-and-rehire alternative, more durable than the headcount-freeze alternative, and more competitive than the wait-and-see alternative.
The only failure mode is the one we have already named: reaching for the headcount lever because it is visible and familiar, while the work-count redesign — harder, slower, and more valuable — sits undone. That failure is a choice. The machinery to make a better choice is already in place. The operators and investors who recognise this first, and act on it with the rigour it demands, will find themselves on the right side of a structural shift that does not reverse.
AI does not replace people. It replaces tasks. The companies that win will be the ones that redesign the work — every task mapped, every handoff designed, every person redeployed into the work only they can do. The headcount spreadsheet is not wrong. It is just not sufficient. The firms that add work-count to their vocabulary — and their budget — are the ones building the operating model that the next decade rewards.