HomePicture a company. It sells real software to real customers. It ships features weekly, answers support tickets in minutes, runs paid campaigns across three markets, closes its books on time, and posts revenue that would have required forty people a decade ago. Now walk through its office. You will find two desks.
Two humans. Fifty agents.
The fifty are not interns. They are software — orchestrated AI agents that write code, triage support, draft contracts, reconcile invoices, qualify leads, monitor systems, and brief the two humans every morning. The humans do not do the work. They direct it. They set the goals, hold the judgment, own the accountability, and decide what the company will not do. The agents do almost everything else.
This is a thought experiment, not a press release. No single company runs exactly this way today, and anyone who tells you otherwise is selling something. But every piece of it already exists in fragments — a solo founder shipping with coding agents here, a three-person team running support and marketing through orchestration there. The fragments are assembling. The two-human, fifty-agent startup is not a prediction. It is an extrapolation of things already happening, asked the only question that matters: what would it take to run a real business this way — and what does it mean for a place like Singapore?
A minimalist modern office with two human desks and a wall of glowing screens representing many AI agents working in parallel
The lazy answer is that it means mass unemployment. The lazy answer is wrong, and expensively so. The real answer is more interesting, more demanding, and far more useful to anyone actually running a company in Singapore right now.
The world-class move: from doing the work to directing it
Start with what is genuinely new, because most commentary gets this wrong by treating AI as a faster typist. It is not. The world-class shift underneath the two-human startup is a change in the unit of work itself. For two hundred years, the unit of work was a task performed by a person. You hired a person to do a thing, and the company's capacity was the sum of the things its people could do in a day. Headcount was capacity. That equation held from the textile mill to the modern open-plan office.
Agentic AI breaks that equation. The unit of work becomes a task delegated to a system that can run thousands of times in parallel, around the clock, in any language, at near-zero marginal cost. Capacity decouples from headcount. A two-person company can now command the task throughput of a forty-person one — not because the two people are superhuman, but because the work has been re-housed.
This is why "AI productivity tool" undersells it. A productivity tool makes a person faster at their existing task. An agent does the task, then the next, then a thousand more, and reports back. The person's job moves up a level — from doing the task to specifying it, supervising it, and deciding what to do with the result. That is not a 10% efficiency gain. It is a different kind of leverage, and it compounds.
Consider what the two humans in our thought experiment actually spend their days on. Not writing the code — an agent does that, against a spec. Not answering the ticket — an agent drafts the reply and escalates the one in fifty that needs a human. Not building the campaign — an agent assembles it from the brief and the brand guide. The humans spend their time on the things that do not delegate cleanly: deciding which customers to pursue, judging whether a risky feature is worth shipping, holding the relationship with the enterprise buyer who wants to look a founder in the eye, and owning the call when something goes wrong. Judgment, taste, accountability, relationships. The residue that does not compress into a prompt.
The winners in this shift are not the people who type fastest. They are the people who learn to direct a fleet of agents the way a conductor directs an orchestra — choosing the piece, setting the tempo, and knowing instantly when a section is off.
There is hard evidence that this shift is structural, not hype. The World Economic Forum's Future of Jobs report for 2025 projects roughly 170 million new roles created and about 92 million displaced globally by 2030 — a net gain of around 78 million, with 86% of employers expecting AI to transform their business in that window. Read those numbers carefully. They do not describe a jobs apocalypse. They describe a churn: enormous destruction and even more enormous creation, happening at once, faster than most organisations can absorb. The risk is not that the work disappears. The risk is that the work moves, and the organisation does not move with it.
Microsoft's 2026 Work Trend Index put a name to the problem the two-human startup quietly solves: the redesign gap. Productivity gains from AI, the research argues, are outpacing the rate at which organisations actually redesign how work flows through them. Companies are buying the capability and bolting it onto the old org chart. The technology sprints; the operating model walks. The gap between them is where the value leaks out — and where, for a few years, the advantage will belong to whoever closes it fastest.
The two-human startup is, in essence, a company built with no redesign gap at all — because it was designed around agents from the first day. It never had a forty-person org chart to retrofit. That is its unfair advantage, and it is precisely the advantage that incumbents, with their headcount and their habits, will struggle to copy. Which is exactly why the incumbents that do copy it will be formidable.
The misread: replacement versus task-automation
Now the dangerous part, because this is where most leaders — and most headlines — take a wrong turn that costs real money.
The misread goes like this: if two humans plus fifty agents can do the work of forty people, then thirty-eight people are now redundant, so cut them, bank the savings, and move on. It is intuitive. It is also a category error, and the companies that act on it tend to discover the mistake the hard way, on the income statement, two quarters later.
Here is the error. AI does not replace people. It replaces tasks. A job is a bundle of tasks, and AI is wildly uneven across that bundle. It is extraordinary at the routine, structured, high-volume tasks — retrieval, drafting, classification, reconciliation, first-pass code. It is mediocre-to-dangerous at the tasks that actually define a senior role — judgment under ambiguity, reading a room, owning a consequential decision, knowing which rule to break and when.
When you replace a person, you delete the whole bundle — including the judgment tasks the AI cannot do — and you lose the institutional knowledge, the customer relationships, and the trust that lived in that person's head. When you automate the tasks, you keep the person and hand them back the hours the routine work was eating. One of these moves makes the company smaller. The other makes it more capable. They are not the same move, and they do not produce the same result.
The most instructive cautionary tales of the past two years were companies that announced sweeping AI-driven replacement, made the headline, and then quietly rehired humans for the work the model handled badly. The pattern repeats because the logic is seductive and the feedback is delayed. The savings show up immediately and visibly. The damage — the churned customer, the missed fraud, the eroded quality, the institutional memory that walked out the door — shows up slowly and is hard to attribute. By the time the dashboard catches it, the cut looks like a win and the decay looks like bad luck.
Make it concrete. Imagine a Singapore insurer that runs the replacement play on its claims team — agents now handle the whole pipeline, headcount drops, the quarter looks great. Three months later, a cluster of complex claims that a seasoned assessor would have flagged as suspicious sails straight through, because the model was trained to optimise throughput, not to feel the prickle of "this doesn't add up." Now layer in the customers who were quietly mishandled at their worst moment — a hospitalisation, a death in the family — and switched insurers, telling everyone they know why. None of that appears in the cost line. All of it appears, eventually, in the loss ratio and the brand. The replacement looked like a 20% cost saving. It was, in truth, a deferred and much larger bill. The task-automation reading would have kept the assessors, handed them the routine intake via agents, and pointed their now-freed judgment at exactly the cases that blew up. Same tools, opposite outcome, and the only variable was whether leadership thought in roles or in tasks.
The two-human startup, read correctly, is not a story about replacement at all. Nobody was fired to create it. It is a company that was born with its tasks already routed to the right executor — humans where judgment lives, agents where volume lives. The forty people it "replaces" were never employed. There is no severance, no lost knowledge, no morale crater. That is the whole point, and it is the part the replacement narrative cannot see: the startup wins not by removing humans but by never building the bloated structure in the first place, and by reserving its two human seats for the highest-leverage work in the company.
For an incumbent, the equivalent move is not "cut to match the startup." It is "redesign so your existing people do what the startup's two humans do — direct, judge, own — while agents absorb the volume." The incumbent that grasps this keeps its people and turns them into a fleet of directors. The incumbent that misreads it cuts its people, keeps the org chart, and wonders why the savings evaporated.
Redesign, not replacement: the three-bucket model
So how do you actually do it? The answer is almost insultingly simple to state and genuinely hard to execute: stop thinking in roles and start thinking in tasks. Take any function — support, finance, marketing, engineering, sales — and decompose it into the tasks people actually perform in a week. Then sort every task into one of three buckets.
A clean three-bucket framework diagram showing tasks for machines, tasks for humans, and tasks done better together
Bucket one — what machines do better. These are the high-volume, structured, rules-tolerant tasks. Status lookups and FAQ responses. Drafting a first version of almost anything. Classifying and routing. Reconciling transactions. Summarising a long case history into three lines. Writing boilerplate code against a clear spec. Monitoring a system and flagging anomalies. Translating across languages. These tasks share a signature: the right answer is largely knowable from the inputs, the volume is high, and the cost of a small error is low or easily caught. Route them to agents and do not feel sentimental about it. A human doing these tasks all day is a human being wasted.
Bucket two — what humans do better. These are the judgment-and-relationship tasks. De-escalating a furious customer. Deciding whether to grant a goodwill exception that breaks policy for a good reason. Reading whether the enterprise prospect is actually going to buy or just enjoys the meetings. Owning a complaint end to end. Spotting the fraud that "looks fine" to a model trained on yesterday's patterns. Holding the line on a decision when the data is ambiguous and someone has to be accountable. These tasks share the opposite signature: the right answer is not derivable from the inputs alone, the stakes are high, and the cost of a confident-but-wrong machine is severe. Keep these firmly with humans, and pay and promote accordingly, because this is now the scarce, valuable work.
Bucket three — what they do better together. This is the bucket everyone forgets, and it is where most of the actual gain lives. It is the human who now handles three times the complex caseload because an agent prepped the context, pulled the history, drafted the options, and cleared everything routine before the human picked up. It is the developer who ships faster because agents handle scaffolding and tests while she designs the architecture. It is the marketer who runs five campaigns instead of one because agents produce the variants and she owns the strategy and the taste. The pairing is not human-or-machine. It is human-plus-machine beating either alone. Get this bucket right and you do not end up with a smaller team doing the same job. You end up with the same team doing a dramatically higher-value job.
Run this exercise on any role and a striking thing happens: very few whole jobs land cleanly in bucket one. Most jobs are 50–70% bucket-one tasks wrapped around a hard core of bucket-two judgment. That structure is exactly why "replace the role" is the wrong instrument and "redesign the role" is the right one. You are not deleting the job. You are dissolving the routine 60% into agents and rebuilding the human's week around the valuable 40% — plus the new bucket-three work of directing the agents that now do the rest.
This is also why the house rule is redesign before you reduce, in that order, deliberately. Reduce first and you cut blindly, severing bucket-two judgment along with bucket-one volume. Redesign first and you discover what each person is actually for, route the routine away, and then — only then — find out whether you genuinely have surplus capacity or whether you have just freed your best people to do your most valuable work. Most companies that redesign honestly find the latter.
What this means for Singapore
Singapore is, in a quiet way, the most interesting place in the world to run this experiment — and not for the reasons a Silicon Valley accelerationist would assume.
The naive view is that Singapore, being small, expensive and labour-constrained, should embrace the two-human startup fastest and hardest: fewer workers needed, problem solved. That view misreads both the country and the moment. Singapore's edge here is not that it can shed labour cheaply. It is that it has spent two decades building the institutional machinery to redesign labour deliberately — which is precisely the capability the AI transition rewards.
Start with the structural facts. Singapore is a high-cost, high-trust, talent-scarce economy with no hinterland of cheap labour to fall back on. That makes operating leverage — output per person — not a nice-to-have but an existential national project. A Singapore SME that learns to run lean with agents is not chasing a fad; it is responding rationally to the highest labour costs and tightest labour market in the region. The two-human startup is, for Singapore, less a threat than a template for survival in sectors where finding and affording the fortieth employee was never realistic anyway.
But — and this is the part that matters — Singapore also cannot afford the social cost of the replacement misread. In a small, dense, reputation-sensitive society with a strong social compact, a wave of "AI cut my job" stories is not just painful; it is destabilising. So the country's instinct, baked into its institutions, is to push every employer toward redesign and reskill rather than cut. That instinct is not soft sentiment. It is, as it happens, also the economically correct move — the one that captures the value instead of leaking it.
Look at how this plays out in the sectors that matter most. Take the banks — DBS, OCBC, UOB — institutions where trust is the entire product and the regulator is in the room. These are exactly the businesses where the two-human logic is most tempting on volume and most dangerous on judgment. A bank's contact centre is full of bucket-one tasks: balance checks, card activations, address changes. It is also full of bucket-two landmines: the distressed customer, the disputed transaction, the fraud pattern, the vulnerable client. A Singapore bank that automates the routine and redesigns its people upward into the judgment work moves with both the economics and the regulator. A bank that simply cuts to chase a headline discovers, the Klarna way, that the last 30% of service is where the relationships — and the deposits — actually live.
The same logic ripples through Singapore's service economy. Customer service is being redesigned rather than deleted, with agents owning tier-one and humans owning the hard and emotional cases — a shift we have written about in detail in our breakdown of customer-service redesign in Singapore. Marketing teams are watching the copywriter's job evolve into an editor-and-director's job, as agents draft and humans judge — explored in how the copywriter becomes the editor. Even the "AI-first" mandates that made headlines, where companies told staff to justify any hire that an agent could do, land differently in Singapore — covered in our look at the AI-first contractor model. Across all of them the pattern holds: the function is redesigned, not erased, and the humans who thrive are the ones who moved up a level into directing the work.
A Singapore Marina Bay skyline at dusk with a modern team blending human professionals and AI interfaces in a calm office
There is a deeper point here about national strategy. The two-human startup will be built somewhere. The only question is whether Singapore's existing companies redesign fast enough to compete with it, or whether they protect the old org chart until a leaner competitor — local or foreign — eats their margin. Singapore's tripartite model gives it a genuine, and frankly unusual, ability to coordinate that redesign at scale rather than leaving every firm to discover it alone and late. That coordination is the enabler — and it deserves its own section.
It is worth being precise about who is exposed and who is insulated, because the headline "AI is coming for jobs" is too blunt to act on. Inside any Singapore firm, the roles most exposed are the ones that are almost entirely bucket-one — the data-entry clerk, the first-line script reader, the junior who spends the day copying numbers between systems. The roles most insulated are the ones that are almost entirely bucket-two — the relationship manager, the senior underwriter, the founder closing the enterprise deal. The genuinely interesting middle is the vast majority: roles that are mixed, where the routine 60% is about to evaporate and the question is whether the organisation rebuilds the person around the valuable 40% or simply declares the whole job redundant. That middle is where Singapore's redesign infrastructure does its work, and where the difference between a thriving workforce and a resentful one will actually be decided. This is also why we keep returning to the theme across our Insights coverage: the same shift looks like opportunity or threat depending entirely on whether a company chooses redesign or replacement.
The Singapore enablers: redesign as national infrastructure
Here is what most countries lack and Singapore has spent twenty years building: the institutional plumbing to turn "AI is coming" from a threat into a managed transition. It is unglamorous, it is acronym-heavy, and it is the most underrated competitive asset in this entire story.
Start with the financial sector, because that is where the stakes and the scrutiny are highest. The Monetary Authority of Singapore's FEAT principles — Fairness, Ethics, Accountability, Transparency — set the expectation that AI driving consequential decisions in finance must be fair, explainable, and answerable to a human. Read through the two-bucket lens, FEAT is not a brake on AI; it is a codification of where the human must stay. It says, in regulatory language, that the bucket-two judgment tasks — the consequential, contestable, trust-bearing decisions — cannot be quietly handed to a model. That is not a constraint that slows Singapore down. It is a constraint that stops Singapore from making the expensive replacement mistake by law. The human-in-the-loop is not optional; it is the design.
Then there is the workforce machinery, and this is the part that genuinely sets Singapore apart. Workforce Singapore (WSG), e2i (the NTUC Employment and Employability Institute), and SkillsFuture run a coordinated set of programmes built precisely for the redesign-not-replace transition:
- Job Redesign initiatives fund employers to literally do the three-bucket exercise — decompose roles, route tasks, and rebuild the human job around higher-value work. The government will co-pay to help a company redesign its jobs rather than cut them. There is no clearer signal of national intent.
- Career Conversion Programmes (CCPs) reskill workers from declining roles into growing ones, with salary support during the transition. This is the institutional answer to "what happens to the person whose routine tasks moved to an agent": you do not release them, you convert them.
- SkillsFuture keeps the whole workforce continuously upskilling, so that "learn to direct agents" becomes a fundable, normal part of a Singaporean career rather than a luxury.
Stack these together and you get something rare: a country where an SME does not have to choose between adopting AI and keeping its people. It can do both, with public co-funding, inside a framework explicitly designed to redesign work rather than destroy it. The reclaimed hours can be reinvested into growth and service quality instead of being banked as a one-time headcount cut — and the state will help pay for the transition.
Singapore's quiet advantage is not cheaper AI or faster chips. It is the only kind of infrastructure that actually matters for this transition: the institutional ability to redesign work at the speed the technology demands, without breaking the social compact.
Underpinning all of it is trust, and trust is the tripartite model's real output. When government, employers and unions move together — when a worker believes that "AI redesign" means a path to a better role rather than a polite word for retrenchment — adoption accelerates instead of stalling. The single biggest blocker to AI adoption inside companies is not technical; it is the entirely rational fear among staff that the tool is there to replace them. Singapore's tripartism is, in effect, a national mechanism for defusing that fear honestly — which means Singaporean companies can adopt faster precisely because their people are less afraid. That is a moat that no amount of compute can buy. Getting the governance and integration right is exactly the kind of work our sister teams do day to day — building the systems safely at NICKTUNG and getting the AI strategy and readiness right at Freemansland.
The operator's playbook: five moves
Enough theory. If you run a company in Singapore — a bank, an SME, an agency, a clinic, a logistics firm — the two-human thought experiment compresses into five concrete moves. Run them in order.
1. Map tasks, not roles. Pull a representative month of real work — tickets, tasks, tickets, deals, documents — and tag each item: routine, complex, emotional, regulated. Do this with the people who actually do the work, not in a strategy offsite. You are looking for the honest split, and you will almost always find that 50–70% of the volume is genuinely routine. That routine slice is your automation surface — not a person, a slice of work. This single exercise prevents the replacement misread, because it forces you to see jobs as bundles of tasks rather than as headcount to be cut.
2. Automate the routine — visibly to staff, invisibly to customers. Deploy agents on the bucket-one tasks, but be ruthlessly transparent internally about why. Tell your team plainly: this clears your queue so you can own the work that actually needs you. The moment staff suspect the agent is there to fire them, adoption collapses and they quietly sabotage it — feeding it bad data, routing around it, withholding the knowledge it needs. Adoption is a trust problem before it is a technology problem. Solve the trust problem first.
3. Redesign the human role upward. This is the step everyone skips and the step that creates all the value. Rewrite the job description around bucket-two and bucket-three work: judgment, complex resolution, relationship ownership, and the new skill of directing agents. Make the role harder and more valuable, and let pay and title reflect that. A redesigned role that is smaller and lower-status is just a layoff with extra steps, and your best people will read it that way and leave. A redesigned role that is more demanding and better rewarded is a promotion, and people fight to earn it.
4. Keep the human firmly in the loop where it counts. Codify, explicitly, which decisions an agent may make alone and which require a human. Anything consequential, contestable, regulated, or reputation-bearing — a dispute, a vulnerable customer, a credit decision, a public statement — is human-decides, AI-assists. This is not just good governance and FEAT compliance; it is good business, because it puts your human judgment exactly where errors are most expensive. The two-human startup works not because the humans do less, but because the two humans are positioned at the two highest-leverage points in the whole system.
5. Reskill, don't release. Move the capacity you reclaim into the redesigned roles and into growth — better service, new lines, deeper relationships — supported by Singapore's WSG, e2i and SkillsFuture infrastructure. Tap Job Redesign funding to do step three properly and Career Conversion Programmes to move people across. The goal is not a smaller company. The goal is the same people producing far more valuable output — operating leverage you captured by redesigning, not leverage you faked by cutting.
Notice what these five moves are not. They are not "buy an AI tool and wait." Closing the redesign gap is organisational work — process, incentives, roles, trust — with technology as the enabler, not the answer. The companies that win the next five years will be the ones that treat AI adoption as a redesign problem they own, not a software purchase they delegate. That is the difference between the two-human startup and the forty-person company that bought the same agents and changed nothing.
The investor close: operating leverage is the whole story
For anyone allocating capital, strip everything above down to one number and one ratio.
The number is revenue per employee. For most of business history it was a slowly-moving figure, bounded by how much one person could do in a day. The two-human, fifty-agent startup is, at its core, a bet that this number is about to become unbounded for the companies that redesign — and a trap for the companies that do not. A business that genuinely decouples capacity from headcount can grow revenue without growing payroll in lockstep. That is operating leverage of a kind the software industry has dreamed about and that AI now extends to companies that were never "software companies" at all — agencies, service firms, distributors, clinics.
But here is the discipline, and it is the part the hype skips. Operating leverage from AI is not automatic. It is conditional on redesign. Two companies can buy the identical agents and post opposite results. The first bolts AI onto the old org chart, keeps the redesign gap wide open, shows a slightly smaller team and a bigger software bill, and watches the promised savings leak out through quality erosion and lost relationships. The second redesigns the work — routes tasks correctly, moves humans up a level, keeps judgment in the loop — and shows rising output, flat-to-lower cost-to-serve, and revenue per employee climbing quarter over quarter. Same technology. Same spend. Completely different income statement.
So the diligence question for any business in this era is no longer "are they using AI?" Everyone will be using AI; the answer is uninformative. The question is sharper: "have they redesigned the work to capture the leverage, or have they just bought the tools and left the org chart alone?" The first kind of company is a compounding machine. The second is a margin story waiting to disappoint. The tell is in the second-derivative metrics — revenue per employee trending up while headcount stays flat, cost-to-serve falling while quality holds or improves. Those are the fingerprints of redesign. Their absence, in a company loudly announcing its AI adoption, is a warning.
For Singapore specifically, the investable thesis is unusually clean. This is a market with the highest labour costs in the region, the tightest labour supply, and — uniquely — the institutional infrastructure to redesign work at scale through MAS, WSG, e2i, SkillsFuture and tripartism. That combination means Singaporean companies that internalise "redesign before you reduce" can build durable operating leverage faster, and with less social friction, than competitors almost anywhere else. The constraint that looked like a weakness — expensive, scarce labour — becomes the forcing function that drives the redesign first.
The two-human startup, in the end, is not really a story about two humans. It is a story about what the humans are for. In the company that wins, they are not the people doing the tasks. They are the people deciding which tasks matter, judging the hard calls, owning the consequences, and directing the fleet. That is the job that does not delegate — and the companies, and the countries, that figure out how to fill those two seats with their best people will quietly run rings around everyone still counting headcount.
AI did not replace the worker. It changed what the most valuable worker does. Redesign the work, and the two seats you keep are worth more than the forty you never needed.