HomeOn a developer-conference stage in 2017, Google's chief executive said a sentence that, in hindsight, reads like a starting gun. The company, he announced, was moving from mobile-first to AI-first. At the time it sounded like the usual keynote vapour — a tagline to bridge two slides. It was not. It was a mandate, and the company spent the next decade acting on it with a seriousness most of the industry only began to feel years later.
AI-first did not mean add a chatbot. It meant something colder and more structural: that every product team, every infrastructure decision, every hiring plan should now assume machine learning was the default tool rather than a special project requested by a research lab. Search, Maps, Photos, Gmail, Android, the ads engine that pays for all of it — each was quietly rebuilt so that intelligence was woven into the substrate rather than bolted onto the surface. Google reorganised its research, poured capital into custom chips to train ever-larger models, and reshaped what it hired for. The visible products changed slowly. The operating philosophy changed overnight.
A vast, dimly lit data centre corridor receding into shallow-focus darkness, rows of server racks glowing faintly with warm amber indicator lights, cinematic and editorial.
Here is why a 30-person company in Singapore — a logistics firm in Tuas, a design studio in Tiong Bahru, a wealth-advisory practice off Cecil Street — should care about a decade-old decision made by one of the largest companies on earth. Not because you will ever build a model. You will not. But because the move Google made is fractal. Strip away the scale, the chips, the research budget, and what remains is a question any business of any size can ask on a Monday morning: what would change if we assumed AI was the default tool for every task, rather than an optional experiment we get to next quarter?
This article decodes that mandate for the company that actually has to make payroll on the 28th. We will look at what the giants really did, why almost everyone misreads it, the model that separates the winners from the headcount-cutters, and — because this is Singapore — the specific local machinery that makes the redesign affordable. The thesis of this house is simple and we will not pretend otherwise: AI does not replace people. It replaces tasks. The winners redesign the work; they do not just cut the headcount. Redesign before you reduce.
The world-class move: what 'AI-first' actually was
The temptation, when a company like Google announces a strategy, is to treat it as a product roadmap. It was not. AI-first was an organisational rewiring, and the distinction is everything for a small business trying to learn from it.
Consider what "mobile-first" had meant a few years earlier. When mobile-first arrived, it did not instruct designers to make a phone app. It instructed every team to assume the phone was the primary surface — to design for the small screen first and the desktop second, to assume the user was distracted, on the move, thumb-driven. It reordered priorities. It changed defaults. It made a whole class of old assumptions illegal inside the building. A team that shipped a desktop-only feature in 2014 was not breaking a rule written down somewhere; it was simply out of step with how the company had decided to think.
AI-first did the same thing, one level deeper. It said: assume the relevant question for any product is no longer what should this software do when the user clicks but what should this software predict, generate or decide on the user's behalf. That is a profound shift in the unit of work. The old unit was the feature — a thing a programmer built and a user operated. The new unit was the judgment — a thing a model learned and a user supervised. Google bet, correctly, that whoever owned the judgment layer would own the next era.
AI-first was never a product. It was a decision about what the company would assume by default — and defaults, repeated across thousands of decisions, become destiny.
What makes this world-class rather than merely early is the discipline underneath. Three things travelled together, and they are the three things a small company must copy:
First, they changed the default question. Before AI-first, a Google product manager asked "how do we build this?" After, they asked "can a model do the hard part of this, and if so, what does the human now do instead?" The default flipped from human-does, software-assists to software-does, human-supervises. This is not a tooling change; it is a change in where you start every conversation. A 30-person firm can adopt this default tomorrow at zero cost. The next time anyone proposes hiring for a repetitive role — data entry, first-draft copywriting, invoice matching, tier-one support triage — the new house rule is that someone must first answer, in writing, why a model cannot do the routine eighty percent of it.
Second, they invested in the boring substrate. Google did not just prompt a model and call it strategy. It built the data pipelines, the custom silicon, the internal platforms that made AI cheap to deploy across hundreds of teams. For a small firm the analogue is unglamorous but real: clean your data, document your processes, and put your knowledge somewhere a model can reach it. The reason most SME AI pilots fail is not the model — the models are extraordinary and getting cheaper monthly. It is that the company's knowledge lives in one person's head, three WhatsApp groups and a shared drive nobody has tidied since 2021. The substrate is the moat, and the substrate is buildable without a billion dollars. This is precisely the unglamorous groundwork our sister practice Freemansland spends most of its time on before a single model is deployed — because the company that organises its own knowledge first is the one that compounds.
Third, they redesigned roles rather than retiring people. This is the part the headlines miss, and it is the part this whole article turns on. As AI absorbed more of the routine, Google did not simply empty the building. It changed what the people inside did. Engineers moved up the stack from writing boilerplate to specifying systems and reviewing machine-generated work. The job title stayed; the job changed. We will return to this, because it is where most companies — large and small — get it catastrophically wrong.
The deeper lesson is about altitude. Every time the giants automated a task, they did not pocket the saved hour as a cost cut and stop there. They redeployed the freed human attention to a higher-value problem the machine could not yet touch. That is the compounding loop: automate a task, lift the human, automate the next task, lift again. A company that only does the first half of that loop — automate, cut, bank the saving — gets a one-time efficiency bump and a hollowed-out team. A company that does both halves gets a flywheel. The mandate was never "do more with fewer people." It was "do far more, with the same people pointed at harder things." That sentence is the whole strategy, and it scales down to thirty people as cleanly as it scaled up to a hundred thousand.
The misread: replacement is not what happened
Here is where almost everyone — boardrooms, headlines, anxious staff, and a frankly embarrassing number of consultants — gets the story wrong. The misread goes like this: AI is coming for jobs, the giants are proving it, so the smart move is to get ahead of the curve and cut staff before your competitors do. It is wrong, it is expensive, and it is worth dismantling carefully because the error is seductive.
Start with the unit of analysis. A job is not an atom. It is a bundle. Take a bookkeeper at a 30-person firm. On any given week that person reconciles bank statements, chases overdue invoices, fields a panicked question from the founder about cash position, spots that a supplier double-billed, explains to a new hire why their claim was rejected, and quietly notices that one client always pays late before a complaint arrives. List those tasks honestly and a pattern appears: perhaps half are routine and rules-based — reconciliation, matching, categorisation — and genuinely automatable today. The other half are judgment, relationship, anomaly-detection and trust. AI can do the first half brilliantly. It cannot, in any real sense, own the second half — because owning it means being accountable for it, and accountability is something only a person or an organisation can hold.
So when a giant automates "tasks worth 4,000 roles," it has not discovered that 4,000 humans are obsolete. It has discovered that across its workforce, the routine fraction of many jobs now adds up to roughly that much labour. What it does next is a choice, not a law of physics. It can shed the headcount and keep the old role design — the lazy path. Or it can redesign the roles so the humans do the judgment half at higher intensity, and reallocate freed capacity to work that was previously starved of attention — the compounding path. The technology forces neither outcome. The technology automates tasks; management chooses whether that becomes a redesign or a layoff.
The evidence at the macro level supports the unglamorous reading. The World Economic Forum's Future of Jobs research projects something like 170 million new roles created and around 92 million displaced globally by 2030 — a net gain of roughly 78 million, even as the churn underneath is enormous. Around 86 percent of employers expect AI to transform their business this decade. Read those numbers together and the replacement narrative collapses. This is not an extinction event; it is the largest reallocation of human work in a generation. The displaced number is real and must be taken seriously — but it sits inside a larger story of creation, not deletion.
And yet companies keep making the misread, because the misread is easy to execute. Cutting headcount is a decision you can make in an afternoon and put in a press release. Redesigning work is a months-long slog of mapping tasks, retraining people, rebuilding processes and absorbing the awkward middle period where the new design is half-built. Faced with a hard slow thing and an easy fast thing that both show up in the same quarter's cost line, tired managers reach for the easy one. They mistake a finished-looking number for a finished strategy. Six months later they discover what they actually cut was not slack — it was the institutional memory, the client relationships and the judgment that the routine tasks were merely the visible surface of. The model handled the invoices. Nobody is left who notices when a client is about to churn.
That is the trap. The rest of this article is about avoiding it.
Redesign, not replacement: the three-bucket model
If "redesign before you reduce" is the slogan, the three-bucket model is the method. It is deliberately simple enough to run in a workshop with a whiteboard and your actual team, which is the only place strategy ever becomes real. The move is this: take every meaningful role in the company, and sort its tasks — not the role, the tasks — into three buckets.
Bucket one: Automate. These are the routine, rules-based, high-volume tasks where the output is judged right or wrong against a clear standard. Data entry, document classification, first-draft generation, scheduling, reconciliation, standard report assembly, tier-one FAQ responses, transcription, basic research collation. For these, AI is not a risk; it is a gift. The goal is to remove them from human hands almost entirely, with light review. Be honest and generous about what goes here — most teams underestimate this bucket because each person is emotionally attached to the parts of their job that feel safe.
A clean overhead flat-lay of three distinct empty trays on a dark walnut desk, soft directional light, one tray catching a warm highlight, minimalist and conceptual, no text.
Bucket two: Augment. These are tasks a human still owns but where AI dramatically raises the ceiling or the speed — drafting a proposal the person then sharpens, analysing a dataset the person then interprets, surfacing options the person then chooses between, reviewing a contract the person then negotiates. Here the human stays in the loop and stays accountable, but works at the top of their licence instead of the bottom. This is the largest and most valuable bucket for most companies, and it is where the real productivity gain lives — not in firing the analyst, but in making one analyst as productive as the three you could never afford to hire.
Bucket three: Reserve for humans. These are the tasks that should stay human not because AI is technically incapable but because accountability, trust, ethics, relationship or final judgment must rest with a person. Firing a client. Approving a loan that breaks a rule for a good reason. Telling a staff member their work missed. Standing in front of a regulator. Deciding the company's values when two of them conflict. The reason a firm reserves these is not nostalgia; it is governance. In any regulated or high-trust context — and most Singapore businesses touch one — the human-in-the-loop is not a courtesy, it is the control.
Sort the tasks, not the jobs. Almost every job survives the sort — it is the contents of the job that change. That single reframing is the difference between a redesign and a redundancy exercise.
Now watch what the model does to the layoff instinct. Once you have sorted tasks, you can see that a role losing its entire first bucket has not disappeared — it has been concentrated into its second and third buckets, which are the high-value parts. The bookkeeper from earlier, freed of reconciliation, becomes the person who reads the cash position, spots the anomalies, manages the supplier relationships and gives the founder a genuine financial steer. That is a better job and a more valuable one. The redesign did not delete the person; it deleted the drudgery and promoted what was left.
There is a discipline to running this well, and it is where the governance partner earns its keep. The sort must be honest (no protecting pet tasks), it must respect the regulatory line (some tasks cannot leave the human bucket regardless of capability), and it must come with a reskilling plan so the human can actually perform at the new altitude. Getting the human-in-the-loop boundaries right, documenting why each task sits where it does, and keeping the whole thing defensible to a regulator or auditor is precisely the kind of risk-and-governance work our sister practice FMC Collective exists to handle — because a redesign that cannot be explained to MAS or a customer is not a redesign, it is a liability waiting for an incident.
The three-bucket model also exposes the real cost of the lazy path. When a company skips the sort and simply cuts a role, it does not cleanly remove bucket one. It removes all three buckets at once — drudgery, judgment and trust together — and then quietly discovers it still needs buckets two and three, now performed worse, by someone already overloaded, or not at all. The saving was an illusion. The capability is gone. This is the mechanism behind every "we cut too deep and had to rehire" story, and it is entirely avoidable with a whiteboard and the honesty to sort properly.
What this means for Singapore
Singapore is, by temperament and by policy, almost perfectly built for the redesign path — and almost uniquely exposed to the temptation of the lazy one. Both things are true, and understanding why is the difference between a Singapore SME that compounds through this decade and one that hollows out.
Start with the exposure. Singapore is a high-cost, talent-constrained economy. Labour is expensive, office space is expensive, and the local talent pool — while world-class — is finite and fiercely competed for. For a 30-person firm, every hire is a serious commitment and every salary is a meaningful line. That cost pressure is exactly what makes the "automate and cut" story so tempting here. When a model can plausibly do the routine eighty percent of an expensive role, the arithmetic of simply not renewing that role is brutally attractive. Singapore's strength — its premium on productivity — is also the gravity well pulling firms toward the wrong half of the loop.
Now the structural advantage, which is considerable. Singapore runs a tripartite model — government, employers and unions working in deliberate concert — that most economies can only envy. This is not a slogan; it is a functioning machine with real money behind it. When a Singapore firm decides to redesign jobs rather than cut them, it is not doing so alone and unfunded. There is an entire apparatus designed to share the cost and de-risk the transition: Workforce Singapore and e2i run Job Redesign and Career Conversion Programmes; SkillsFuture underwrites the training; the unions sit at the table to keep the redesign fair rather than extractive. The state has, in effect, put its thumb on the scale in favour of redesigning over discarding. A firm that cuts headcount captures a one-time saving and walks away from all of this support. A firm that redesigns gets co-funded.
There is also a trust dimension that is sharper in Singapore than almost anywhere. This is a small, dense, high-reputation market where word travels and where regulators are taken seriously. In finance, the Monetary Authority of Singapore's FEAT principles — Fairness, Ethics, Accountability and Transparency — set explicit expectations for how AI and data analytics are used. FEAT is not a finance-only curiosity; it is a preview of the standard every serious Singapore business will be held to. It makes the human-in-the-loop and explainability into design constraints rather than nice-to-haves. A bank here cannot let a model make a consequential decision unsupervised and unexplained, and increasingly neither can a healthcare provider, an HR firm or anyone touching personal data under the PDPA. This sounds like friction. It is actually a moat. The discipline FEAT forces — knowing which tasks a human must own, documenting why, keeping a person accountable — is exactly the discipline that produces a good redesign instead of a reckless automation. Singapore's regulatory seriousness quietly pushes its firms toward the better path.
Put the pieces together and a clear picture emerges. The big local institutions — the DBS, OCBC and UOB tier — are already living this. When a major Singapore bank announces that AI will absorb work equivalent to thousands of contract roles over several years while it creates a smaller number of new AI roles, the lazy reading is "the cuts are coming." The accurate reading is that the most sophisticated employers in the country are doing the three-bucket sort at industrial scale, under FEAT, with tripartite support, redesigning faster than they reduce. They are showing every smaller firm in the country the template. The giants' playbook is not a foreign import here; it is being run, openly, a few MRT stops away.
The mistake a Singapore SME makes is assuming this is a big-company game. It is the opposite. A 30-person firm has structural advantages the banks would kill for: no legacy systems to unwind, no committee to convince, no quarter-by-quarter analyst pressure, and the ability to redesign an entire function in a week because the entire function is four people who all sit within shouting distance. The constraint on the small Singapore firm is never capability or even capital — the support is co-funded. The constraint is attention and discipline: the willingness to stop, map the work, and redesign on purpose rather than drift into reactive cutting. That is a leadership problem, not a technology problem, and it is the most solvable problem in this entire essay.
The Singapore enablers: the machinery that makes redesign affordable
It is worth being concrete about the support, because the gap between firms that use it and firms that do not is enormous, and the only thing separating them is usually awareness. A founder who knows this machinery exists thinks "redesign, co-funded." A founder who does not thinks "redesign, fully self-funded, can't afford it" — and reaches for the layoff. The machinery, in plain terms:
Job Redesign support. Workforce Singapore and e2i run programmes specifically designed to help employers redesign roles around new technology — exactly the three-bucket move, with funding and consultancy attached. The premise is that redesign is hard and expensive to do well, so the state co-invests to make it happen rather than leaving firms to either struggle or default to cuts. For an SME, this can mean co-funded help to map the work, identify what moves to AI, and rebuild the human roles around it.
Career Conversion Programmes. When a role's centre of gravity shifts — say a customer-service rep whose routine queries are now handled by AI and who needs to become a customer-success specialist managing relationships and supervising the AI — Career Conversion Programmes co-fund the salary and training during the conversion. This is the mechanism that makes "reskill instead of replace" financially rational rather than charitable. The freed bookkeeper, the converted support rep, the analyst moving up the stack — these transitions have a funding path. That changes the maths of every redesign decision.
SkillsFuture. The broader skills infrastructure underwrites the training itself — the courses, the credentials, the actual capability-building that lets a person operate at their new, higher altitude. A redesign without reskilling is just a reorganisation that sets people up to fail; SkillsFuture is the piece that makes the new role real.
Tripartism as a trust layer. The presence of unions and government alongside employers does something subtle and valuable: it keeps redesign honest. A redesign run purely by an employer under cost pressure can quietly curdle into a cut dressed up in nicer language. The tripartite structure exists partly to ensure that when a firm says "redesign," people actually land in better jobs rather than out the door. For the firm, this is not a constraint to resent; it is reputational cover. "We redesigned with WSG support and converted our team into higher-value roles" is a far stronger story — to staff, to clients, to the market — than "we cut headcount to fund AI."
In most economies, redesigning work around AI is a cost the firm bears alone. In Singapore, it is a cost the country has volunteered to share. The firms that win this decade are the ones who actually pick up the offer.
MAS FEAT and the governance frame. For firms in or adjacent to finance, FEAT is the standard, but its logic — fairness, ethics, accountability, transparency, human accountability for consequential decisions — is becoming the de facto bar across sectors. Treating it as a design input from day one, rather than a compliance afterthought, is what separates a redesign that survives scrutiny from one that blows up in an incident. This is the governance-and-grants intersection where firms most often need a steady hand: knowing which programmes apply, how to qualify, how to document the human-in-the-loop, and how to keep the whole thing defensible. It is unglamorous, it is decisive, and it is precisely the work that turns an AI ambition into a funded, compliant, real transformation.
The summary for the Singapore operator is blunt: you are not doing this alone, and you are not paying for it alone. The country has built the rails. The only question is whether you board.
The operator's playbook: five moves
Enough principle. Here is what a founder or managing director of a 30-person Singapore company should actually do, in order, starting this quarter. These are not aspirations; they are moves with first steps you can take this week.
1. Issue your own AI-first mandate — and mean it. Google's mandate worked because it changed the default question across the whole company, not because it funded one project. Do the same at your scale. Tell every team that from now on, the first question for any meaningful workflow is "what is the AI-default version of this?" Make it a standing rule that new headcount for a repetitive role requires a written answer to why a model cannot do the routine majority of it. This costs nothing and changes everything, because it stops the slow accretion of human drudgery before it starts. A mandate that lives only in your head is a wish; write it down, say it in a meeting, and apply it to the next real decision.
2. Run the three-bucket sort on your top five roles. Get the actual people in a room — they know their tasks better than you do — and sort every task into Automate, Augment, Reserve. Do it honestly; the goal is not to threaten anyone but to find the drudgery worth removing and the judgment worth protecting. You will be surprised how much sits in bucket one and how relieved people are to lose it. Output: a one-page task map per role and a ranked list of automation candidates. This is a half-day workshop, not a consulting engagement, and it is the single highest-leverage thing in this list.
3. Automate one painful, high-volume task end-to-end — and measure it. Resist the urge to boil the ocean. Pick one task from bucket one that everyone hates and that happens constantly — invoice matching, first-draft proposals, support triage, report assembly. Automate it properly, with a human reviewing the output until trust is earned, and measure the hours returned. One clean win does more for adoption than ten slide decks, because it converts AI from an abstraction your team fears into a tool that just gave them their Friday afternoons back. The metric matters: hours returned per week is your proof, and your mandate's fuel.
4. Redeploy the freed time upward — do not bank it as a cut. This is the move that separates the flywheel from the one-time efficiency bump, and it is the move most companies skip. When bucket-one work disappears, the freed hours must be pointed at higher-value work the machine cannot do: the client relationship that was being neglected, the analysis nobody had time for, the new service line, the quality the team always wanted to reach. If you simply pocket the saving and cut, you get a smaller, more brittle company. If you redeploy, you get a more capable one at the same cost. Decide, explicitly and in advance, where the freed attention goes — before the efficiency tempts you to just bank it.
5. Fund the redesign through the rails — and reskill on purpose. Before you spend a dollar of your own on the transition, find out which Singapore programme applies — Job Redesign support, a Career Conversion Programme, SkillsFuture funding. Map the roles that are shifting and build an actual reskilling plan so people can operate at their new altitude. This is where a governance-and-grants partner pays for itself many times over: knowing the landscape, qualifying for the support, and documenting the human-in-the-loop so the whole thing is defensible. A redesign you self-fund is expensive; a redesign the country co-funds is a competitive advantage. Pick up the offer.
A single chess knight piece in warm focus on a dark slate board, the surrounding pieces falling away into shallow-focus shadow, cinematic editorial lighting, a sense of one deliberate move.
Run those five and you have, at your scale, executed exactly what the giants did: changed the default, sorted the work, automated the routine, lifted the humans, and funded the transition. None of it requires a research lab. All of it requires the discipline to do the slow thing instead of the easy one. For more on how the other giants are running this same play, the rest of our Insights series maps each one to the Singapore SME: Microsoft pushing an AI copilot into every job, Meta flattening the org chart around AI, and Salesforce making everyone a manager of agents. The connective tissue is always the same: redesign before you reduce.
The investor close: operating leverage is the whole point
Strip away the workforce language and AI-first is, at bottom, a bet about a single number: revenue per employee. That is the number an investor, an acquirer or a serious founder should watch, because it is where the entire strategy either shows up or evaporates.
Here is the logic. A company's value is, crudely, a function of how much output it generates per unit of cost. Headcount is the largest cost line for most service businesses. For decades, growing revenue meant growing headcount roughly in proportion — more clients, more people, more or less linearly. The promise of an AI-first redesign is to break that link: to let revenue grow while headcount stays flat or grows far more slowly, because each employee, freed of drudgery and augmented by AI, simply produces more. That is operating leverage, and it is the most valuable thing a business can manufacture, because it compounds. A firm that doubles revenue per employee has not just cut cost; it has changed the slope of its own growth curve.
This is precisely why the lazy path is so destructive to enterprise value, and why an investor should be suspicious of it. Cutting headcount to fund AI looks like it improves the ratio — fewer employees, same revenue, higher revenue-per-head. But it is a one-time, non-repeatable trick that quietly removes the capability needed to grow the numerator. You improved the ratio by shrinking the denominator and damaging the numerator's future. The redesign path improves the same ratio by growing the numerator — more output per person — which is repeatable, compounding and exactly what a discerning buyer pays a premium for. Same headline metric, opposite quality. The investor's job is to tell them apart.
Microsoft's recent Work Trend research named the gap precisely: a "redesign gap," where the productivity that AI makes available is outpacing the organisational redesign needed to actually capture it. That gap is the single biggest source of value lying on the table right now. The firms closing it — sorting tasks, redesigning roles, lifting people, redeploying freed capacity — are converting raw AI capability into durable operating leverage. The firms not closing it are either ignoring AI entirely or, worse, "capturing" it through cuts that look good for two quarters and hollow out the franchise by the third.
For an investor, the tell is simple. A company cutting headcount to fund AI is shrinking to look efficient. A company redesigning work to fund growth is building leverage that compounds. Bet on the second one.
So the close for the operator and the investor is the same sentence the house has been making all along, now in financial language. The point of AI-first was never to do the same work with fewer people. It was to do far more valuable work with the people you have, and to make that the engine of growth rather than the source of a one-time saving. Revenue per employee is where that bet either lands or fails. Redesign grows it durably. Cutting fakes it temporarily. The whole strategy — Google's, the banks', and the one available to your 30-person firm in Singapore this quarter — comes down to choosing which kind of number you are actually building.
The mandate Google issued on a stage in 2017 has finally arrived for everyone else. The good news for the Singapore operator is that the playbook is known, the rails are funded, and the constraint is not money or technology but the discipline to redesign before you reduce. The giants have shown the move. The country has built the support. The only thing left is to make the decision — and to make it the right way round.