HomeIn April 2025, a single internal email did more to crystallise the anxiety of an entire generation of content workers than a year of think-pieces had managed. Duolingo's co-founder and chief executive, Luis von Ahn, told his company it was going "AI-first." The most quoted line was the operational one: Duolingo would gradually stop using contractors to do work that AI can handle.
It was, on paper, a sober logistics decision about contract labour. It detonated like a cultural event.
Within days the green owl — a brand built on personality, mischief and the conspicuous warmth of human craft — was being roasted by the same internet that had made it famous. Creators who admired the company felt betrayed. Language enthusiasts who had spent years inside the app read "AI-first" as "humans last." And every freelance writer, translator, illustrator and editor watching from the sidelines heard the quiet part out loud: the work you do is now a line item a model can absorb.
Here is the uncomfortable truth and the reassuring one, in the same breath. Duolingo did not lay off its core team and hand the company to a chatbot. It began phasing out contractors for tasks AI could do — and even that, framed badly, cost it a wave of goodwill it is still repaying. The story everyone remembered was "AI replaced people." The story that actually matters is which work moved, why, and what got rebuilt around it.
For Singapore — a small, high-trust, content-hungry economy where most businesses run lean and outsource a great deal of their marketing, translation and creative production — that distinction is not academic. It is the whole game. It is also the recurring theme of these Insights: the companies making headlines for "replacing people with AI" are almost never doing what the headline says, and the ones quietly winning are doing something the headline never captures.
A content studio at dusk where a human editor reviews work alongside glowing AI-assisted screens, warm light against deep navy
The world-class move, decoded
To understand why Duolingo's announcement mattered, you have to understand what it actually was — and it was not "fire the writers." It was a strategic declaration about the unit economics of content.
Duolingo's business depends on producing a staggering volume of language material: exercises, sentences, audio prompts, illustrations, course content across dozens of languages, each needing localisation, review and constant refresh. Historically, a meaningful slice of that production ran on contract labour — freelancers and agencies spun up and down as courses expanded. That is the classic content operating model of the last twenty years: scale your output by scaling your headcount, mostly variable and mostly external.
Generative AI broke that equation. A capable model can now draft an exercise, propose a sentence, translate a phrase, generate a first-pass illustration or a synthetic voice line in seconds, at a marginal cost approaching zero. When von Ahn said AI would do "work that AI can handle," he was naming the genuinely automatable surface of content production — the repetitive, high-volume, pattern-based drafting that used to require an army of hands.
This is the world-class move that every content-heavy company is now making, whether they say it out loud or not: decouple the volume of content from the size of the team. For a decade, more content meant more people. Now, more content can mean the same people directing a fleet of models. That is not a marginal efficiency. It is a different physics.
Von Ahn was explicit that this was structural, not cosmetic. The follow-on details that emerged — that AI proficiency would factor into hiring, that headcount would only be granted to teams that had already automated what they could, that AI use would be built into performance — describe a company trying to rewire how work gets done, not one trimming a budget line. He was redesigning the operating model. The contractor cut was a symptom of the redesign, not the strategy itself.
And the results, by the company's own subsequent reporting, were not trivial. Through 2025 Duolingo continued to grow revenue and users at pace while leaning hard into AI-assisted production — the operating-leverage story that makes investors lean forward and makes labour markets nervous in equal measure.
But here is what made the move world-class rather than merely aggressive, and it is the part most coverage missed entirely. After the backlash, von Ahn did not double down on the soundbite. He clarified — repeatedly and publicly — that "AI-first" did not mean replacing the full-time team, that AI was a tool to make existing employees more productive, that the company was not planning mass layoffs of its core staff. By late 2025 he was on record framing AI as something that made his people more productive rather than redundant.
Duolingo didn't prove that AI replaces content people. It proved that AI replaces content tasks — and that saying it the wrong way can cost you more brand equity than the automation ever saves.
That correction is not a footnote. It is the lesson. The market handed Duolingo a real-time education in the difference between automating tasks and announcing the removal of humans — and the gap between those two things, measured in goodwill, was enormous. The technology decision was probably right. The communication decision taught a more durable lesson than the technology ever could.
It is worth dwelling on why the framing detonated so badly, because the mechanism is universal. Duolingo's entire brand equity is built on the perception of human craft and personality — the unhinged owl, the meme-literate social account, the sense that real, funny, opinionated people are behind the product. When a brand whose whole promise is "made by people who care" announces it is going "AI-first," it is not just changing a workflow. It is appearing to repudiate the thing customers loved it for. The backlash was not really about jobs in the abstract; it was about a brand seeming to betray its own value proposition. That is a specific trap, and it is one that every premium, craft-forward, personality-led business — of which Singapore has many — needs to understand before it copies the playbook.
There is a second, quieter reason the move was world-class despite the stumble: it forced an organisational discipline most companies never impose on themselves. The reported internal rule — that a team could only get more headcount once it had already automated what it could — is a genuinely powerful constraint. It flips the default. In most companies, the answer to "we are overstretched" is "hire more people," and capability silently calcifies around manual work. Duolingo's rule says: prove you have redesigned the work before you are allowed to scale the team. Whatever you think of how it was announced, that is a disciplined way to stop an organisation from carrying forward yesterday's operating model into a world where the tools have fundamentally changed.
For any Singapore business owner watching, the takeaway is not "be braver about cutting people." It is the opposite: the automation is the easy part; the redesign and the framing are where you win or lose. The owl had the technology right and the story wrong — and in a small, reputation-dense market, getting the story wrong is the more expensive mistake.
The misread: replacement versus task-automation
The single most expensive mistake in the entire AI-and-work conversation is a category error, and Duolingo's news triggered it at scale. People heard "AI replaces my job" when the accurate sentence is "AI replaces some of the tasks inside my job."
These are not the same claim, and the difference is worth real money.
A job is a bundle of tasks. A content writer does not simply "write." Across a week she researches a topic, interviews a stakeholder, sketches an angle, drafts copy, fact-checks it, fits it to a brand voice, optimises it for search, formats it for three channels, responds to feedback, and decides — using judgment no brief fully captures — what is actually good enough to ship. Generative AI is genuinely excellent at some of those tasks (drafting, formatting, repurposing, first-pass translation) and genuinely poor at others (original judgment, accountability for accuracy, taste, knowing what not to say, owning the relationship with the person who signs off the work).
When you automate the drafting but still need the judgment, you have not eliminated the role. You have changed its centre of gravity. The writer who used to spend 70% of her time generating words and 30% deciding which words were right now spends 30% directing the machine and 70% on the judgment that was always the valuable part. Same person. Higher-value work. That is not replacement. That is a role being redesigned by the tasks moving underneath it.
The reason this gets misread so consistently is that task-automation and job-replacement look identical from the outside in the short term — especially when the people whose tasks were automated happen to be contractors who can simply not be rehired. Duolingo could phase out contract work precisely because that work was structured as discrete, externalised tasks. The full-time team, whose roles were thicker bundles of judgment and ownership, stayed. The structure of the employment relationship determined what looked "replaced" — and that is a clue, not a coincidence.
The global data tells the same story at macro scale. The World Economic Forum's Future of Jobs 2025 report projects that by 2030 AI and broader trends will displace roughly 92 million roles while creating around 170 million new ones — a net gain of about 78 million. Around 86% of employers expect AI to transform their business. Read that carefully: the dominant signal is not annihilation. It is churn and recomposition — tasks dissolving and recombining into new roles faster than anyone's job description can keep up. The people who lose are the ones whose roles were never redesigned to absorb the shift. The people who win are the ones who moved up the task stack.
But honesty demands a calibration here, because the aggregate hides a brutal local truth. A net gain of 78 million globally is cold comfort to the specific freelance translator in Singapore whose volume work just vanished, or the junior copywriter at a small agency whose entire job was bucket one. The net number is real and the individual dislocation is also real — both at once. The displaced 92 million and the created 170 million are rarely the same people, in the same place, with the same skills. The role that disappears is concrete and immediate; the role that appears is often abstract, elsewhere, and needs skills the displaced worker does not yet have. This is exactly why the managed approach matters so much — and why "the market will sort it out" is not a serious answer for the person whose task bundle just got hollowed out. The macro optimism is warranted. The micro complacency is not.
There is also a subtler version of the misread that catches sophisticated operators. It is not just "AI replaces jobs" versus "AI replaces tasks" — it is which tasks, and how the leftover tasks recompose. When you strip the routine drafting out of a writer's week, you do not simply leave the judgment untouched and waiting. You change the texture of the whole role: the rhythm of the day, the skills that get exercised, the cognitive load. A writer who used to warm up on easy drafting before tackling the hard angle now faces the hard angle cold, all day, with no on-ramp. Redesign is not subtraction; it is recomposition — and a recomposed role can be more demanding, not just more valuable. Pretending otherwise is how well-intentioned automation produces burnt-out "elevated" workers who quietly resent the upgrade. The operators who get this right design the new day deliberately, rather than just deleting the old tasks and hoping the role still hangs together.
The misread is dangerous because it produces the wrong action. If you believe AI "replaces jobs," your move is to cut headcount and bank the saving. If you understand AI "replaces tasks," your move is to redesign roles to capture the freed-up capacity as higher-value output. The first produces a smaller team and a one-time cost cut. The second produces the same team doing more valuable work — and a business that compounds. Duolingo, to its credit, ended up closer to the second. The internet only ever reported the first.
Redesign, not replacement: the three-bucket model
So what does "redesign the work" actually mean when you are staring at a content function and a row of new AI tools? It means resisting the urge to ask "how many writers can we cut?" and asking the far better question:
"If the machine clears the routine production, what could our people finally do with the judgment, taste and originality that was always the point?"
That reframing produces a concrete, repeatable operating model. Take any content role — writer, translator, designer, social manager — and decompose it into tasks rather than treating the job title as atomic. Then sort every task into three buckets.
A clean three-bucket framework diagram showing tasks for machines, tasks for humans, and tasks done better together, rendered in navy and warm gold
Bucket one — what machines do better. First-draft copy, formatting and reformatting, repurposing one asset into ten, first-pass translation and localisation, generating variations for testing, summarising research, producing alt text and metadata, drafting the obvious sections of a long document. These are high-volume, pattern-based, and tolerant of a human edit. Route them to AI without sentiment. This is exactly the surface Duolingo was naming.
Bucket two — what humans do better. Deciding the angle that will actually land. Knowing the cultural nuance that makes a Singlish-inflected line work or fail. Fact-checking and standing accountable for accuracy. Holding the brand voice across a thousand pieces. Exercising taste — the thousand small judgments about what is good, not just correct. Owning the relationship with the client or stakeholder who approves the work. Spotting the thing that is technically fine and strategically wrong. AI cannot own any of these, because ownership and accountability are human by definition.
Bucket three — what they do better together. This is the bucket most businesses miss, and it is where the real leverage lives. The editor who now ships five times the volume because AI drafts and she directs. The translator who reviews and elevates machine output across ten languages instead of hand-crafting one. The strategist who tests twenty headlines in the time it used to take to write two. The human is not removed from the loop — the human is amplified by it.
Done properly, you do not end up with a smaller content team doing the same job. You end up with the same team — or a re-shaped one — doing a higher-value job. The writer becomes an editor and strategist. The production hand becomes a quality owner. The role gets harder and more valuable, not smaller. This is precisely the transition we explore in from copywriter to editor: the marketing role that survives AI — the centre of gravity moves from making to judging, and the people who make that move become more defensible, not less.
This is also the connective tissue to a broader shift we keep seeing: the startup running on two humans and fifty agents is not a fantasy about firing everyone. It is the three-bucket model taken to its logical end — a tiny human core owning judgment and accountability, directing a large fleet of machines on the production. The lean team is not a smaller version of the old team. It is a redesigned one.
Duolingo's stumble is instructive here precisely because it announced bucket one ("AI will do work AI can handle") without telling a convincing story about buckets two and three. It led with what the machines would absorb and trailed on what the humans would become. The fix — von Ahn's later "AI makes my people more productive" framing — was the three-bucket story arriving a few news cycles too late. Redesign before you reduce. And tell the redesign story before the reduction story, not after.
What this means for Singapore
Singapore should read the Duolingo episode not as foreign tech drama but as a near-term mirror. The structural conditions that made Duolingo's move possible — and its framing so combustible — are amplified here.
Start with how Singapore businesses actually buy content. The economy runs on small and medium enterprises, most of them lean, most of them outsourcing marketing, copywriting, translation, design and social production to freelancers and agencies on exactly the contract basis Duolingo was unwinding. Singapore's content supply chain is disproportionately contractor-shaped — which means the "phase out contractors for work AI can handle" logic lands here with unusual force. The agency that staffed up on junior copywriters to hit client volume is looking at the same equation Duolingo solved, and the freelancer who priced their living on volume drafting is feeling the same floor shift.
This is not a reason for fatalism. It is a reason for deliberate redesign, and Singapore is unusually well-equipped to do it deliberately — because it does not leave workforce transitions to the market alone.
Consider the trust dimension, which in a small market is everything. Singaporeans — and Singaporean businesses buying services — are reputation-sensitive in a way large markets are not. Word travels. A brand that is seen to slap "AI-first" on its door and quietly hollow out its creative team pays a reputational tax that a brand in a giant anonymous market might shrug off. Duolingo's backlash would, if anything, be sharper in Singapore's dense, interconnected, screenshot-everything business community. The framing risk von Ahn ran into is not a Silicon Valley quirk; it is a live constraint for any local business owner planning the same move.
Then there is the regulated layer. In financial services — DBS, OCBC, UOB and the wider ecosystem — content is not just marketing; it is disclosure, advice, customer communication, all of it accountable. The Monetary Authority of Singapore's FEAT principles (Fairness, Ethics, Accountability, Transparency) for AI in finance mean that content touching customers cannot simply be model-generated and shipped. A human must own accuracy and accountability. That is bucket two, codified into a regulatory expectation. For any Singapore business in or adjacent to finance, the three-bucket model is not a nice framework — it is the only compliant way to deploy AI on content at all.
A Marina Bay creative and marketing team in a bright modern office, human strategists collaborating with AI tools on screens, warm gold accents against deep blue
And consider the redesign gap that Microsoft's 2026 Work Trend Index named directly: across the economy, AI-driven productivity gains are outpacing organisational redesign. Companies are buying the tools faster than they are rebuilding the work around them. In Singapore's SME-heavy economy, that gap is the single biggest risk and the single biggest opportunity. The businesses that merely buy AI content tools will get a flood of mediocre output, a thinner team, and a brand that slowly loses its voice. The businesses that redesign the work — that move their people up the task stack and keep humans on judgment — will out-produce and out-trust their competitors with the same headcount.
The honest version of this is not "AI will create more jobs than it destroys in Singapore, relax." That is a national-aggregate comfort that means nothing to the specific freelance translator whose volume work just evaporated. At the level of the individual worker and the individual SME, the transition is real and it is happening now. The question Singapore gets to answer — better than almost anywhere, because of how it is built — is whether that transition is managed or brutal.
The Singapore enablers
Here is where Singapore's deliberate, institutional approach becomes a genuine competitive advantage — and where the contrast with the Duolingo discourse is sharpest. The Valley narrative is binary: automate or be automated, win or lose, founders and shareholders on one side, displaced workers on the other. Singapore's model is tripartite by design — government, employers and unions building the bridge together — and that is not soft idealism. It is hard infrastructure for exactly this moment.
Start with the institutions built for precisely this transition. Workforce Singapore (WSG) and NTUC's e2i run job-redesign and Career Conversion Programmes that do the unglamorous, essential work of moving a person from a role that AI has hollowed out into one it has not. These are not retraining-in-the-abstract schemes; they fund and structure the redesign of actual jobs — the copywriter into a content strategist, the production hand into an AI-assisted editor, the translator into a localisation lead who directs machine output. The three-bucket model has a national funding mechanism behind it. A Singapore SME does not have to absorb the full cost of redesigning its content function alone, the way Duolingo did internally and a US freelancer does not at all.
Then there is SkillsFuture, which puts reskilling capacity in the hands of the individual worker, not just the employer. For the content professional watching the Duolingo news with dread, this matters enormously: the floor is not "lose your role and fend for yourself." It is "your role is changing, and there is a national system whose entire purpose is to help you change with it." That is the difference between a workforce that resists AI and one that absorbs it — and it is why Singapore can run this transition with less of the scorched-earth resentment the Duolingo episode generated.
The tripartite trust layer does something subtler and more valuable still. Because unions, employers and government are in the room together, a Singapore business that frames its AI move as "redesign and reskill" rather than "cut" is not just being nice — it is moving with the national grain, qualifying for support, and protecting the goodwill of its people and the public. Duolingo's "AI-first" email created an adversarial frame: the company versus the humans. Singapore's institutions make the cooperative frame the default and the funded one. The smart operator here does not have to choose between efficiency and goodwill; the system is built to let you have both, if you do the redesign honestly.
This is, frankly, the strategic gift that Singapore's much-discussed "managed" approach hands to a business owner. In a market that prized pure speed-to-cut, the Duolingo move would be the obvious play and the brand damage just a cost of doing business. Here, the trust environment and the institutional support change the optimal strategy itself — they make redesign cheaper, faster and less risky than raw replacement. A business that ignores that and copies the crude version of the Duolingo playbook is leaving real, subsidised advantage on the table while inviting a reputational tax it did not need to pay.
For Singapore SMEs that do not have an internal team to run this, the enabling layer extends to specialist partners. Getting the redesign right — the workflow mapping, the brand-voice guardrails, the CX and process changes so the new content engine actually produces better work, not just faster slop — is a craft in itself (Freemansland Creatives), as is standing up the AI implementation safely and in line with FEAT-grade governance (Freemansland). The point is that no Singapore business has to do this alone or do it badly. The enablers — public and private — are unusually thick here.
The operator's playbook: five moves
Enough principle. If you run a Singapore business — an agency, an SME marketing team, an in-house content function — the Duolingo saga compresses into five moves you can start this quarter. They are deliberately ordered: do them out of sequence and you get Duolingo's backlash without Duolingo's growth.
1. Map tasks, not roles — before you touch a single headcount decision. Pull a month of your actual content output and decompose it. What share is routine drafting, formatting, repurposing, first-pass translation? What share is genuine judgment — strategy, fact-checking, brand voice, client ownership? You will almost always find that 40–70% of the volume is routine production, but a much smaller share of the value. That ratio is the entire strategy. The routine volume is your automation surface. The judgment is your human moat. You cannot redesign what you have not decomposed.
2. Automate the production layer — visibly to your team, invisibly to your audience. Deploy AI on bucket-one tasks: drafting, variation, reformatting, repurposing one asset into ten. But tell your people plainly what Duolingo learned to say only after the backlash: this clears the grind so you can own the work that actually matters. Adoption dies the moment your team believes AI is there to fire them. And keep it invisible to your audience — the reader should feel more craft, not less. AI that produces obvious slop is not an efficiency; it is a brand liability with a fast turnaround time.
3. Redesign the human role upward — and pay for the new shape. Rewrite the job around judgment, taste, accountability and direction-of-machines. The writer becomes an editor-strategist; the producer becomes a quality and brand owner. Make explicit that the role got harder and more valuable, not smaller — and let title and pay reflect that. This is the move that turns a feared transition into a wanted one, and it is the one most businesses skip because it costs more than a layoff in the short term and far less in the long one.
4. Keep humans firmly accountable where trust and regulation live. Anything customer-facing, anything with factual or financial weight, anything regulated — AI drafts, a human decides and owns it. In finance this is FEAT compliance; everywhere else it is just good business. The human-in-the-loop is not a brake on the machine; it is the thing that makes the machine's output safe to ship under your name. This is the bucket-two discipline, and in Singapore's trust-dense market it is also your differentiator: be the brand whose content a customer — and an AI search engine — can actually rely on.
5. Reskill into the redesign, and use the national rails to do it. Move your displaced production capacity into the redesigned higher-value roles, supported by SkillsFuture, WSG and e2i job-redesign and Career Conversion Programmes. Do not release the institutional knowledge you spent years building — your people understand your customers, your voice and your market in ways no model does. Reskill them into the roles the redesign created. The reclaimed hours then become growth, quality and retention — not a one-time cost cut that quietly erodes your capability.
One discipline runs through all five: this is a content question, but it is increasingly a visibility question too. As search shifts toward AI answer engines, the content that wins is original, trustworthy, well-structured and accountable — exactly the human layer the three-bucket model protects. A business that automates volume but loses its human judgment does not just lose brand voice; it loses its rankings and its AI citations. The redesign and the discoverability are the same project. We unpack the parallel version of this for service teams in redesigning customer service in Singapore — same logic, different function: automate the routine, elevate the human, never outsource the trust.
The investor close: where the leverage actually shows up
For anyone allocating capital — or running a business like an owner who one day wants to sell it — the Duolingo story is a lesson in where AI value lands on the income statement, and it is not where the headlines point.
The naïve reading is "Duolingo cut contractor costs, margins up, done." The operator's reading is sharper: AI created enormous operating leverage in content production — the ability to grow output without growing the team proportionally — and that leverage showed up, or leaked away, depending entirely on whether the organisation was redesigned to capture it. Duolingo's continued revenue and user growth through 2025 alongside aggressive AI adoption is the leverage captured. A clumsier company that bought the same tools, cut the same contractors, and let its content quality drift would show the same lower costs and a slow, invisible erosion of the brand that drives its revenue. Same technology. Opposite result.
The metric that should actually move is revenue per employee. This is the cleanest read on whether AI is being used to redesign or merely to cut. A business that simply removes people and banks the saving gets a one-time bump and a smaller, more fragile organisation. A business that redesigns — keeps its people, moves them up the task stack, and lets AI multiply their output — shows revenue per employee climbing and continuing to climb, because the same humans now direct ever-larger volumes of valuable work. The first is a cost story that ends. The second is a leverage story that compounds. Investors should learn to tell them apart, because on a balance sheet at the moment of the cut they look identical.
The question to ask any company touting an "AI-first" transformation is brutally simple: did your revenue per employee rise because you cut the denominator, or because you grew the numerator? One is a diet. The other is a redesign.
For the Singapore context this is doubly true, because the cost of getting it wrong is higher. In a trust-dense, reputation-sensitive market, the brand erosion from crude automation is not a slow leak — it is a fast one, and it shows up in revenue, not just in NPS. The Singapore business that redesigns around AI will show rising content quality, flat-to-lower production cost, climbing revenue per employee, and a brand that customers and AI search engines still trust. The one that copies Duolingo's first move without its second will show a thinner team, a bigger software bill, and a quiet drift in the one thing that was never on the dashboard: whether the work was still any good.
Duolingo went "AI-first," cut contractors, took a wave of backlash, and then spent months re-learning — in public — what every serious content business eventually learns. You do not win by removing the humans. You win by redesigning the work so the humans and the machines each do what they are best at — and by telling that story before you make the cut, not after.
The Singapore business that internalises that won't go viral for declaring itself "AI-first." It'll do something more valuable: keep its voice, keep its people's judgment, keep its customers' trust — and quietly build the operating leverage that the loud version was chasing all along.
Redesign before you reduce. The owl learned it the hard way. You don't have to.