HomeCustomer service fell first. Not because it mattered least, but because it was the most exposed.
For two years the debate about which jobs AI would touch ran on intuition and anxiety. Would it come for the coders, the lawyers, the radiologists, the creatives? The honest answer, visible now in hindsight, was hiding in plain sight on every company's org chart. The function AI absorbed first, fastest and at the largest scale was the one built on high-volume, repeatable, well-documented work with a clear right answer and instant feedback — and that function is customer service.
Think about why. A contact centre is, from a machine's point of view, almost a purpose-built training environment. The questions repeat. The good answers are written down. Every interaction is logged, transcribed and scored. The feedback loop — did this resolve the ticket, did the customer come back angry — is immediate and measurable. There is no other corporate function that hands an AI system this much clean, labelled, outcome-tagged data about exactly what "good" looks like. Where a model has to guess at a marketer's taste or a strategist's judgment, in support it can simply learn from a million resolved tickets what the right move usually is.
So the wave hit support before it hit anywhere else. By late 2025 it was no longer a pilot story. Airlines, telcos, banks, e-commerce platforms and fintechs across the region were routing the bulk of their first-line contacts through AI before a human ever saw them. The headlines followed the familiar grammar of fear: AI replaces customer service teams. And in Singapore — a small, service-dense, high-trust economy where a single badly handled complaint travels faster than any ad campaign — the question landed harder than almost anywhere.
But the companies winning this shift in Singapore are not the ones cutting their service teams to the bone. They are the ones redesigning the work — moving the routine to machines and re-pointing their people at the hard, human, valuable third that no model reliably owns. That distinction is the entire game. This is how they are doing it, and why it is the only move that actually fits Singapore.
The world-class move: why service was always going to fall first
To understand what is happening to customer service, you have to stop thinking about it as a department and start thinking about it as a distribution of tasks. And once you do, the reason it fell first becomes almost mechanical.
Every job is a bundle of tasks sitting somewhere on a spectrum from pure routine to pure judgment. The further a task sits toward the routine end — high frequency, bounded inputs, a knowable correct response, abundant recorded examples — the more legible it is to a machine. Customer service, more than any other corporate function, is loaded toward that end of the spectrum. A staggering share of contacts to any support operation are variations on a small number of questions: where is my order, how do I reset this, what is my balance, why was I charged, how do I cancel, can you explain this line on my bill. These are not edge cases. They are the bulk of the volume, and they are exactly the shape of work large language models were born to do.
Now layer on the three structural advantages that no other function offers a machine as cleanly.
First, the data is already labelled. A decade of support tickets is, in effect, a decade of supervised training examples: question in, resolution out, satisfaction score attached. Most functions cannot hand an AI anything close to this. A legal team's reasoning lives in heads and emails; a strategist's judgment is rarely written down at all. Support, by contrast, has been quietly building the perfect training corpus for years without realising it. The CRM was a dataset all along.
Second, the feedback is immediate and quantitative. Did the customer's issue get resolved? Did they reopen the ticket? Did the conversation escalate? Did the survey come back at one star or five? In most jobs you wait months to learn whether a decision was good. In support you learn in minutes, at scale, on every single interaction. That tight loop is what lets an AI service system actually improve in production rather than plateau — it gets corrected thousands of times a day.
Third, the questions are bounded. A support agent for a specific product answers questions about that product, within known policies, against a finite knowledge base. The universe of possible queries is large but not infinite, and it is anchored to documentation the company already owns. That boundedness is what keeps the hallucination risk manageable and the answers grounded — the model is not free-associating about the universe, it is retrieving and rephrasing a company's own truth.
Put those three together — abundant labelled data, instant measurable feedback, bounded scope — and you have described the single most favourable environment for applied AI that exists inside a normal company. It was never a question of whether service would be automated first. The economics and the data made it close to inevitable.
Customer service was not the weakest function. It was the most legible one. AI goes first where the work is most written-down, most measured and most repeated — and nothing is more written-down, measured and repeated than support.
This is also why the early results were genuinely impressive, and why the cut-headcount instinct is so seductive. When a company routes first-line contacts through a competent AI layer, the visible numbers move fast: instant responses at 3am, queues that evaporate, resolution times measured in seconds, support available in every language a market speaks. A customer who used to wait forty minutes for a password reset now waits zero. For the routine tier, the machine is not merely cheaper — it is often a better experience, available always, never impatient, never having a bad day. That is real, and pretending otherwise is the kind of denial that gets a business disrupted.
But the impressive part contains the trap. Because the very legibility that made the routine tier easy to automate does the opposite to the rest of the function. When the machine clears the easy 60 to 70 percent, what remains is not a smaller version of the same job. It is a concentrate of the hardest, most emotional, most consequential contacts a business ever has. The frightened customer who just spotted a fraudulent charge. The grieving family trying to close an account. The furious client whose order failed before a once-in-a-lifetime event. The vulnerable caller who needs patience no script contains. The complaint that is one wrong word away from a regulator, a lawyer or a viral post. That residue is precisely the work AI handles worst and humans handle best — and it is now the entire job of whoever remains.
The companies that read only the first half of this story — the impressive automation of the routine — and cut their teams to match are the ones quietly degrading their service while congratulating themselves on a smaller payroll. We have watched this movie before in adjacent functions. Duolingo's AI-first move on its contractor base became a cautionary tale precisely because the hardest, most human slice of the work resisted clean automation. The companies reading the whole story do something far more interesting, and it starts with refusing the wrong frame.
A modern Singapore customer-service operations floor where AI dashboards and human agents share the same space
The misread: replacement is the wrong word
Here is the mistake, made in operations reviews across the region right now, and it is expensive.
A leadership team sees a competitor route 70 percent of contacts through AI and reasons backward to a headcount target. If they can deflect 70 percent, we can cut 70 percent of the team. The CFO models the salary savings. The number is large and attractive. A target is set. And the entire programme is framed, from its first slide, as a cost-reduction exercise wearing AI clothing.
This fails for a reason that is almost arithmetic, and it is the single most important thing to understand about AI and work.
AI does not replace jobs. It replaces tasks. A service role is a bundle of tasks — retrieval, logging, templated replies, status updates, and judgment calls, de-escalation, goodwill exceptions, fraud intuition, the human warmth that turns a furious customer into a loyal one. When you point AI at the role, it does not vaporise the role. It dissolves the automatable tasks within it and leaves the rest standing — often more exposed and more valuable than before. Automate the 60 percent of a service agent's week that is genuinely routine and you do not get 60 percent of a person back to cut. You get a person whose remaining 40 percent just became the most valuable 40 percent in the building.
The leader who frames this as headcount reduction makes two errors at once. First, they cut for the wrong number — chasing salary savings instead of the operating leverage that comes from redeploying freed capacity onto higher-value contacts. Second, and more insidiously, they automate the wrong tasks, because a cost-first mindset is impatient and reaches for the visible, emotional, customer-facing roles precisely because they feel like overhead. They save a little on the income statement and quietly destroy a lot on the balance sheet of trust.
There is a third error, quieter and more damaging over time: a cut-first programme poisons its own data supply. AI service systems improve through use — through the corrections, the edge cases and the tacit knowledge that frontline agents feed back. When those same agents have been told, implicitly or explicitly, that the AI exists to replace them, they stop feeding it. They route around it, withhold the knowledge that would make it good, and wait, not unreasonably, for it to fail. The replacement framing sabotages the very flywheel that would have made the AI excellent. The redesign framing does the opposite: agents who believe the AI clears their drudgery become its most patient trainers, and the system compounds. The frame you choose is not a communications decision. It is an input to whether the technology works at all.
And in customer service specifically, the replacement framing carries a unique poison: it automates the moments that matter most. The whole economic value of a service function is concentrated in a handful of high-stakes interactions — the save call, the complaint that could go viral, the fraud victim, the loyal customer at the moment they decide to stay or leave. Those moments are bucket-two, human-owned work, and a cut-first programme is structurally blind to them. It sees average handling time and deflection rate; it does not see the relationship that walked out the door. So replacement is not just the wrong word ethically. It is the wrong word strategically — it describes a programme that subtracts from the spreadsheet while quietly bleeding the asset.
Redesign, not replacement: the three-bucket model
If "cut the team" is the wrong frame, what is the right one? It begins with the most useful question any operator can ask of AI:
"If the machine clears the routine contacts, what could our people finally do with the time?"
That reframing produces a concrete, repeatable model. You do not start with job titles or a headcount target. You start with tasks. Pull a representative month of contacts — tickets, calls, chats, emails — and decompose the function into the discrete things people actually do. Then sort every task into one of three buckets.
Bucket one — what machines do better
These are the tasks where a competent model genuinely outperforms a human on speed, consistency, availability and cost. Order and status lookups. Password and access resets. First-line FAQ. Balance and transaction queries. Summarising a long case history into a brief. Drafting a first-pass reply. Routing a contact to the right team. Translating across the languages a market speaks. In most service operations this bucket is the majority of volume — and it is exactly where the early automation wins came from. Route this work to the machine without apology and without nostalgia. It is genuinely better at it, and the customer is better served by it.
Bucket two — what humans do better
These are the tasks where the human is not merely preferable but load-bearing. De-escalating a customer who has just discovered fraud and is frightened. Owning a complaint end to end and being accountable for the outcome. Handling a vulnerable, distressed or grieving caller with genuine care. Making the goodwill exception no policy quite covers. Catching the scam pattern that "looks fine" to a model trained on yesterday's scams. Holding the save conversation when a high-value customer is one sentence from leaving. This bucket is small in volume and enormous in value. Automate it carelessly and you do not save money — you bleed it, one churned relationship and one viral complaint at a time.
Bucket three — what they do better together
This is the bucket most companies forget exists, and it is where the real upside lives. It is the agent who now resolves the hard 40 percent brilliantly because a machine handled the 60 percent that used to eat their day. It is the support specialist who walks into every complex case with the full history already summarised, the relevant policy already surfaced, the draft already written — and spends their energy on judgment and empathy rather than retrieval. It is the human-plus-AI pair that turns a frustrating ticket into a moment of loyalty. Together they are not a smaller team doing the same job. They are the same team doing a far higher-value job.
The reason bucket three is so easy to forget is that it does not show up in the first round of cost modelling. A spreadsheet that asks "how many agents can we remove?" finds buckets one and two and stops. It has no column for "value created when a freed agent is pointed at the contacts that decide whether a customer stays," because that value is diffuse, shows up later, and lands on the revenue and retention lines rather than the cost line. So the cost-first analysis structurally undercounts the upside and overcounts the savings. The redesign-first analysis inverts this: it treats freed capacity as fuel for retention and growth, not as a line item to delete.
When you run this exercise honestly across a service function, the headcount question stops being the point. Bucket one shrinks the routine tier. Bucket three grows the value of the people who remain. Bucket two is fiercely, deliberately protected. That is redesign. Any change in headcount is a by-product of the task migration, not the goal of it. The house rule is simple enough to put on a wall: redesign before you reduce. Reduce first and you cut blind, automate the wrong contacts, and spend the following year rehiring and apologising. Redesign first and the rest takes care of itself — cleanly, defensibly, and without setting fire to the trust you spent years building. This is the same redesign discipline reshaping every function AI touches, from the contact centre to the marketing desk, where the writer becomes the editor of machine output rather than its casualty.
A clean conceptual diagram of three task buckets — machines, humans, and the two working together
What this means for Singapore
It is worth dwelling on why Singapore, specifically, makes redesign not just the wiser path but very nearly the only viable one for customer service. Three local forces converge, and they all push the same way.
First, trust here is dense, local and unforgiving. Singapore is a small, high-trust, reputation-rich market where word travels fast and switching is easy. A Singaporean customer will happily let a bot reset a password — and will switch banks, telcos or platforms over a single badly handled dispute, then tell everyone why, on Google, on social, in the group chat. In a market this tight, the "last third" of service — the bucket-two work — is disproportionately valuable, because the cost of getting it wrong is not one ticket. It is a relationship, a public review, and a reputation that compounds. This is exactly why the automate-everything-then-repair-the-damage approach is a worse idea here than almost anywhere: the trust you would burn is denser and slower to rebuild. The companies that win Singapore service do not deflect the hard contacts. They use AI to clear the queue so a human can lavish attention on the contacts that decide loyalty.
Second, the regulatory and sectoral frame rewards keeping humans accountable. In financial services, the Monetary Authority of Singapore has set clear expectations through its FEAT principles — Fairness, Ethics, Accountability and Transparency — for how institutions use AI and data. In practice, FEAT turns a human-in-the-loop and explainability from nice-to-haves into design constraints. A bank here cannot let a model make a consequential, customer-impacting decision unsupervised and call it efficiency; it must show the decision was fair, that someone is accountable, and that it can be explained. For a service operation that touches money, disputes and fraud, that is a direct instruction about which contacts belong in bucket two — owned by a human who can answer for the outcome. The regulation is, in effect, a specification for the redesign. It tells you exactly where the human must stay.
Third, and most decisively, the labour model is tripartite. This is the part overseas playbooks simply do not have, and it is a structural advantage for any operator willing to use it. Singapore's entire approach to economic change runs through tripartism — government, employers and unions moving together — and through an institutional machine purpose-built to redesign workers into new roles rather than discard them. When a US or European firm automates its contact centre, the displaced agent is largely the firm's problem or the individual's. In Singapore there is a dense, funded, deliberately constructed system for moving a person from a role AI is shrinking into a role the economy is growing. A company that frames its service-AI shift as "redesign and reskill" rather than "cut" does not merely look better in the press. It moves with the national grain, becomes eligible for real support, keeps the goodwill of its people and the public, and executes the smarter strategy anyway.
Put these three together and the picture is unambiguous. Singapore is close to the worst possible market in which to run a crude automate-and-cut service programme, and close to the best possible market in which to run a disciplined redesign. The trust dynamics punish the cut. The regulation constrains it. The labour model subsidises the alternative. An operator who reads only the global headline — AI replaces support teams — and imports it wholesale will collide with all three forces at once. An operator who reads the local physics will find that the smarter move and the supported move are the same move.
This is also where the small-team economics get genuinely exciting, because redesign is not just for the DBS-scale players. A lean Singapore business can now run a service operation that would once have required a department several times its size — a handful of skilled humans orchestrating an AI layer that handles the volume. That is the same pattern we see in the rise of the two-human, fifty-agent company: leverage moving to the operators disciplined enough to redesign the work rather than simply staff it. For SMEs especially, the lesson is liberating rather than threatening. You no longer need scale to deliver world-class service. You need design.
The Singapore enablers: the redesign machine already exists
The most underused fact in this entire conversation is that Singapore has already built the infrastructure to make service redesign the path of least resistance. Most economies are improvising their response to AI. Singapore spent two decades building a workforce-transition system, and AI is simply the next wave it was designed to handle.
Start with the reskilling and redesign machinery. Workforce Singapore and e2i run Career Conversion Programmes and Jobs Redesign support — precisely the mechanisms for moving a worker from a shrinking role into a growing one, with funding and structure behind the journey. SkillsFuture underwrites the reskilling itself, turning "your routine tasks are being automated" into "here is the funded path to the higher-value role on the other side." For a service leader, this changes the maths of the entire transition. The freed capacity from automating bucket one does not have to become a layoff line in a quarterly report. It can become a funded redesign of agents into complex-case specialists, retention owners, and AI-supervision roles — supported, in part, by national programmes built for exactly this.
Then there is tripartism, which is not decoration but the operating system of Singapore's response to disruption. Government, employers and unions move together, and they have navigated wave after wave of change — manufacturing offshoring, the financial crisis, the platform economy — without the social fractures other economies suffered. AI is the next wave, and the tripartite instinct is the same: transparent, gradual, reskilling-led change at a pace the social contract can metabolise. A service operation that handles its AI shift in this spirit — engaging its people, being honest about which tasks are migrating, funding the redesign rather than booking a one-time cut — protects something larger than its own brand. It protects the trust that makes the whole system work, and it earns the public goodwill that, in a market this small, is itself a competitive asset.
And in regulated sectors, MAS FEAT completes the picture by making the redesign defensible by construction. Because FEAT demands accountability and explainability on consequential decisions, the human-in-the-loop is not a compliance afterthought bolted on later — it is the design itself. A financial-services operator who redesigns service around "machines on the routine, humans on the accountable" satisfies FEAT as a by-product of doing the smart thing. Getting that posture right — so a regulator never asks a question you cannot answer, and so the brand and customer experience stay coherent through the change — is precisely the kind of work our sister practice Freemansland Creatives builds into a service redesign from the start, pairing the customer-experience and process design with the AI strategy and implementation that Freemansland leads.
The deeper point is that these enablers tilt the entire incentive landscape. In most countries the cheap, fast, default move is to cut. In Singapore, the cheap, fast, supported move is to redesign — because the funding, the regulation and the social contract are all pointed that way on purpose. The companies that lose here will not be the ones that automated too slowly. They will be the ones that ignored a redesign system the country had already built and paid for, and chose the crude cut instead. The infrastructure is sitting there. The only question is whether an operator is disciplined enough to use it.
A Singapore tripartite scene blending public agencies, employers and workers with subtle technology and reskilling motifs
The operator's playbook: five moves to run now
Strategy is only as good as the next action it produces. If you run a service operation — a contact centre, a support team, a customer-success function — in any Singapore business, the redesign compresses into five concrete moves. Run them in order.
1. Map contacts to tasks, not agents to seats
Pull a representative month of contacts and tag every interaction by the task it really is: routine retrieval, status, FAQ, transaction — or judgment, de-escalation, complaint ownership, regulated decision. Do not start from the headcount; start from what customers actually contact you about. You will almost always find that a clear majority of volume is genuinely routine — and that the routine hides inside contacts that felt like they needed a person but did not. This map is the single most important artefact in the programme. Skip it and every later decision is a guess dressed as a strategy.
2. Automate the routine — visibly to staff, invisibly to customers
Deploy AI against bucket one. But how you communicate it to your own people determines whether it works. Tell agents plainly: this clears your queue so you can own the hard cases that actually matter. Adoption collapses the moment staff believe the AI is in the building to fire them — they will route around it, withhold the tacit knowledge that makes it good, and wait for it to fail. Frame it as the thing that finally takes the drudgery off their desk and they become its best trainers. To the customer, the automation should be invisible: faster resolution, instant answers, a seamless hand-off to a human when needed, and never the sense of being trapped in a bot maze with no human exit.
3. Redesign the human role upward
This is the move almost everyone skips, and it is the one that makes the difference. Once the routine is gone, rewrite the role around judgment, complex resolution, relationship ownership and retention. The job did not get smaller; it got harder and more valuable. Title, pay and expectations should reflect that. If you automate 60 percent of an agent's tasks and leave their role and reward untouched, you have created a confused, under-paid, over-exposed employee handling nothing but your angriest contacts. If you redesign the role around its new high-value core — complaint specialist, fraud-and-trust handler, customer-success owner — you have created your most productive and most loyal worker.
4. Keep the human firmly in the loop where it counts
Bucket two is sacred. Disputes, vulnerable customers, fraud, anything carrying regulatory, financial or reputational weight — AI assists, the human decides and is accountable. Under MAS FEAT in financial services this is not optional, and outside the regulation it is simply good business. Design the workflow so the AI does the preparation — pulling the history, surfacing the policy, drafting the option — and the human does the deciding, with a clear record of who owned the call. Build the escape hatch too: a customer who needs a person must always be able to reach one quickly. The fastest way to destroy trust in Singapore is to trap a frustrated customer in an AI loop with no human door.
5. Reskill, don't release — and use the system built for it
Move freed capacity into the redesigned roles, supported by Singapore's reskilling infrastructure: Career Conversion Programmes and Jobs Redesign support through Workforce Singapore and e2i, and reskilling through SkillsFuture. The reclaimed hours should become retention, service quality and growth — not merely a one-time cost cut booked in a single quarter. A company that releases people banks a small saving once and loses the institutional knowledge that walks out the door. A company that reskills them compounds a capability advantage for years, keeps the goodwill of its people and the public, and moves with the national grain rather than against it. In Singapore, "reskill, don't release" is not just the kind move. It is the funded one.
Run these five and the AI service shift stops being something that happens to your business and becomes something you execute deliberately, on your own terms, with the regulator, the unions and your own people moving alongside you rather than against you.
The investor's close: the number that should actually move
Now for anyone allocating capital, because this is where the whole argument cashes out on an income statement.
The naïve reading of the AI service wave is "these companies will save the cost of their support teams." It is the wrong number to watch, and watching it will lead you to back the wrong businesses. The number that should actually move is revenue per employee — and beneath it, the operating leverage of the whole service function.
Here is the mechanism. Service has historically scaled the way a galley scaled: more customers meant more agents, more seats, more cost. Cost-to-serve and headcount marched together; growth and support labour were chained. AI breaks that chain. When the routine tier migrates to machines that cost a fraction of an agent's salary and scale without hiring, the relationship between customer growth and support headcount finally decouples. A business can serve more customers, in more languages, around the clock, without the labour curve rising in lockstep. That is operating leverage of a kind service businesses have rarely had — closer to software economics than to traditional support.
But — and this is the crux for an investor — the leverage only appears on the income statement if the organisation is redesigned to capture it. Two companies can buy the identical AI and land in opposite financial places.
The company that merely buys AI for its contact centre will show, eighteen months later, a slightly smaller support team, a meaningfully larger software and cloud bill, and — if it copied the cut without the redesign — a quiet, corrosive drift in satisfaction, retention and complaint resolution. Its cost-to-serve barely moves, because the savings were eaten by the technology spend and the churn. On paper it "did AI." In reality it spent money to slide.
The company that redesigns around AI shows something categorically different: rising service quality, flat-to-falling cost-to-serve, climbing retention, and revenue per employee rising as the same people — freed from the routine — handle the high-value contacts, save the at-risk customers and deepen the relationships that grow the book. Same technology. Same starting headcount trajectory. Completely different result on the income statement. One bought a tool. The other rebuilt the machine around it.
The question for an investor is no longer "is this company using AI in service?" Everyone is. The question is "is this company redesigning its service around AI, or just buying it?" Only one of those shows up as durable operating leverage — and only one of those keeps the customers it claims to be serving more cheaply.
Customer service fell first because it was the most legible function in the company — the most written-down, the most measured, the most repeated. That same legibility is why the cut-first instinct is so tempting and so wrong. The routine was always going to migrate to machines. The businesses that copied only that half of the story are spending this year rehiring and apologising, their service quietly worse and their customers quietly leaving. The businesses that read the whole story — automate the routine, protect the human third, redesign the role upward, and use the redesign machine Singapore already built — are becoming more capable, more trusted and more profitable than the rest. The lesson was never in the cut. It is in the redesign — and in customer service, it is sitting in plain sight, waiting for the rest of Singapore's operators to read it the right way. For more on how the AI workforce shift is being decoded for Singapore, explore the rest of our Insights.