HomeAsk most people outside the insurance industry which function AI has reshaped most dramatically, and the likely answer involves a chatbot handling claims queries or a customer service line automating policy renewals. Inside Singapore's insurance industry, practitioners tell a different story: the function that has changed most substantially, with the least public visibility, is underwriting, the risk-assessment discipline that decides whether and on what terms an insurer takes on a given policy.
This matters for Singapore's broader workforce transformation conversation because underwriting sits precisely at the intersection this whole discourse keeps circling back to: genuinely judgment-heavy work, that until recently required years of accumulated pattern-recognition experience to do well, now substantially supported, and in some categories largely automated, by AI systems capable of processing far more risk-relevant data than a human underwriter ever could manually. The story is not simple replacement. It is a genuine redesign of what underwriting judgment actually consists of, and where in the process it gets applied.
For related context on how other MAS-regulated functions are navigating this shift, see our coverage of DBS's approach to AI reshaping roles across thousands of positions and JPMorgan's AI assistant rollout and its governance lessons for Singapore.
What Underwriting Actually Involved Before AI
Traditional underwriting, particularly for personal lines like motor and health, and for straightforward commercial policies, involved a substantial amount of data-gathering and pattern-matching work: pulling an applicant's history, checking it against actuarial tables and internal risk guidelines, flagging anomalies that warranted closer human attention, and arriving at a pricing and coverage recommendation. A meaningful share of this work, for well-understood, lower-complexity risk categories, was pattern application rather than genuine novel judgment: experienced underwriters had, in effect, internalised the actuarial logic well enough to apply it quickly and consistently.
This is precisely the category of work that modern AI risk-scoring systems now handle with genuine competence, often processing a wider range of relevant signals, more consistently, and considerably faster, than a human underwriter working through the same case manually. The volume-driven, pattern-recognisable share of underwriting has shifted substantially to AI-assisted or AI-led assessment, freeing human underwriters to spend a larger share of their time on the harder, genuinely judgment-dependent share of the caseload.
Where Human Underwriting Judgment Still Firmly Sits
The cases that remain squarely human territory are precisely the ones where the pattern-matching that AI does well breaks down: genuinely novel risk profiles that don't map cleanly onto historical data, high-value commercial policies where the financial consequences of a wrong call are severe enough to warrant deep human scrutiny, and cases where the underlying data itself is ambiguous or contradictory in ways that require judgment to resolve rather than a confident-sounding model output to paper over.
There is also a category of underwriting decision that remains human by design rather than by current AI capability limits: decisions where accountability, not just accuracy, is the governing requirement. MAS's FEAT principles, Fairness, Ethics, Accountability, Transparency, establish that a human must remain accountable for consequential underwriting outcomes, which means even in cases where an AI model could technically produce a defensible risk assessment, the accountability structure requires a named human underwriter who owns the decision and can explain its reasoning under scrutiny. This is not merely regulatory friction to be minimised; it is, properly understood, a design requirement that shapes how underwriting workflows should be built from the outset, model-assisted, human-accountable, rather than either fully automated or entirely manual.
The Fairness Question Insurance Cannot Avoid
Underwriting carries a bias risk that deserves more explicit attention than it typically receives in general workforce-transformation commentary, because insurance risk data has a long, well-documented history of encoding patterns that correlate with protected characteristics, sometimes lawfully as genuine risk factors, sometimes as historical artefacts of biased human underwriting practices that a model trained on that history would simply learn to replicate at scale, faster and less visibly than a human ever could.
A Singapore insurer deploying AI-assisted underwriting has a genuine, non-optional obligation to actively test whether the model's outputs replicate unfair historical patterns, not simply assume that a data-driven system is automatically more objective than the human judgment it partially replaces. This is precisely the kind of governance work that MAS's FEAT framework is designed to force insurers to do deliberately, and insurers who treat this testing as a compliance checkbox rather than genuine ongoing model auditing are taking on a risk that eventually surfaces, whether through regulatory action or, in a market as reputation-sensitive as Singapore's, through the kind of public trust damage that is expensive to repair.
What Happens to the Underwriter Career Path
The traditional underwriting career path, years spent on high-volume, straightforward policies building pattern recognition before progressing to complex commercial risk, is compressing in a way that creates both an opportunity and a genuine training challenge. Junior underwriters today are exposed to AI-assisted risk assessment much earlier in their careers than their predecessors were exposed to complex cases, because the routine volume that used to occupy their early years is now largely AI-handled.
This can accelerate genuine skill development, junior underwriters spending more time supervising and validating AI-generated assessments, and less time on repetitive data entry, potentially builds judgment faster than the old apprenticeship model did. But it only accelerates skill development if the training structure is deliberately redesigned to make it so. An insurer that simply removes the routine volume without redesigning what junior underwriters actually do with their time risks producing a generation of underwriters who supervise AI output without ever building the foundational pattern-recognition skill that supervision actually depends on. This is a genuine, under-discussed risk parallel to the junior-analyst concerns raised elsewhere across professional services functions facing similar AI-driven compression.
The Data Quality Problem Nobody Wants to Discuss Publicly
AI-assisted underwriting is only as good as the data it's trained and operated on, and Singapore's insurance data landscape, like most markets, carries real quality issues: inconsistent historical record-keeping across legacy policy administration systems, data silos between different lines of business that were never designed to share risk-relevant signals, and genuine gaps in coverage for newer, less-established risk categories where historical loss data simply doesn't exist yet in sufficient volume to train a reliable model.
Insurers who invest seriously in data infrastructure, cleaning and integrating historical records, building the pipelines that let AI risk-scoring systems actually access the full range of relevant signals, get materially better underwriting outcomes than insurers who bolt an AI tool onto an unchanged, siloed data environment and expect transformation to follow automatically. This is not a minor technical footnote; it is frequently the actual bottleneck determining whether an insurer's AI underwriting investment produces genuine improvement or an expensive, underwhelming pilot that quietly gets shelved.
What Distinguishes Insurers Getting This Right From Those Struggling
Across our conversations with Singapore's insurance and broader financial services sector, a consistent pattern separates insurers who have genuinely captured value from AI-assisted underwriting from those running an expensive, underwhelming pilot. The insurers succeeding treat the AI model's risk score as one input into a documented decision process, not the decision itself, with the underwriter's own reasoning, including any departure from the model's recommendation, captured as part of the case record. This might sound like a small procedural detail, but it is precisely what makes the eventual accountability review, whether internal, by MAS, or in the rare case of a disputed claim reaching a tribunal, defensible rather than a black box nobody can explain after the fact.
The insurers struggling, by contrast, tend to have deployed AI risk-scoring as a genuine black box that underwriters are implicitly discouraged from second-guessing, either through subtle performance incentives that reward fast processing over careful override, or simply because nobody built the workflow to make overriding the model's recommendation easy and well-documented when an underwriter's judgment genuinely differs. This latter pattern is a governance failure waiting to surface, and it tends to surface at the worst possible moment, during a regulatory review or a high-profile disputed claim, rather than during the quieter period when it would have been cheaper to fix.
The Reinsurance Dimension Worth Naming
A detail that rarely makes it into public commentary on AI underwriting, but matters genuinely to Singapore's insurance market given its role as a regional reinsurance hub, is how reinsurers are beginning to scrutinise the AI governance practices of the primary insurers they work with. A primary insurer that can demonstrate a well-documented, human-accountable AI-assisted underwriting process is, in the assessment of several reinsurance professionals we've spoken with, an easier and more favourably priced counterparty than one whose AI practices are opaque or poorly governed, because the reinsurer is ultimately exposed to the same risk-assessment quality the primary insurer's process produces. This is a genuine commercial incentive, beyond pure regulatory compliance, for Singapore insurers to invest properly in the governance discipline described throughout this piece.
The Governance Infrastructure Singapore Insurers Can Draw On
Singapore's institutional stack for navigating this transition responsibly is genuinely well developed relative to many markets. MAS's guidance on AI and data analytics in the financial sector, alongside the FEAT principles specifically, gives insurers a clear governance framework to build against rather than having to invent responsible-AI practice from first principles. The Institute of Banking and Finance has been expanding AI-literacy training relevant to insurance risk roles, addressing exactly the skill-building gap that compressed career paths create. And Workforce Singapore's job-redesign consultancy support can co-fund the underlying task-mapping exercise that identifies, function by function, which underwriting tasks genuinely shift to AI-assisted workflows and what the redesigned human role around them should actually look like.
The advisory and grant-structuring work that helps insurers document this redesign properly, in a form that satisfies both MAS's accountability expectations and the co-funding requirements of workforce transition schemes, is precisely the territory FMC Collective specialises in for Singapore's regulated financial services sector.
A Note on Customer Trust in a Small Market
Singapore's insurance customers, like the broader Singapore market, are a relatively tight, word-of-mouth-sensitive community, and a publicly disputed AI underwriting decision, a customer who feels unfairly assessed and takes that grievance to social media or the press, carries a reputational cost that a larger, more anonymous market might absorb more easily. This is a genuine, practical reason, beyond the regulatory obligation itself, for Singapore insurers to invest seriously in the fairness testing and human accountability structure described throughout this piece: the cost of getting underwriting fairness wrong in a market this size and this connected is measured not just in a single disputed claim, but in the broader trust erosion that follows when customers see a pattern rather than an isolated incident.
The Honest Bottom Line
Singapore's insurance industry offers one of the clearer illustrations, precisely because it operates under an explicit regulatory accountability framework, of what a properly governed AI workforce transition actually looks like: substantial automation of the volume-driven, pattern-recognisable share of a genuinely judgment-heavy profession, paired with deliberate protection and redesign of the human accountability that remains non-negotiable both by regulation and by the nature of what insurance actually requires. The insurers getting this right are not the ones automating fastest. They are the ones that have been most honest about which decisions genuinely require a human name attached, and have built their AI-assisted workflows around that honesty rather than around minimising it.