There is a number in this week’s reporting that should stop any Singapore hiring manager mid-sentence. Not the spending figure — the deployment figure. Because one of them is a budget line and the other is a verdict on how those budgets are being staffed.
What was reported on 7 September
Fintech News Singapore published “Financial Services AI Spending Rises, Yet Most Initiatives Still Can’t Show Tangible Business Value” on 7 September 2026, drawing on the Quinlan and Associates report From Pie in the Sky to ROI (July 2026).
The headline economics are unambiguous. Global financial services AI spending reached US$75 billion in 2025 and is projected to hit US$132 billion by 2030, growing at 12% annually. Adoption is effectively universal: 95% of traditional financial institutions are exploring or have already adopted AI.
Then the numbers turn. 71% of those institutions face challenges achieving clear returns. Only 56.6% of AI initiatives are successfully deployed at all. And of the ones that do reach deployment, only 21.1% fully achieve their objectives, with 35.5% achieving them partially.
Boston Consulting Group’s 2025 research lands in the same place from a different angle: 5% of organisations polled had achieved AI value at scale, while 60% reported no material value at all.
Our Expert Take (1 of 3)
Run the arithmetic on those two numbers and it reframes the whole conversation. If 56.6% deploy and 21.1% of those fully achieve their objectives, then roughly twelve out of every hundred funded AI initiatives fully do what they were funded to do. That is not a distribution you get from a hard modelling problem, because modelling difficulty does not fail that consistently in the same places. It is the distribution you get when the work dies in the gaps between teams: data access, systems integration, and the approval path nobody owns. Every one of those gaps has a job title attached to it, which is why I read this as a staffing report rather than a technology report.
The failure is not where the budget assumes it is
Split the failure into its two halves, because they have different causes and different hires attached.
The 43% that never deploy. In a regulated institution these projects almost never die because the model was inaccurate. They die on data access — the data exists but not in a permissioned, production-grade, lineage-traceable form — or on integration, or because the path from a working prototype to a production approval was owned by nobody in particular.
The deployed projects that miss their objectives. These usually shipped without an agreed definition of success set before the build started. Nobody can prove the thing worked, and nobody is willing to say it did not, so it persists in an expensive limbo.
Neither of those is a data science failure. Both are engineering and ownership failures — and that makes them hiring decisions.
About to fund an AI initiative in Singapore?
We benchmark the team shape around the budget — not just the headcount number — and shortlist engineers who have taken AI systems into production inside regulated institutions.
Let us talk it throughThe three roles I hire before the budget is approved
In this order, and the order is the point.
1. A data engineer who can reach production data. Not someone who builds pipelines in the abstract, but someone who has navigated permissions, lineage and retention rules inside a regulated environment and got clean data to a model without a six-month exception process. This role alone addresses the single biggest cause of the 43% that never deploy.
2. An integration or platform engineer. The output must land inside the system where the decision is actually made — the case management tool, the underwriting flow, the monitoring queue. A prediction nobody consumes creates no value by definition, and the distance between a working endpoint and a used one is almost always integration work.
3. An engineer who thinks in evaluation. Someone who insists on the success metric before the build begins and instruments it afterwards, so the project can be honestly judged. This is what converts a partially-achieving project into either a full success or a clean, cheap stop.
Only after these three does an additional applied scientist reliably change your outcomes. Most institutions hire in exactly the reverse order — scientist first, plumbing later — and then find themselves in the 43%.
Our Expert Take (2 of 3)
Singapore has a specific version of this problem that I do not see as sharply elsewhere. The city has genuine depth in quantitative and modelling talent because the financial sector has been recruiting for it for two decades. What it does not have in equal depth is engineers who have carried an AI system through a regulated production approval and out the other side. Those two populations look identical on paper and behave completely differently on a project. When I screen for the second group, I ignore the frameworks section of the CV entirely and ask one question: what did you ship, who used it, and what changed. Most candidates cannot answer the third part.
What this means for hiring in Singapore right now
Spending is rising and projected to keep rising to 2030. Competition for genuinely scarce profiles will not ease — so the question is not whether to hire, but what to stop hiring for.
| Where budgets go | Where the failures actually are | The hire that closes it |
|---|---|---|
| Model and tooling licences | Data not permissioned or production-grade | Data engineer with regulated-environment experience |
| Applied scientists | Output never reaches the decision system | Integration / platform engineer |
| Pilot programmes | No success metric agreed before build | Evaluation-minded engineer or tech lead |
| More pilots after the first stalls | Nobody owns the production approval path | A named owner, not another vendor |
If you are building the surrounding engineering team rather than a pure AI function, the same sequencing logic applies to hiring AI engineers in Singapore, and the delivery-side structure is covered in our approach to hiring developers into a growing team.
It is worth watching the Gulf on this too. The financial infrastructure build-out we track around engineering hiring in Dubai is competing for the same small population of production-hardened AI engineers, and it has state capital behind it. That competition is not theoretical: we see the same candidates in both pipelines.
Our Expert Take (3 of 3)
The most useful thing an employer can take from this reporting is permission to be unfashionable. When 60% of organisations report no material value, the differentiator is no longer ambition — everyone has that, and 95% of institutions are already spending. The differentiator is boring completion: data that arrives clean, output that lands in the system where someone acts on it, and a metric agreed before anyone writes code. Those are not exciting hires to justify to a board that has just read about frontier models. They are the hires that put you in the 12% rather than the 43%.
What to do in the next thirty days
1. Audit your stalled pilots against the two halves. For each one, write down whether it failed to deploy or deployed without hitting objectives. The answer points directly at which of the three roles you are missing.
2. Do not fund a new initiative without a named owner of the production approval path. Not a committee. A person, with the authority to clear it.
3. Write the success metric into the funding request. If it cannot be written before the build, the project is not ready to be funded — and that conversation is far cheaper now than in eight months.
4. Rewrite one AI job description around shipping rather than stack. Ask for evidence of a system in production and what changed as a result. The framework list filters nothing; that question filters most of the market.
Frequently asked questions
What did the 7 September 2026 report actually say about financial services AI spending?
Fintech News Singapore published a piece on 7 September 2026 titled Financial Services AI Spending Rises, Yet Most Initiatives Still Can’t Show Tangible Business Value. It draws on the Quinlan and Associates report From Pie in the Sky to ROI, published in July 2026, which puts global financial services AI spending at US$75 billion in 2025 and projects US$132 billion by 2030, growing at 12% annually. The same research finds 95% of traditional financial institutions are exploring or have adopted AI, while 71% face challenges achieving clear returns. The two numbers that matter most operationally are these: only 56.6% of AI initiatives are successfully deployed, and of those deployed only 21.1% fully achieve their objectives, with 35.5% achieving them partially.
Why do most financial services AI projects fail to show value?
The failure is rarely in the model and almost always in everything around it. A project that does not deploy usually died on data access, integration with core systems, or an approval path nobody owned. A project that deploys but does not hit its objectives usually had no agreed measure of success defined before the work began, so nobody could prove it worked or admit that it did not. Boston Consulting Group’s 2025 research points the same way: only 5% of organisations polled had achieved AI value at scale while 60% reported no material value at all. That distribution is not what a modelling problem looks like. It is what an engineering and ownership problem looks like.
Which roles should Singapore employers hire for AI initiatives?
Three, and typically in this order. First, a data engineer who can get clean, permissioned, production-grade data to the model, because data access is the most common reason a project never deploys at all. Second, an integration or platform engineer who can put the output inside the system where the decision is actually made, since a model that produces a prediction nobody consumes generates no value by definition. Third, an evaluation-minded engineer who defines the success metric before the build starts and instruments it afterwards. Only after those three does an additional applied scientist reliably change your outcomes. Most institutions hire in exactly the reverse order and then wonder why the pilot never reached production.
Is this a good or bad moment to hire AI engineers in Singapore?
It is a good moment to hire the right profile and a poor moment to hire the fashionable one. Spending is rising and is projected to keep rising to 2030, so competition for genuinely scarce profiles is not going to ease. But the shortage in Singapore is not in people who can fine-tune a model; it is in engineers who have taken an AI system into production inside a regulated institution and can show what changed as a result. Those people are rare, they are usually employed, and they are not found through keyword searches on a CV. Ask candidates what they shipped, who used it, and what measurably changed. That single line of questioning filters most of the market.
Hiring for an AI initiative in Singapore this quarter?
We shortlist engineers who have shipped AI systems into regulated production — and tell you honestly which of your three roles to fill first.
Let us talk — see vetted candidates