On 26 September 2026, TechCrunch published “Insurers claim AI is already increasing healthcare costs”, reporting a Blue Cross Blue Shield Association analysis that attributes an extra $942 million of healthcare spending over two years to hospitals’ use of AI tools in insurance claims. I have read a lot of “AI causes harm” stories this year and dismissed most of them. This one stopped me, because the mechanism is not a model behaving badly. It is a model behaving exactly as specified, in a system where nobody was accountable for measuring what else changed. That is a hiring problem, and it is one I have been getting wrong in Singapore.
What the Analysis Actually Found
The headline number is $942 million in additional spending over a two-year period. The finding underneath it is more interesting: BCBSA reported a sharp increase in patients being documented as having complex conditions, alongside no evidence of a corresponding change in the care delivered.
Read that twice. The patients did not get sicker. The treatment did not change. What changed was the description of the patient, and the description is what determines the payment. AI tools that help clinicians document more completely and code more accurately did what they were asked to do, and the aggregate effect was a transfer of money.
The quotes in the piece capture how differently the two sides read this. Luke Chalker, a senior vice president at BCBSA, said: “It’s not a war. It’s a completely one-sided blood bath.” Dr Shiv Rao, founder of the AI medical documentation company Abridge, described the trajectory as “a horrible dystopic future nobody wants to live in, with bots fighting bots, agents fighting agents.”
Note that Rao is not an outside critic. He builds the category. When the people shipping the tools and the people paying for the outcomes both describe the direction as bad, the disagreement is not about whether the systems work.
Our Expert View 1: This Is an Evaluation Failure, Not an AI Failure
If you gave this brief to any competent engineering team — “help clinicians document conditions more completely and accurately” — they would ship something that raises documented complexity. That is the specification. Success would be measured as completeness, coding accuracy, clinician time saved, maybe claim rejection rates. Every one of those metrics would look excellent.
The metric that was missing is the one that compares the documentation against the care. Nobody owned it, because it sits between two organisations that are adversaries. The hospital has no incentive to build it and the insurer has no access to build it.
I raise this because it is a structural pattern, not a healthcare quirk. Any AI system that optimises a description of reality, in a context where the description determines money, will drift in the direction of money unless something explicitly measures the drift. Fraud scoring, credit decisioning, procurement classification, insurance underwriting — Singapore teams are shipping all of these right now.
Our Expert View 2: “AI Experience” on a CV Has Stopped Discriminating
Here is where I have to be honest about my own screening. For the last eighteen months, when a Singapore healthtech or insurtech client asked for AI engineering capability, I was effectively filtering on model and framework exposure: has this person fine-tuned, has this person shipped a RAG pipeline, do they know the tooling. That filter is now close to useless, because almost everyone clears it.
What that filter never tested is whether the candidate can design an evaluation that would catch their own system doing something unintended. In my experience that skill is rare, it is uncorrelated with model expertise, and it is the entire difference between the two outcomes in this story.
The question I use now is simple, and I ask it about a system they actually shipped: “What metric would have told you this was working in a way you did not intend?”
Strong candidates answer within seconds, usually because they had already built a counter-metric and can tell you what it caught. Weak candidates restate the primary success metric in different words. That restatement is precisely the blind spot BCBSA measured at $942 million.
If you are recalibrating a technical screen for this, our Singapore method for screening developers in seven steps has the general structure, and the QA automation engineer guide is the closest existing playbook for hiring someone whose job is to prove a system is wrong.
Hiring for a regulated AI workflow in Singapore?
We screen for evaluation design before candidates reach you, so your shortlist contains people who have measured their own systems going wrong.
Let’s discuss your shortlistOur Expert View 3: Singapore Is Exposed Differently, Not Less
The obvious objection is that this is a US story about a US payer model, and Singapore’s system is structured differently. That is true, and it changes the shape of the risk rather than removing it.
Singapore healthtech and insurtech teams build for a mixed environment: public institutions, private hospital groups, and a competitive private insurance market on top of national schemes. The financial consequence of a documentation or coding change does not disappear in that structure, it just lands in different places — in private claims, in integrated shield plan pricing, in the disputes that follow.
There is also a local factor that cuts the other way, and it is genuinely favourable. Singapore engineering teams tend to operate under closer regulatory attention than their US equivalents, and data protection obligations mean that questions about accountability for automated decisions are already familiar in the room. The vocabulary exists. What is often missing is someone whose actual job is to hold the measurement, rather than it being everyone’s implicit responsibility and therefore nobody’s.
That is a role, and it is one I am now recommending clients staff explicitly rather than distribute.
The Role I Now Tell Clients to Open
It does not need a new title, and inventing one usually makes it harder to fill. In practice it is a senior engineer or a staff-level data person with three characteristics:
- They have worked inside a regulated workflow. Claims, underwriting, clinical, or financial compliance. They understand that a change in how something is described is a financial event, not a documentation detail.
- They design evaluations before they design models. Ask what they measured before shipping; the sequence in their answer tells you a lot.
- They have institutional standing to stop a launch. This one is organisational, not personal. If the person holding the counter-metric reports to the person whose bonus depends on the primary metric, you have not created the role, you have created its appearance.
On cost: a mid-level engineer on healthtech or insurtech data and AI systems in Singapore typically sits at 7,000 to 10,500 SGD per month, and a senior with real regulated-domain and evaluation experience at 11,000 to 16,000 SGD per month, before employer CPF for citizens and permanent residents, and before Employment Pass thresholds for foreign hires. The salary calculator gives current bands, and if you are scoping the wider build, our SaaS development in Singapore breakdown shows where the role sits in a full team.
The Same Conversation in Dubai
Our UAE colleagues have been having a near-identical argument with insurers and health groups there, with one difference worth noting: in the Gulf the constraint arriving first is data engineering maturity rather than evaluation maturity, because more of the underlying claims data is still being consolidated. Their guide to hiring data engineers in Dubai is the better starting point if that is your stage, and the QA automation team guide covers building the function whose job is to disprove your own systems.
What I Would Do This Week
Take one AI or automation system you already have in production. Write down its success metric. Then write down, honestly, the way that metric could improve while the thing you actually care about gets worse. If you cannot describe that mechanism in two sentences, you do not yet understand the system well enough to have shipped it.
Then find out who in your organisation is currently measuring that second thing. If the answer is nobody, that is your next hire, and it is a more urgent one than another model-building engineer. The BCBSA number is what the gap looks like after two years of nobody looking.
Building an AI team for a regulated Singapore workflow?
We place engineers who can design the evaluation as well as the model — and we will say plainly when the role you have written will not catch the problem you have.
Let’s discuss it with our Singapore teamFrequently Asked Questions
What did the Blue Cross Blue Shield Association actually report?
On 26 September 2026 TechCrunch reported a Blue Cross Blue Shield Association analysis finding that hospitals using AI tools in insurance claims processing generated an additional 942 million dollars in healthcare spending over a two-year period. The association said it found a sharp increase in patients being documented as having complex conditions, but no evidence of a corresponding change in the care actually delivered. In other words, the documentation changed and the medicine did not. BCBSA senior vice president Luke Chalker described the dynamic as not a war but a completely one-sided blood bath, while Abridge founder Dr Shiv Rao warned of a future of bots fighting bots and agents fighting agents.
Why does a US insurance story matter for hiring engineers in Singapore?
Because the failure is not an American one, it is a measurement one. The AI tools worked exactly as specified: they improved documentation completeness and coding accuracy. Nobody was asked to measure whether the system changed patient care or simply moved money. Singapore healthtech and insurtech teams are shipping the same class of system right now under the same incentives, and the engineers being hired to build them are generally screened on model and framework experience rather than on whether they can define a metric that would reveal this kind of drift. That screening gap is portable to any market.
What should I screen for instead of AI experience?
Screen for evaluation design and for the willingness to measure second-order effects. The single most useful question is to ask a candidate about an AI or automation system they shipped, then ask what metric would have told them it was working in a way they did not intend. Strong candidates answer immediately and often describe a counter-metric they put in place deliberately. Weaker candidates restate the primary success metric, which is exactly the blind spot the BCBSA analysis exposed. Also ask who owned the decision to ship, and what the rollback criteria were.
What do these engineers cost in Singapore?
In Singapore a mid-level engineer working on healthtech or insurtech data and AI systems typically sits between 7,000 and 10,500 SGD per month, and a senior engineer with genuine regulated-domain and evaluation experience sits between 11,000 and 16,000 SGD per month. Add employer CPF for Singaporeans and permanent residents, and factor Employment Pass qualifying salary thresholds for foreign hires. The premium worth paying is not for model-building skill, which is increasingly commoditised, but for people who have worked inside a regulated claims or clinical workflow and understand why a documentation change is a financial event.
Bryan
Delivery & Offshore Teams Expert at HireDeveloper.sg. Builds and runs distributed engineering teams for Singapore companies, with a focus on regulated delivery environments.