Novo AI's Julien Condamines: The Industry's 10% Leakage Number Is a Gross Underestimate

Julien Condamines, co-founder of Novo AI, didn't set out to fix insurance claims. As a consultant, he pitched an early GPT model to an insurer shortly after OpenAI's first public release; the insurer said no. He called a former Google colleague, a machine learning engineer, and built a prototype instead — the insurer bought in, and that engagement became a company. What Condamines found once inside the claims function is the real story: the industry's standard leakage estimate is wrong, and the reason claims has been ignored for decades has nothing to do with technical difficulty.

The leakage number every insurer quotes is too low

Condamines opens with a figure most of the industry treats as gospel: global claims leakage runs around 10%. He calls it "a gross underestimation." Novo AI's own portfolio data puts recoverable leakage closer to 15% in more developed health markets, and as high as 35% in Thailand and Vietnam — markets where the company started before expanding into Hong Kong and Singapore.

The context makes the gap matter. Singapore's medical inflation, he says, was reported at 16.9% in the most recent cycle — a number that has held for years while most health insurers in the market aren't turning a profit. Regulators have responded by capping premium increases rather than letting them track inflation; Condamines points to Malaysia's most recent cycle, where premiums were capped "at a few percent" even as medical costs kept climbing. Payouts are rising faster than the revenue meant to cover them, and insurers have no regulatory lever to close that gap themselves — which is what pushes leakage from an operational nuisance to a solvency question.

Claims stayed manual because cheap labor made the problem invisible

The more interesting claim isn't the size of the leak — it's why nobody plugged it. Condamines' answer is structural, not technological. In much of Asia, seasonal spikes in claims volume were solved by hiring temporary staff in low-cost markets. "That doesn't even make a dent on a P&L," he says — and a problem that doesn't show up on the P&L never becomes urgent.

That fed into a longer-running dynamic: claims has never been the most influential function inside an insurer, because insurers spent decades chasing top-line growth — more clients, more market penetration — rather than operational resilience. Claims was, in his words, "that dirty, dark basement where no one ever wanted to go because there was like, this kind of like very manual, clunky operation." As markets mature and new-client acquisition gets harder, insurers are being forced to open that door and look at what's inside.

The other half of the answer is that the tooling genuinely didn't exist until recently. Condamines is direct about this: Novo AI's core product wasn't feasible two or three years ago, so the industry's instinct until now was reasonable — look away, and improve something else instead.

What automated claims processing is actually replacing

Strip away the AI framing and the current state of claims is mundane and manual. In many markets, claims teams still work from stacks of paper, or PDFs at best; reviewers check line by line under time pressure because providers and patients both want to be paid fast. That pressure produces incomplete, low-quality data — which then makes it impossible to build any intelligence on top of it.

The specific gap Condamines returns to is medical coding. It's mandatory from providers in the US, but in most other markets it either doesn't happen or happens badly, leaving insurers to do it themselves — inconsistently. He gives a concrete illustration: teams end up using roughly 250 codes out of the roughly 70,000 diagnosis codes that exist, which caps how precisely anyone can assess whether a treatment plan was appropriate. He's careful to frame this as a minority problem, though: he estimates 80–90% of medical providers do good work, and the point isn't to treat every claim as suspect, but to give clinical teams the accuracy to spot the smaller group gaming the system.

Novo AI's model is to automate extraction and coding so that 80% or more of claims — the clean ones — require no manual intervention at all, freeing clinical reviewers to spend their time on the remainder: claims with billing patterns, length-of-stay, or procedures that don't match clinical guidelines. A separate monitoring layer watches provider billing patterns continuously to surface new fraud strategies as they emerge, rather than waiting for a periodic audit to catch them. The goal, as Condamines frames it, is to screen 100% of claims accurately instead of sampling a fraction of them manually.

Selling AI into insurers means selling patience

Condamines is unsentimental about how hard this is to sell. Insurance is a small, tightly networked, heavily regulated industry that has "been sold miracles for years," which makes buyers more skeptical than the average enterprise customer, not less. Half the job, he says, is building the technology; the other half is navigating internal stakeholder maps that function "almost like an escape game" — figuring out who to approach first, who will kill the deal on sight, and who actually owns the pain point. Asked later in the episode's rapid-fire round for his biggest career lesson, Condamines didn't reach for anything abstract: "sometimes making your stakeholders happy is more important than doing the job well" — an approach, he added, that once cost him a job.

The numbers back up the difficulty. The fastest deal Novo AI has closed went from first conversation to proof of concept in three months; others have taken up to two years, with implementation stretching a further six to eighteen months after signature. Pricing is structured as a percentage of the claims book rather than a flat fee, alongside a modest implementation charge mainly meant to secure stakeholder commitment. That structure is also why the ROI case only lands once leakage is quantified: as Condamines puts it, "...a third of that is actually thrown out of the window and I tell you, I can help you recover seven to 10 to 15%, then all of a sudden you start counting in millions."

He also pushes back on the idea that speed alone is the differentiator insurers should chase. His market research found that consumers largely can't distinguish one insurer from another — the product is one nobody wants to spend money on and hopes never to need. Fast claims payment isn't a competitive edge so much as a baseline that most insurers still fail to hit; he cites insurers quoting a 14-day turnaround time while actually averaging three to four weeks. The differentiation opportunity he sees is becoming an active guide at the point of vulnerability — helping a policyholder who breaks a leg abroad figure out which hospital to trust, rather than just eventually paying the bill. One early skeptic, he says, told him: "when I saw you coming and talking about AI [it] was like, yeah, here another snake oil seller" — before concluding Novo AI had actually done the work.

What this means for the region

Novo AI hasn't closed a Middle East deal yet — Condamines confirms discussions are underway but nothing has come through — which makes this a market to watch rather than one where the model is already proven. Its footprint has expanded quickly beyond its Singapore base: the company started in Thailand and Vietnam, markets Condamines describes as having "bigger headaches" for insurers, before adding Hong Kong and Singapore. A contract with an assistance company handling claims from ninety countries then pulled the business into Europe and Latin America.

His advice to Gulf insurers, offered cautiously, centers on an asset he thinks the region underuses: centralized, more open health data infrastructure, which he cites the UAE as having built out further than many other markets. His second point is more pointed: don't try to build this internally. AI talent doesn't stay in organizations that move slowly, and insurers structurally move slowly. His recommendation is co-design with vendors — insurers bringing deep domain expertise on regulation and risk, technology partners bringing the build speed — rather than insurers trying to stand up an in-house AI function to compete with startups and Big Tech for the same talent.

The dynamics he describes — medical inflation running ahead of regulator-capped premiums, claims data too poor to measure leakage, a claims function with little internal influence — are not unique to Southeast Asia. GCC health insurers face similar loss-ratio pressure and similar underinvestment in claims infrastructure. The difference, per Condamines, is that this region still has time to build the plumbing before the leakage becomes as entrenched as it is in the markets where Novo AI started.

The lesson worth taking from this conversation isn't "adopt AI for claims." It's that the claims function has been under-resourced for a reason — it was never on anyone's growth roadmap — and that reason is now expiring as premium growth slows and inflation keeps running. Insurers that treat leakage reduction as the new growth lever, rather than a cost center to trim, are the ones positioned to turn the claims moment into the trust advantage Condamines describes, instead of another friction point.

This post draws on the FS Brew episode Breaking Barriers in Insurance: Julien Condamines' Vision for Claims with Novo AI.