Jan Eirik Olsen

Transformation Program Manager at Tryg. Claims. Core systems. Fraud. AI. The human side of large-scale change across the Nordics. Views my own.

[Jan Eirik Olsen @jaeiol]

Fraud is where insurers have run AI in production for years. It is the mature case, not the frontier case.

The real question is not whether the model detects it. It is who is allowed to overrule the model, and on what grounds.

Governance, not capability.

[Jan Eirik Olsen @jaeiol]

Those checks are not always available. Not every purchase leaves a record you can reach.

Detection asks whether the evidence looks real. Verification asks where it came from. A good fraud defense needs both, and knows which answer weakens as the fakes get better.

[Jan Eirik Olsen @jaeiol]

The AI-generated claim photos that make the news are the ones that got caught. The extra finger. The receipt where the text turns to gibberish.

We hear about them because they failed. The fakes that matter are internally consistent.

[Jan Eirik Olsen @jaeiol]

No single control solves this. Fraud defense works in layers: the documents themselves, patterns across claims, identity checks, and verification against sources the claimant does not control.

AI-generated evidence weakens the first layer more than the others.

[Jan Eirik Olsen @jaeiol]

The reflex response is to ask everyone for more proof: more photos, original files, a video call with an adjuster.

That slows the fraudster down. It also punishes the honest majority and undoes the straight-through processing the industry spent years building.

[Jan Eirik Olsen @jaeiol]

In a Verisk survey of 300 US claims professionals, two thirds said manipulated media fraud often goes undetected.

https://t.co/mg8UQogHl7

[Jan Eirik Olsen @jaeiol]

The next gains are in the surge plan, not the digital channel: repair capacity agreed before the storm, triage to the worst-hit homes first, and an honest timeline on day one.

Nobody remembers how fast the form submitted. They remember how long the dehumidifier ran.

[Jan Eirik Olsen @jaeiol]

Reporting a water damage claim takes minutes. Drying out the basement takes months.

Insurers have spent a decade making the first part faster. The customer's experience is decided by the second.

[Jan Eirik Olsen @jaeiol]

What decides those months is physical: drying equipment, an assessor, and a contractor with free capacity.

After a big weather event, all three are booked at once. The form went in the same evening. Then the customer waits.

[Jan Eirik Olsen @jaeiol]

Better questions, and harder to game:

What share of claims had a human touch that added nothing?

What did that touch cost, in cycle time and in money?

Neither produces a flattering slide.

[Jan Eirik Olsen @jaeiol]

Every insurer has an STP number. Almost none of them mean the same thing.

Some count claims paid with no human touch at all. Some count claims that cleared automated validation and then went to an adjuster anyway. Some exclude the lines that would drag the figure down.

[Jan Eirik Olsen @jaeiol]

So 70 percent and 12 percent can be the same operation measured two ways.

That matters, because the number ends up in targets. A target built on a denominator you picked yourself is one you can hit without anything changing for the customer.

[Jan Eirik Olsen @jaeiol]

So a good deal of what gets called "high-risk AI" in non-life insurance does not meet the Act's definition of it. That says nothing about whether it is risky.

Sixteen months is enough time to build the controls. Also enough time to forget them.

https://t.co/To7cIaU5BR

[Jan Eirik Olsen @jaeiol]

"The EU delayed the AI Act" has been the headline since the summer. It is half right.

Regulation (EU) 2026/1744 came into force on 27 July and moved Annex III high-risk obligations from 2 Aug 2026 to 2 Dec 2027.

[Jan Eirik Olsen @jaeiol]

Scope gets even less attention than the dates.

Annex III 5(c) names risk assessment and pricing for life and health insurance. Non-life pricing is not named. Claims adjudication is not listed. Fraud detection is expressly carved out of the credit-scoring point.

[Jan Eirik Olsen @jaeiol]

The maneuver that makes the first interstellar probe affordable came from an AI lab, PSI. No maneuver, no mission under $15M, they say.

Announced the same day as the lab's $58M raise.

Hard to find an AI capability claim that isn't also a funding announcement.

[Jan Eirik Olsen @jaeiol]

Breakthrough Starshot, 2016: $100M pledged, roughly $4.5M actually spent, program now on hold.

Fermi Explorer, this week: under $15M, launching 2029, arriving in 79,500 years.

The slow one is the one that flies.

[Jan Eirik Olsen @jaeiol]

Full detail, including the radioisotope proposal: https://t.co/TqEYiLkZqq

[Jan Eirik Olsen @jaeiol]

Proposed payload on @FermiExplorer's 80,000-year probe: sealed samples of Ni-63, C-14 and Cl-36.

Half-lives: 101, 5,730 and 301,000 years.

Whoever finds it can date the launch from the decay ratios. A self-authenticating clock for a mission nobody alive will see land.