
Let’s start with the obvious: Every company and their mother has automated every thinkable process with AI. (Or, if they haven’t done so already, somebody’s using a personal ChatGPT / Claude / Perplexity / Manus / you-name-it to make their job a bit easier.)
Before you click away to something more interesting, consider this: Plenty of these AI projects are messing up. Can we fix some of these — using AI?
The answer, as with anything business, is “it depends.” Which is not a satisfying answer. So let’s break it down to a use case we can actually work with, to develop a better answer.
I’ve worked in various functions in insurance, and I’ve seen my share of insurance claims data. I also know that, even before AI became a thing, insurances quite routinely mess up claims payments. And I don’t mean the instances where the insurance doesn’t pay — most regular people I know get frustrated that theirs doesn’t every year or so! — but the instances where the insurance company pays too much for a claim.
Companies don’t like to lose money. However, because claims processing is such a lengthy and not-fun process, it’s also one of those activities that are first in line for AI automation. And, AI playing “nice” so often, there’s a real risk that claims overpayments happen more, not less, in an AI-powered world.
There’s a myriad ways that this could happen, by human or AI, but a few examples are:
Overpayments: The insurance company paid more than the policy asked for, or — in rare cases — more than the claim itself asked for.
Duplicate payments: The insurance company accidentally pays the same claim twice.
Missed recoveries: Although the original claim payment was correct, the insurance could have gotten back some of that money through its own reinsurance, the accident partner, or re-selling a wrecked asset.
Fraud: The claim should have never been made, but the insurance didn’t spot that.
Process errors: A cat jumped onto the case worker’s keyboard and clicked “send money.” Too bad.
What all these categories have in common is not only that they cost insurances real money, but also that they can get more frequent because AI automations of such processes are less than perfect.
That begs the question: Can we use AI to mitigate this? In other words, can we use AI to spot where humans or AI are costing the company undue money?
How AI safeguards against hemorrhaging money
Let’s start with the technology. “AI” is a bit fuzzy a word, especially these days — so what specifically can help detect where an insurance is paying out too much money?
The most boring component of this are leakage rules. This is basically the complex job of ingesting a policy, the claims, and any surrounding context, and deciding whether or not something leaked — and if yes, determining exactly which rule it broke. Language models are really good at this, but also error-prone. An AI checker would ideally need a detailed bank of known rules in order to catch the 80%-or-so worst offenders. Finer, more flexible logic can then take care of detecting the rest.
The second tech brick isn’t actually AI in itself: comparative evidence. In other words, what did the company pay for similar claims? Assessing similarity is a job AI is increasingly good at, but frankly, in many cases a good old machine learning model can get you further faster.
Third, causal attribution. In other words, okay, there was a claims leakage — but why? And I don’t just mean which rule was violated, but is it because this claim is particularly tricky, or because the office cat needed to hug a keyboard? Of note: causal attribution isn’t AI in itself, but it used to be a very artisanal craft (for a data scientist). These days, causal algorithms can increasingly be run by AI agents, which makes the whole task more scalable (sorry, office cats).
So this is not “let’s train a language model on claims leakage data.” Even if insurances partnered up to pool all claims leakage data that had ever existed (a gargantuan task that would keep leagues of insurance staff busy for months), my intuition tells me that the resultant model wouldn’t be that great. Rather, we start with what is already in front of our noses (a moderate-sized dataset of existing claims, some of which were overpaid). We then use some common-sense data science and combine it with language AI where that gives real leverage.
Sounds less sexy, but works better in practice. And it’s much cheaper to implement as well, because it requires one dataset and maybe 2-3 people instead of thousands of data-collectors and model-trainers, not to mention GPUs.
How to automate this
This is the point where innovation teams and C-level executives usually ask “cool, can we buy this off-the-shelf?”
Frankly, not yet. Not from me and not of anybody I know of or researched.
That’s not to say that there aren’t solutions out there — this one, this one, and this one, amongst others. However, with all these vendors it’s less like you’re buying washing powder that you stick in the machine so you can admire spot-free washing afterwards. It’s more like a bunch of people who know a lot about dirty laundry, so they’ll take a look at your stains and figure out what’s best for your specific machine, especially because you’re constantly upgrading it like some kind of, I don’t know, washing machine addict? (Insurances constantly upgrade certain tech — and keep Excel sheets going for less sexy areas.)
To put it into more general terms: There are so many moving parts in enterprise AI right now that building a solution that works off the shelf and works reliably for all kinds of use cases is still virtually impossible. Paradoxically, this means that enterprises thus need to rely on human expertise — their own staff or external consultants — to figure out the next course of action. As with every technological wave, step by step a picture emerges of what works and what doesn’t. It typically takes anywhere from 3-7 years for even halfway-established solutions to become available “off the shelf.” Even those vendors who sincerely try to build and sell off-the-shelf-products are in reality still fixing a lot of things manually.
Wangari is not unusual in that regard: Like many other vendors, we have some platformized solutions, but lots of the real value delivery still happens in case-by-case human labor. Even when we’re increasingly handling AI-generated output!
What this tells us about the future of AI
Zooming out of this one use case, the emerging pattern I see is this: We’re building AI automations, and they mess up in various ways. Sometimes in human ways, sometimes in ways nobody anticipated. (Such AI systems are on the market galore.)
Then we build other AI systems that correct these first-wave AI automations. Those secondary AI systems currently require plenty of human oversight, because no-one really knows yet how exactly AI will mess up, when, and how badly.
Once we’ve properly characterized the primary AI automations using secondary AI tech, I see a fork in the road:
We use insights we got from the secondary AI to improve the primary AI, thus rendering the secondary AI obsolete over time.
We build a perpetual tug-of-war between primary and secondary AI. Like lawyers in court, the duplication has a cost — but each AI also improves the performance of the AI players around it.
I see different scenarios happening in different industries. Where getting things roughly right and keeping costs fairly low matters, option 1 is better — we improve the primary tech and kill secondary tech. Where AI budgets are pretty limitless and costly mistakes do matter, option 2 happens — this tends to be the case in bigger organizations, or rather lucrative industries such as (you guessed it) insurance.
Either way, I don’t see humans going away any time soon. Somebody needs to maintain that primary AI, or orchestrate the tug-of-war between primary and secondary AI. Increasingly, we’ll just see humans getting busier and busier correcting finer and finer mistakes that AI tries to avoid.
Meanwhile, at Wangari
Germany’s in one week! At Machine Learning Week Europe. I’ll be running two sessions:
A workshop on how to create reliable, high-quality sports predictions — applied to this season’s Bundesliga. I co-authored the book Soccer Analytics with Machine Learning (O’Reilly, 2026); this is the practical version of it.
A table discussion titled How to get AI projects production-ready. I’d run a course on that topic last summer; it’s just getting timelier now.
If you’re still in for a stay in Munich from Nov 16-18 — tickets are still on sale and you can get 15% off with the code AriMLW on checkout. See you there!
Reads of the Week
Xavier de Gracia sees a huge opportunity in AI for specialty practices. Providing adequate care for patients, while keeping costs for the patients themselves, can actually be solved with AI, he posits. The number is a whopper: That opportunity could be worth $108 billion.
We just talked about primary and secondary AI systems in the piece above. Merlyn Shelley goes a step further and shows that AI agents can become a Frankenstein-style mess. In enterprise contexts, avoiding this means being very deliberate about observability, access constraints, and model usage costs. While a one-size-fits-all solution doesn’t (yet) exist, these are some very useful thoughts to keep in the back of your mind as you’re building.
Zooming out a bit further, Roder Mortega points out that, although companies probably have plenty of data on you, you’re allowed to become something unexpected. Your past is data, not your destiny. It’s good to let AI predict your behavior, but it’s ethically questionable to use that prediction to further influence your behavior. Whatever happens — human dignity is beyond measurement. This is a well-written reminder thereof.



The post focuses on AI overpaying claims, but the mirror-image failure is already documented: UnitedHealth’s nH Predict tool, which plaintiffs say carries a 90% error rate (nine of ten denials reversed on appeal), stayed in use because only about 0.2% of policyholders ever appeal. That suggests a leakage checker is only half the audit, since an insurer’s incentives make wrongful denials far less likely to get flagged than overpayments. Causal attribution could help there too, because a denial pattern that survives only because nobody appeals looks very different from random process noise. I summarized the UnitedHealth case, including its lobbying and the 5% Medicare Advantage rate boost, in a related piece: https://wdrury.substack.com/p/wicked-shady-list-s-and-p-500-kickoff