We are in the middle of one of the most exciting periods in technology. AI is reshaping industries, automating tasks that took teams weeks, and making things possible that weren’t even on the table a few years ago. Insurance is no different, and the appetite to make use of that is genuine and well-placed.

Generative AI in particular has caught a lot of attention, and it’s easy to see why. It’s powerful, easy to access, and improving quickly. But that momentum has also led to it being applied to very different kinds of problems it was never designed to solve on its own. Data cleaning is one of them.

What data cleaning actually involves

Getting insurance data from a raw submission to something usable downstream isn’t one step. It’s a chain of parsing, extracting, cleaning, standardising, and validating, and each stage builds on the last. Get any of them wrong, and everything downstream is affected.

GenAI has been positioned as a solution across much of this chain, and at certain stages that’s reasonable. But the complexity isn’t evenly distributed. The earlier stages of parsing a document and extracting values into the right fields are more structured and forgiving. GenAI works well here because all the context it needs is already sitting in the prompt. It’s reorganising what’s in front of it.

The difficulty starts at validation, checking that the output is actually correct. And validation happens at two levels. There’s checking whether individual cells are right: is this value in the right format, does it fall within an expected range, does it match a known list. And then there’s checking whether values work together, a state, city and country that all pass individually but form a wrong combination. A construction code and occupancy code that are both technically correct but make no sense as a pair.

Where GenAI falls short

GenAI models are built for flexibility. They handle unstructured inputs, generate plausible and probable outputs, and perform well across a wide range of tasks. That’s a genuine capability with real applications in insurance and beyond.

But flexibility and consistency are different things. The same input can produce a different output depending on when you ask, which model version you’re running, and how that version has been updated. For writing a summary or drafting a response, that variability (sometimes confused for creativity) is usually fine (and often desired). For structured data feeding downstream systems, it isn’t.

These models are built by compressing petabytes of internet data into something a fraction of that size. In that compression, detail is lost, and what the model retains is the average. So it’ll never get the capital of France wrong. But ask it to validate an obscure street name against a specific zip code and it’ll give you something plausible with the first few digits right, and the rest a best guess. It’s not looking anything up. It’s predicting what a correct answer probably looks like.

You need a fact checker, and GenAI is a text generator.

We’ve written about the different types of AI and where each one fits here. That level of complexity needs deterministic AI, an approach that validates against known facts and known relationships rather than predicting what a correct answer probably looks like. Same input, same output, every time. For insurance data that feeds everything downstream, that should be a given.

The cost overlooked in the business case

Because GenAI has improved so dramatically, getting to 80 or 90% accuracy on cleaning tasks has become genuinely accessible. You can build something that looks like it’s working relatively quickly, and that’s made the decision to build in-house feel more attractive than it used to. What’s harder to see is what comes next.

GenAI model versions are typically retired every 12 months or so, with 60 to 90 days’ notice to migrate. What comes next may perform quite differently on your specific use case. The prompt engineering that took months to get right doesn’t carry over, instead you start again. And that last stretch: the edge cases, the unusual combinations, the inputs that fall outside familiar patterns, that’s where errors accumulate and where the cost of getting it wrong is highest. Those errors tend to look right, which in insurance is often more dangerous than an obvious mistake.

That’s a rolling cost that never makes it into the original business case, because it’s invisible until you’re already deep in the trenches.

If you’re working through the build vs buy question, we covered it in depth with InsTech on their recent podcast, worth a listen if the decision is live for you.

Using the right tool for the job

Data cleaning isn’t a competitive differentiator for most insurance companies. The carriers and MGAs at the front of the market are there because of how they underwrite, price, and manage risk. Data quality enables all of that, but the infrastructure to deliver it isn’t where they’re competing.

GenAI deserves its moment and there are real problems it solves well. But can it deliver clean insurance data on its own? Every pricing decision, every risk model, every portfolio view starts with the data. If that data is only probably right, so is everything built on top of it. At its core, this needs deterministic AI, making sure what comes out is actually right. It comes down to choosing the right tool for the job and being willing to question the default.