Everyone Wants AI. Few Can Provide the Data It Needs.
Inside the quiet shift happening in insurers: from model obsession to reliability obsession.
Every insurer I speak with is planning something ambitious around AI — reserving copilots, automated actuarial workflows, model governance agents, you name it. The appetite is real. The pressure from boards is real. And the promise is real too. In theory, at least.
But when you work inside the data — not in the slides, not in the strategy documents, but in the actual million-row IFRS17 tables — a very different picture emerges.
Just last week, I opened a large European insurer’s dataset. Before I even touched a model, the first failure came from something as mundane as file encoding. The second came from inconsistent movement types between business units. The third came from an undocumented override that someone had quietly added months ago, shifting the development tail without anyone truly noticing. By the time the pipeline ran end-to-end, half the input contradicted its own definition.
And yet the question everyone keeps asking is why AI isn’t delivering.
The answer is simple: it can’t deliver on top of data foundations that don’t hold. AI isn’t the bottleneck. Reliability is.
The Illusion of AI Readiness
Insurance is unusual among industries because it is both data-rich and data-fragile.
Large insurers have decades of transaction history, intricate reserving frameworks, and entire departments dedicated to data quality. On paper, this should make them ideal candidates for AI transformation.
But anyone who has worked hands-on with actuarial or claims data knows the hidden truth: volume is not the same as readiness, and documentation is not the same as reliability.
A company can have hundreds of tables, thousands of fields, and millions of records —
but if the business units interpret movement types differently, or if one country uses a patch that no one else knows about, AI becomes nothing more than a very expensive amplifier of inconsistency.
Insurers think they are ready for AI because they have data.
In practice, they are ready only when that data is consistent, predictable, and reproducible.
The Real Bottleneck: Reliability, Not Models
The most striking pattern across actuarial datasets — and this holds across geographies, business models, and maturity levels — is how often the pipeline fails before the model even begins.
And I don’t mean “fail” as in throw an error. I mean it fails silently.
The schema shifts because someone added a column.
A field’s definition changes but the documentation doesn’t.
A transformation is applied in one reporting cycle and forgotten in the next.
A CSV comes with a different encoding because a local team used another export method.
None of these things technically “break” a model. But they break the foundations on which the model relies.
AI is not fragile — it can handle noise. What it cannot handle is semantic drift: when the meaning of data changes without anyone noticing.
That’s why the real bottleneck is not the model. It’s the reliability of the pipeline that feeds it.
Four Failure Modes That Sink Projects Before They Begin
Across insurers, four reliability failure modes appear again and again:
1. Inconsistent definitions across entities: The same movement type or reserve category behaves differently country to country. Models don’t just struggle — they learn contradictory truths.
2. Silent overrides and one-off patches: Someone fixes a problem “for this quarter only,” but the fix becomes permanent and undocumented. Six months later, nobody knows why the numbers changed.
3. Structural drift (schema, encoding, formatting): Small formatting differences create huge downstream effects. UTF16 vs UTF8, decimal vs comma separators, shifting column orders, missing headers.
4. Divergent semantics over time: A column originally meant “reported claims” gradually becomes “reported + negatives,” but only in specific lines of business. The field name stays the same. The meaning no longer does.
Each of these issues is survivable on its own. Together, they create a situation where pipelines no longer behave predictably — and AI becomes impossible to trust, no matter how good the model is.
What Reliability Looks Like in Practice
People often ask me: “What does reliability even mean in a data pipeline?” It’s simpler than it sounds. Reliability means:
Consistent: the same input always produces the same output.
Documented: every transformation is explained in language the business understands.
Reproducible: the pipeline can be run from scratch with identical results.
Testable: small changes trigger explicit tests, not surprises.
When reliability is missing, every model sits on a foundation of sand. When it’s present, everything accelerates — especially AI.
This is the paradox: The more reliable the data, the simpler the AI required. Clean, consistent data makes even basic models unexpectedly powerful.
The Hidden ROI of Reliability
Companies often underestimate how much reliability reduces cost and increases speed. Here is the hidden ROI:
Faster closings: When pipelines run cleanly, month-end doesn’t feel like a triage zone.
Lower audit and regulatory risk: Stable definitions mean fewer surprises in submissions.
Cheaper AI projects: Reliable data reduces the number of iterations, the engineering overhead, and the need for custom logic around edge cases.
Better business decisions: Executives don’t have to wonder if a shift in reserves is real or caused by a hidden data quirk.
Reliability isn’t glamorous. But it is extraordinarily valuable.
The Bottom Line: AI Is a Mirror of Your Data
AI is not going to transform insurance because it is intelligent. It will transform insurance because it is consistent — once the data allows it to be.
The insurers who win the next decade will not be the ones with the fanciest models or the most ambitious AI roadmap. They will be the ones whose data lineage is clean, whose definitions are stable, and whose pipelines are boring in the best possible way.
When reliability becomes a default, AI finally becomes predictable. And when AI becomes predictable, it finally becomes useful.
In the end, AI is not magic.
It’s a mirror.
It reflects back the quality of the data you give it — with brutal honesty and perfect fidelity.
If the foundations are reliable, AI becomes a multiplier.
If they are not, AI becomes an amplifier of chaos.
And that is why the real race in insurance today is not to deploy AI.
It is to produce data that can be trusted.
Reads of The Week
In this punchy and provocative essay, Jurgen Appelo argues that today's AI—flawed as it is—often delivers clearer, more coherent answers than the average human. For Wangari Digest readers navigating misinformation, digital overwhelm, and late-night curiosity, his piece is a timely call to sharpen our critical thinking and rethink who we turn to for insight. It's not about glorifying machines—it’s about confronting just how often we accept sloppy reasoning from people while holding AIs to impossibly high standards.
This dense roundup from Lux Opes is a goldmine for readers tracking how European companies are navigating economic crosswinds—from aerospace surges and AI-driven tech pivots to legal minefields and ad market slumps. It’s a sharp snapshot of how strategy, supply chains, and financial discipline shape resilience across sectors. There’s something here for anyone watching how real-world companies adapt under pressure.
For readers curious about how to spot hidden gems in the financial world, this deep dive into Markel and Kinsale offers a masterclass in what makes a great insurance business—and why some deserve a permanent spot in a smart investor’s portfolio. The Dutch Investors break down the concept of “float,” the power of disciplined underwriting, and how tech and management alignment create lasting advantages. It’s a reminder that some of the most profitable long-term plays lie in boring, brilliantly-run businesses that quietly compound wealth behind the scenes.