Every insurer is talking about AI right now — reserving copilots, automated workflows, faster closings, smarter governance, you name it. The appetite is real, the pressure is real, and the promise is real too. At least in theory.
But if you move from strategy decks into the actual data — the million-row IFRS17 tables that power these companies — a very different story appears.
Last week, I opened a dataset from a large European insurer. Before I touched a model, the first issue came from something as mundane as file encoding. The second came from inconsistent movement types across two business units. The third came from an undocumented override someone had quietly added months ago, shifting the development tail without anyone noticing.
By the time the pipeline ran end-to-end, half the input contradicted its own definition.
So when people ask why AI isn’t delivering, the answer is simple:
AI isn’t the bottleneck. Reliability is.
The Illusion of AI Readiness
Insurance is data-rich but data-fragile.
Decades of claims history, complex reserving frameworks, and layers of documentation make insurers look like ideal candidates for AI transformation.
But volume is not readiness.
Documentation is not consistency.
And a data lake is not a guarantee that data behaves predictably.
I’ve seen files with 1.2 million rows that are perfectly usable — and files with 10,000 rows that are nearly impossible to trust. The difference isn’t size. It’s reliability.
What Breaks First (Hint: It’s Not the Model)
When AI projects fail, people blame the model.
But the first thing that actually breaks — and breaks most often — is the pipeline feeding the model.
Common culprits include:
schema changes no one documented,
definitions that drift over time,
manual patches added “just for this quarter,”
and tiny formatting variations that cause massive downstream issues.
None of these fail visibly.
They fail silently — which is far worse.
AI can handle noise.
It can’t handle semantic drift.
Four Reliability Failure Modes
Across insurers, four patterns appear again and again:
Inconsistent definitions across entities
Same column name, different meaning.Silent overrides and patches
Temporary fixes that become permanent surprises.Structural drift
Encoding changes, formatting inconsistencies, column order swaps.Semantic drift over time
A field name stays the same while its meaning subtly changes.
One issue is manageable.
All four together make reliable AI impossible.
What Reliability Actually Means
Reliability is simpler than it sounds. It means that your pipeline is:
consistent (same input → same output),
documented (in business language),
reproducible (can be run from scratch),
and testable (breaks loudly, not silently).
When reliability is absent, every model is built on sand.
When it’s present, everything accelerates.
The Real Race
The insurers who win the next decade won’t be the ones with the most advanced models.
They’ll be the ones whose data is boring in the best possible way: stable, predictable, and trustworthy.
AI is not the magic.
AI is the mirror.
It reflects the quality of the data you give it — with perfect fidelity.
And that’s why the real transformation begins not with algorithms, but with reliability.
