Many insurers are rushing to deploy AI before they have taken the time to ensure they can trust what it will say.
The industry is sitting on decades of expertise, but it is scattered. It lives in underwriting guidelines that have been updated over time, claims notes written by different adjusters, engineering reports, inspection writeups, loss history, and a lot of individual judgment that never made it into a system at all. Then, we layer AI on top of it and expect clean, confident answers.
Because of this, users are hitting massive roadblocks when the underlying knowledge is inconsistent, disconnected, or hard to interpret. What many still don't realize is that AI is simply not designed to fix human errors stemming from baseline data inputs. In fact, it only makes it easier to surface those incomplete or misleading answers faster. And in insurance, those answers can show up in major underwriting decisions, claims handling, client advice, and risk recommendations, which can have a negative domino effect on critical decision making in a highly regulated industry.
That said, the issue is not whether AI is "ready." The question is whether insurers have done the groundwork to make their knowledge usable and trustworthy in the first place.
A recent initiative at FM is a solid example of this foundational work in practice. The company enabled more than 1,500 engineers to access decades of engineering knowledge through an AI-assisted platform. The real impact was consistency and confidence, with speed as an added benefit. Engineers could get to the right answer faster, but more importantly, they could trust what they were seeing.
This is a persistent challenge that exists across insurance. Whether it is underwriting appetite, claims precedent, loss control recommendations, or how you respond to a client in a complex scenario, the success of AI depends entirely on what lies beneath the surface.
Here are three things insurers should consider before these tools start influencing real work:
1. Stop organizing knowledge for storage. Start organizing it for judgment.
Most knowledge environments are built like filing cabinets, which works until someone actually needs to make a decision.
An underwriter is typically not thinking, "let me go find one document." They are weighing signals. Prior experience, similar risks, exceptions, engineering insight, loss trends. The same goes for claims, risk engineering, etc. The work is contextual versus linear.
If your knowledge is stored in disconnected places, AI will pull from it in disconnected ways. The answer might look polished, but it will not reflect how real decisions get made.
What worked in the FM deployment is that they did not just make content searchable. They connected it in a way that matched how engineers think about risk. That required a different level of effort.
Insurers need to take the same approach. If you want AI to support underwriting or claims, then your knowledge has to be structured around real decision paths that move beyond document repositories.
2. Assume your answers will be wrong and build for that reality.
This is where a lot of organizations get too comfortable too quickly. There is an assumption that if you are pulling from approved content, the answers will align. That is not a safe assumption once you are generating responses dynamically.
In insurance, being "mostly right" is not good enough. If an AI tool points an underwriter in the wrong direction, or gives a claims professional an incomplete answer, the downstream effect is real. It affects decisions, consistency, and in some cases client outcomes.
The teams that are getting this right treat accuracy as something that has to be earned every day. They test outputs, create feedback loops, and involve subject matter experts in refining results. They also assume gaps will exist and design for continuous improvement.
At FM, trust came from validating how their system performed in real-world scenarios and improving it over time.
That is the mindset insurers need. Not "is this ready to launch," but "how will we know when this is wrong and how fast can we correct it."
3. Decide up front where AI stops and people take over.
A lot of governance conversations sound good on paper but fall apart in practice because they're too vague. In insurance, you need clear, concrete lines to protect the integrity of the work. Where is AI allowed to assist? Where is human judgment required? What needs to be reviewed before it is used in a decision? What level of confidence is acceptable?
If an AI tool is helping draft an initial view of a risk, that is one thing. If it starts affecting pricing, coverage interpretation, or claims outcomes without clear oversight, that is something else entirely.
The strongest implementations treat AI as an extension of the professional. That means building in transparency, showing where information came from, and making it easy for people to challenge or verify what they are seeing.
The closer these tools get to core insurance functions, the tighter that control needs to be.
In summary, there is a lot of energy around AI right now, and for good reason, with tremendous opportunity on the horizon. In parallel, industry rigor should never be sacrificed for speed, and demands an established, accurate data foundation prior to deploying AI tools.
The firms that will get this right are not the ones rolling out the most tools the fastest. They are the ones doing the less visible work, taking the time to structure knowledge properly, putting real discipline around accuracy, and being explicit about where AI fits into decisions and where it does not.
Done right, AI becomes a force multiplier for your best people, makes expertise more accessible, and makes decisions more consistent.
If you skip it, you are just introducing a new layer of risk into an industry that already understands how costly that can be. That is the tradeoff. And it is one insurers need to think through carefully before they deploy at scale.
