Consider the following quote from George Box: “All models are wrong, some are useful.”
Most of today’s models are free. And models power AI outcomes. When considering your enterprise strategy, AI should fit snuggly into business processes that successfully execute tasks supporting the overall strategic vision – underwriting policies, pricing risk, settling claims or helping customers.
Each of these areas represents tens, if not hundreds, of thousands of decisions made every single day by a single insurer. When those decisions are made by models inserted into business processes, it introduces risk.
Unfortunately, research from SAS, “Breaking Silos,” found that nearly one-third (31%) of respondents do not consider the impact on other areas of the business when making decisions.
It’s reasonable to expect the models we build (or those built for us) will deliver outcomes similar to those we have come to expect across various business areas. But the difference is the power of today’s AI – it can render millions of decisions in just a second, introducing downside risk that can irreparably damage your brand.
Listen to insurance experts discussing their outlooks on the industry – and how you can respond – in the latest episode of Brewing Curiosity – Insurance Unfiltered.
Can you ask your model why it made a decision?
You could ask a person why they made a decision if it was later scrutinized (say, during a Market Conduct Examination). But what about a decision made by a model?
The technical aspects of model risk management and capital solvency fit hand in glove. Risk is accepted and claims are paid as part of the routine business of insurance. Models performing that work will either add to capital reserves or take from it.
If the enterprise experiences adverse risk selection or delivers a poor claims experience, capital deteriorates. And that puts the organization at risk:
- Their financial outlook could be downgraded.
- They could face unnecessary scrutiny by the regulator.
- They might even become insolvent.
Gartner expects worldwide IT spend to reach $6.31 trillion in 2026 (up 13.5% from 2025), and we expect AI (and the models that power decisions) to be increasingly embedded into our processes. (For context, this estimated spend is more than the entire annual GDP of the United Kingdom.)
So, business leaders and executives are right to ask the question: “How do I safely use AI and models in my business without introducing (potentially fatal) downside risk into my enterprise?”
When governance is spread across spreadsheets and disconnected tools, executives and decision makers believe risk is under control – until a regulatory review, stress event or model failure exposes critical gaps.
Models, mistakes and managing outcomes
In the most recent episode of “Brewing Curiosity – Insurance Unfiltered,” Alex Semenov shares his expertise on model risk, capital adequacy and enterprise governance. He discusses how his role sits at the intersection of all critical functions of any insurance company.
It’s not enough to determine model effectiveness. Insurers must guarantee model control, explainability and defensibility.”
“Models sit everywhere [in an insurance company],” he explains. “They exist in pricing, reserving, capital, investments, underwriting and claims. Yet governance is often fragmented. We still see critical risk processes managed through Excel-based inventories and manual controls, which work…until they don’t.”
The reality Semenov describes carries significant risk. As a model deteriorates, so does its performance. And if controls are inadequate, visibility into model failure can be limited or missing entirely.
“When governance is spread across spreadsheets and disconnected tools, executives and decision makers believe risk is under control – until a regulatory review, stress event or model failure exposes critical gaps.
“From a capital and solvency standpoint, that’s dangerous. Regulators like OSFI and AMF are less concerned with whether a model is complex and more concerned with whether it’s controlled, explainable and governed.”
And it’s not just OSFI and AMF – recently, the NAIC Third-Party Data and Models (H) Working Group announced its main focus would be developing and proposing a framework for the regulatory oversight of third-party data and predictive models. When adopted, model failures will be scrutinized immediately.
“The worst-case scenario isn’t just a finding,” Semenov says. “It’s loss of credibility with regulators and then the board of directors. Once that happens, insurers are forced into costly remediation programs, re-documenting years of model decisions after the fact. That’s time, money and trust you don’t get back easily, if ever.”
AI governance resource
Curious about practical tools to manage model health and operation? Model cards provide transparency into an AI model's training data, development process, accuracy, model drift, fairness assessments and governance.
A step-by-step approach to governance
Trust represents the bedrock of the insurance industry – it takes years to earn and mere moments to lose.
According to Max Samadov, a SAS expert in model governance and risk management, governance needs to be baked into how decisions are made. Doing so ensures that consumers’ trust in the business remains earned and intact.
In the Brewing Curiosity episode, he provides a step-by-step approach for shifting the enterprise mindset on model governance toward an AI-native future.
“First, be honest about where your decisions come from,” he shares. “Be model-driven, but own the truth that adjustments, overrides and judgment happen. That’s fine! If you don’t accept this truth, you create blind spots.”
In a new SAS report, 57% of users expressed significant worries about transparency and explainability. That concern could potentially lead them to override AI outcomes – a reality that represents a competitive advantage for AI-ready organizations.
Further, when human oversight is paired with financial outcomes, the business rationale for judgment becomes crystal clear.
Samadov continues: “Always connect model risk directly to financial consequences. If a model is wrong, what breaks? Capital ratios? Dividend plans? Reinsurance strategy? Too often model risk lives in its own world, separate from balance-sheet reality.”
This lack of line of sight to business results and model decisions can carry catastrophic consequences. Executive-level alignment to model outcomes (good or bad) allows an enterprise to pivot faster and identify issues before irreparable damage occurs.
“Finally, stop relying on heroes,” Samadov concludes. “Many insurers still depend on a handful of people who just know how things really work. That works until they leave, or until things change fast. Real protection means building repeatable processes that don’t leave in someone’s head or their personal spreadsheet.”
Customer story
See how Fidelidade unified its asset and liability management process to create a single source of truth across departments, improving financial oversight and decision-making.
AI is the road ahead
By 2040, every company will be an AI company. And for insurers, the use of AI will be inevitable.
Twenty years ago, companies that did not understand internet search banned it. Today, that decision feels insane. “Googling” is a literal step in many new business underwriting processes.
We will see the same evolution with AI, just much, much faster.
Like traditional internet search, AI will eventually fade into the background. But that does not mean the road ahead is free of peril.
Consider the following outlook: Goldman Sachs forecasts a 24x increase in AI token consumption by 2030 (reaching 120 quadrillion tokens per month globally). We are beginning to see the out-of-control, “always on” use of AI and “token maxing.” As a result, organizations are blowing through their annual AI budgets in just a few months.
Curiously, the average token cost is down (about 80%). Why the difference? Why the out-of-control cost?
The result represents a perfect case study of Jevons Paradox – first coined by British economist William Stanley Jevons while studying coal consumption in the 19th century. He found that efficiency improvements lead to higher demand, driving up costs.
We see the same outcome today. In SAS’ webinar, 2025 Insurance Trends That Matter, Aaron Stout, the Director of Digital Workplace Experience at Nationwide, correctly called this prediction. Those same costs will feed into expenditures that insurers are making today to deploy AI.
Good governance as protection
Deploying capital for AI investments will require higher returns as costs increase. So, predictable and dependable outcomes will not be “nice-to-haves,” they will be critical requirements the business should expect before making such an investment.
That investment should be protected.
If we return to Max Samadov's thoughts on governance, we find that good governance can and does protect investment. Not just for AI, but for all enterprise processes.
“When Excel becomes your system of record for capital calculations and key assumptions, you're not managing risk anymore; you're managing institutional memory. And that's a terrible strategy.”
Samadov believes real transformation comes from drawing a clear line between where (human) thinking happens and where control is required. He concludes: “Automate the boring stuff, so your experts can focus on judgment.”
