A quick search through headlines reveals a range of AI-related disappointments. Consider that 95% of GenAI pilots fail, according to MIT. Amazon’s Kiro agent recently sparked a 13-hour outage by deleting a production environment.

And we can’t forget that the resource and energy strain from a new wave of AI data centers is equivalent to that of NYC and San Diego combined.

So – why is it that AI investments fail? And what can be done about it?

For data-heavy industries like insurance, it turns out that the best route to overcoming AI failures and achieving competitive advantage is hiding right there in the data – still your organization’s most treasured asset. The challenge is to have reliable data that you can trust, access, manage and use effectively – when, where and how you need it.

Let’s look at where we are today with AI and where we’re headed next.

Discover how your organization can use SAS Retrieval Agent Manager

Improve – don’t replace – with AI

As many have learned – and research proves – trying to replace knowledge workers with AI can deliver bad results or unintended outcomes and widen gaps in business processes. In fact, using AI to reduce headcount has been called the least imaginative use case. As we learned in an IDC report, commissioned by SAS, “cost reduction” is the least substantiated reason for investing in AI.

The top area for ROI is actually customer service. Improvements in this area make customers stickier and drive policy-in-force (PIF) growth – leading us back to an insurance industry example.

The insurance claims process

Insurance adjusters ingest a multitude of information in the process of settling a claim. That’s especially true in times of increased loss frequency, which drives more incoming claims (like during a catastrophe).

During these times, insurance staff are overwhelmed by the number of losses they must review and adjust. Documents, receipts, interviews, third-party reports and medical information pour in every single day.

Instead of using AI to reduce headcount, insurers can incorporate AI capabilities to improve the loss-adjusting process and accelerate customer value.

Data’s the hardest part

Insurers have treasure troves of data in their underwriting files, claims notes and customer service logs. This data – the fuel for AI – tells an important story and represents a unique relationship with each customer the insurer protects.

But insurance data is varied and often scattered across multiple silos in the business. A recent study showed that 51% of insurers surveyed lack a centralized or optimized data foundation. Carriers that have not taken the steps needed to address their data foundations before using AI face significant roadblocks.

Clean, centralized and well-labeled data – backed by solid data management practices and supported by AI governance frameworks and ethical oversight – deliver trustworthy AI outputs. Such outputs are accurate, auditable and aligned with organizational standards.

Lacking this type of data foundation, many AI implementations will fail.

Data overload: The role of computer vision and GenAI

It takes a lot of time to sort through the vast amounts of information involved in claims to understand things like why a customer purchased a certain insurance product, experienced a loss or had a certain question.

With computer vision, a form of artificial intelligence that allows machines to interpret and understand the visual world (including many types of documents), tens of thousands of documents can be ingested rapidly. This speeds and improves the claims process in several ways.

• Computer vision can help determine if a document, receipt or image was forged (through GenAI).
• It can read the information and extract it into a usable format.

For example, the adjuster does not have to review a roof repair or replacement estimate if the document comes in for automated review, is cleared as “not fake,” and contains the appropriate information to settle the claim. In these cases, an AI model can score the outcomes and pass the information to the adjuster to make a final decision – no time-consuming document reviews required.

Spotting fake documents is more important than ever in the insurance industry. Research suggests up to 25% of insurance claims now involve fake documents. But claims values being targeted are often lower amounts (around $2,500) – because lower amounts are typically handled by less experienced claims handlers or bypassed altogether.

The AI boost from retrieval augmented generation

Normal text models, like large language models (LLMs), are trained on general-purpose information, not enterprise data. Because of that, AI agents using LLMs to answer questions can and do fall short.

But LLMs can learn and ingest enterprise information to improve agentic outcomes. This is where retrieval-augmented generation (RAG) comes into play. The RAG method combines two AI capabilities – retrieval and generation – to strengthen the quality of AI outputs.

Harvesting enterprise data and “fine-tuning” LLM outputs is how RAG improves outcomes. Rather than relying solely on pretrained AI models, RAG pairs semantic search with LLMs to retrieve relevant information from unstructured data. Due to this capability, RAG is now a cornerstone of agentic AI for organizations looking to transform insights into action.

The promise of agentic AI for insurance – outcomes that outperform

Every insurer has millions of customer service logs with information from customers who contacted them to inquire about billing, discounts or claims status.

Using SAS Retrieval Agent Manager, insurers can extract knowledge trapped in tens of thousands of unstructured documents and turn it into usable data for the LLM, training the model to respond to questions at a “segmentation-of-one” level of granularity. This capability pushes insurers beyond simple chatbot interfaces to truly agentic capabilities.

Consider a simple question: “Can you tell me the status of my claim?” This question can be answered in real time with real information. But with agentic capabilities, the answer goes beyond a passive response to trusted advice.

Imagine this answer: "Your claim has been approved by our adjusting team. But I see you haven’t used your rental coverage to secure a vehicle while your repair is made in our blue-ribbon repair network. Here are three rental options in your area. May I schedule your appointment to pick up a rental vehicle?"

A business of trust

For insurance, trust takes years to earn and just moments to break. Breaches in trust and the downside risk of bad decisions can be devastating – and costly. And the application of technologies like machine learning, generative AI or agentic AI can exponentially multiply unintended customer outcomes, like requested cancellations due to poor experiences.

Accenture estimates the impact of poor decisions (claims experiences) at approximately $170 billion over a five-year period, and the underwriting impact at $160 billion. This customer churn prediction was based on a survey of 6,700 customers across 25 countries who reported dissatisfaction with their claims experience.

Moving to an agentic enterprise

In the SAS report, we found that insurers approach agentic AI in what could be called a “wait and see” mode.

The results showed that 43% of insurers fell into one of two quadrants:

• Underutilization – which means low trust in reliable AI systems.
• Overreliance – which means high trust in unproven systems.

Only 9% fell in the ideal quadrant of high trust and high capability, and only 7% described themselves as transformative. But more than half (52%) have made AI training or reskilling their top priority – notable in light of the industry’s looming “silver tsunami.”

Clearly, bad outcomes will happen if insurers fail to effectively adopt and deploy AI technology. But that should not compel insurers to shy away from this transformative technology. Instead, they should continue with a careful approach to adoption.

For example, insurers could fill the talent gap by repositioning insurance careers so that employees are no longer underwriters, claims professionals or sales reps. Instead, they could become agent supervisors facilitating the functions of the entire insurance value chain via underwriting agents, claims agents and sales agents (no pun intended).

These new employees could have backgrounds in change management, technology or policy ethics, legal issues or data science. And they will manage a billion-dollar portfolio in a four-day work week.

When you move beyond the limitations of an automation mindset, you begin to see the possibilities of agentic AI. It’s not a faster, cheaper insurance model. It’s a new type of ecosystem that enables serving customers for life across the entire value chain.

Watch a demo to see how agentic AI can boost claims processes using SAS Viya.