The Practical Realities of Today's Generative AI

Few technologies in the last two decades have generated as much anticipation—and as much noise—as Generative AI (AI). In late 2023, the commercial release of ChatGPT ignited a wave of experimentation that swept through nearly every industry, including insurance. By 2024, carrier pilots were underway and a veritable global army of middling chatbots appeared…and then most everyone stopped using them just as quickly. By 2025, many enterprises had released their first limited production use cases. Entering 2026, expectations have escalated again—this time powered by semi-autonomous agents, as well as visions of large-scale cost takeout and step change productivity gains.

The reality on the ground, however, remains far more measured, while still quite promising.

GenAI is proving to be genuinely powerful—but not magical. Value is real, but uneven. While adoption is now widespread, creating a sustained competitive advantage based on AI is still rare, as are truly enterprise-scale bankable AI-based solutions. For leaders making investment decisions today, separating applied reality from aspiration has never been more important.

To understand both the challenges and opportunities surrounding AI, it is critical to avoid the trap of thinking of AI—specifically generative AI-powered solutions—as a “product.” A lot of time and goodwill gets wasted around exactly this point. The way many speak of AI—especially those NOT in the thick of bringing enterprise technology and digital solutions to fruition—treats it as an app, a widget, a “thingy” that is created and exists as would your favorite social media app or tip calculator. But those things are finite collections of features and functions, the development of which has a described beginning, middle, and end. AI—generative AI—is far better described as a set of CAPABILITIES and METHODS, tools even, which can be applied to myriad business and technical challenges in largely unlimited ways, resulting in varying degrees of risk and reward. Put differently, AI is a means, not an end.

Adding to this confusion is the never-ending fount of “AI Product” coming to market, which is to say products, apps, and tools which package or expose AI capabilities and methods for consumption by mere mortals. These are, indeed, products, but they exist to facilitate the means of using AI; they are not AI itself.

This whole distinction becomes critical when assessing investment focus and expected outcomes in areas powered by AI. At this delicate point in the lifecycle of AI adoption at companies, where little exists beyond copilots and groovy chatbots, AI reference points are limited, leaving open a kilometer-wide chasm filled with misunderstanding. To wit: when a CIO or CDO (or even a CAIO) is asked for an “AI Rollout Roadmap,” more often than not a business person has in mind the delivery of AI results—i.e. products or “thingies”—while the CIO/CDO is thinking of the rollout as the enablement of platforms and solutions which present AI capabilities that can be used to create AI-driven outcomes. The CIO is implementing Bedrock or Foundry, but the CEO is thinking of an app or a widget.

Technologists deal with this regularly. We use many solutions—database platforms, data lake solutions, integration suites—that are a similar packaging of capabilities that we then weave into products and solutions for our businesses. All these AI solutions—including emerging and very promising Agentic platforms—fall into a similar bucket. Getting this distinction aired out avoids a lot of hassle.

Understanding AI as capabilities provides a consistent explanation for what appears to be a LACK of progress on implementing AI at scale in an enterprise class form. Acquiring and learning capabilities to then turn them loose to create products and solutions takes time, even where the totality of so-called “AI Analysts” is screeching that our corporate lives depend on being agentic by next Tuesday.

One of the clearest signals cutting through GenAI hype is the persistent gap between experimentation and scale. A widely cited MIT NANDA study released in mid 2025 found that roughly 95% of custom AI pilots either failed outright or failed to deliver expected ROI. While provocative, this result resonated because it aligned with lived enterprise experience: pilots are easy; production is hard.

More recent global surveys reinforce this pattern rather than contradict it. By late 2025, nearly nine in ten organizations reported using AI in at least one business function (McKinsey), yet only a minority had scaled those capabilities broadly across the enterprise. Even for agentic AI—systems designed to reason, plan, and act—which is extraordinarily promising and answers many issues associated with injecting AI in-line into complex transactions, only about one quarter of organizations report scaling, while roughly four in ten remain in experimentation (McKinsey).

The growing body of data clearly shows that adoption metrics alone are misleading. The true bottleneck is not access to models or new tools, but operationalization; namely governance, data readiness, security, integration, and change management, all of which require understanding the complex intersections of these new capabilities.

Another useful reality check comes from examining how frequently AI is used by employees themselves, not just approved by leadership in a stream of capital projects.
A large US workforce survey conducted in late 2025 found that only about 10% of employees use AI tools daily in their jobs, with just 23% using AI even a few times per week (Gallup). Nearly half of respondents reported some level of occasional use, but consistent daily reliance remains limited.

This gap between executive enthusiasm (coupled with the bulging eyes of shrieking analysts) and actual day to day usage highlights an important truth: GenAI’s productivity impact is still highly role dependent and somewhat rare. Knowledge workers in writing, research, and analysis heavy roles see immediate benefits. Many operational, frontline, and—especially—regulated functions see far less.

Despite these constraints, GenAI is in fact already delivering tangible value in specific, repeatable patterns. Across industries, the most successful production deployments cluster around:

• Summarization and synthesis of large document sets
  Intelligent document ingestion and classification
  Drafting assistance for customer communications, reports, and internal documentation
  Workflow orchestration and routing, rather than autonomous decision making

These use cases may not be glamorous, but they scale. They reduce cycle time, lower error rates, and free human capacity without triggering regulatory alarm bells.

Insurance—particularly P&C and Life—illustrates both the promise and limits of applied GenAI.
Multiple post 2025 insurance focused studies, including one by Boston Consulting Group (BCG) converge on the same conclusion: interest is high, experimentation is widespread, but scaled production remains rare. One global insurer study published in 2025 found that only about 7% of insurers had successfully brought AI systems to scale, while roughly two thirds remained stuck in pilot or limited deployment mode.

Where insurers are seeing value is telling. The most common production use cases include:

• Policy and claims document summarization
  Submission intake and triage
  Underwriting workbench augmentation (not replacement)
  Customer correspondence drafting and review
  A host of technology-focused AI uses in coding, QA, and documentation
  
  These actually track well with the cross-industry common use cases noted prior. Notably absent are fully autonomous underwriting or claims adjudication systems, despite many analysts and pontificators holding out these are the path to industry salvation. Regulatory scrutiny, explainability requirements, and data lineage concerns continue to enforce human in the loop designs, and our existing incumbent (don’t call it legacy) architectures make AI injection challenging.
  
  Recent underwriting specific research released by Accenture in late 2025 suggests adoption will continue to grow—but incrementally. In a survey of several hundred underwriting leaders across Life, Commercial, and Personal Lines, respondents estimated that AI augmented underwriting would grow from low teens adoption today to roughly two thirds penetration over the next three years (Accenture). While this may prove to be an overly ambitious estimate, it is certainly directionally correct. But critically, the wording matters: most respondents described augmentation rather than automation—faster risk assessment, better information synthesis, and improved consistency—while final decisions remain human. Augmented, but not yet autonomous.
  
  This aligns with the study’s broader Agentic AI findings. Surveys of senior executives in early 2026 show that roughly half of Agentic AI initiatives remain in pilot, with security, compliance, and operational risk cited as primary blockers. This can be forgiven as Agentic itself is barely one year old at this time…but evolution can happen quickly, and it is. Even among deployed AI systems, the vast majority of AI driven decisions are still reviewed or approved by humans. There is very little that is truly “straight-through.”

Agentic AI is already defining the next chapter of enterprise AI—but not on the timelines implied by vendor marketing. Insurance specific research indicates that only about one fifth of insurers expect to have agentic AI solutions in production by the end of 2026 (Celent), which reflects a pragmatic understanding of what must come first: data quality, model governance, control frameworks, and organizational trust.

In the near term, the most effective agentic patterns resemble guided automation and targeted augmentation, not autonomy—AI systems that propose actions, surface risks, and orchestrate workflows, while humans retain accountability.

GenAI is no longer experimental. It is real, useful, and increasingly embedded in enterprise workflows. But the evidence is equally clear that AI-powered revolutionary transformation in the enterprise is actually evolutionary, not explosive.

For insurers and other regulated industries, the winners over the next several years will not be those chasing headline autonomy, but those quietly scaling durable, well governed, human centered AI capabilities.

The hype cycle may be peaking. The real work—and the real value—is just beginning.

The information in this article is timely as of its writing on 29 January 2026

The following sources are referenced in this article.