Optimized Satisficing: Why AI Makes Legacy Insurance Better at the Wrong Things
Between October and January, Anthropic signed enterprise partnerships covering more than one million employees across Deloitte, Cognizant, Accenture, and Allianz. In the same period, Accenture announced near-identical AI partnerships with both Anthropic and OpenAI, eight days apart, equipping tens of thousands of consultants with competing AI backends and parallel implementation playbooks. AIG put Anthropic’s CEO and Palantir’s CEO on stage together at its Investor Day, then launched a Lloyd’s syndicate with Palantir in December. State Farm signed on as a launch partner for OpenAI. Nationwide committed $1.5 billion in technology investment through 2028, including $100 million annually for AI.
The press releases all say “transform.” Read the fine print, and every one of these deals describes the same thing: giving existing employees access to AI tools that process existing workflows through existing business models, faster. This is not transformation. This is optimization. And the distinction matters – because we have seen this movie before.
The Ecosystem
A supply chain is forming around legacy carrier AI adoption, and each layer tells part of the story.
At the foundation sit the LLM providers, Anthropic and OpenAI, building the models and signing carrier deals. Anthropic now commands 40% of the enterprise LLM market, up from 12% in 2023, according to a December 2025 Menlo Ventures survey of approximately 500 U.S. enterprise decision-makers. (Menlo Ventures is an Anthropic investor, relevant context for evaluating the data.) Above the models sit analytics platforms like Palantir, whose Foundry platform now serves as AIG’s data backbone, accessing more than four million industry data points through an AIG-built ontology. Above that sits the consulting intermediary layer, Accenture, Deloitte, KPMG, PwC, mediating the implementation at senior consultant rates that industry benchmarks place at $250-$500+ per hour for AI-specific work. Then come the platform modernization vendors like Equisoft, embedding Claude directly into life insurance systems. And at the top sit the carriers themselves, Allianz, AIG, State Farm, Nationwide, Zurich, buying all of it.
Each layer extracts rent. None of them is transforming the underlying business model. Together they represent an enormous infrastructure of technology companies selling to legacy carriers, not competing with them.
The Picks and Shovels
That last point deserves emphasis, because it reveals something important about where insurance stands in the technology cycle.
Amazon did not partner with Borders. Google did not sell search tools to newspaper classified departments. Netflix did not consult for Blockbuster. They competed directly and won. The technology companies driving disruption in retail, media, and entertainment built alternatives that replaced incumbents. In insurance, the pattern reversed for the giants. Not long ago, the industry fear was that Google, Amazon, or another technology giant would enter insurance directly and disrupt carriers the way they disrupted retail and media. That did not happen. Some of the largest technology companies in the world looked at insurance and chose to be vendors. Anthropic, OpenAI, Palantir, and Accenture are all content to sell tools to legacy carriers rather than build competing insurance products. They view insurance as a customer base, not a market to transform.
The venture capital flowing into insurtech tells the same story. As I documented in Dot-Com Survivors and the Insurtech Parallel, most of the investment capital in insurance technology is going to B2B infrastructure: companies that will sell tools and platforms to carriers, not companies building full-stack alternatives that compete with them. Most of the capital structure of insurance technology, corporate and venture, has chosen the picks-and-shovels position.
The picks-and-shovels position is familiar from the dot-com era. Sun Microsystems sold servers to internet startups and to the legacy companies those startups were about to disrupt. Sun’s slogan was “We put the dot in dot-com,” but they also put servers in newspaper data centers. The hardware did not care who won. The analogy here is not to Anthropic’s technology, Claude is a fundamentally different kind of product than a SPARC server, but to the role the vendor ecosystem is playing. Infrastructure providers sell to everyone digging. They do not pick sides in the transformation battle.
Accenture’s dual partnerships make this explicit: OpenAI on December 1, Anthropic on December 9, selling the same consulting methodology with two different AI backends to whoever will buy. When the consulting layer is model-agnostic, the product is not the model. The product is the consulting fee. And the carriers buying it, whether they are building internally, hiring Accenture to build, or purchasing off the shelf, are still building on top of legacy systems. Allianz’s thousands of developers now have access to Claude Code. They are using it to write better code on the same architecture. The tool is powerful. The foundation it is being applied to is not new.
Optimized Satisficing
Satisficing, Herbert Simon’s term for choosing “good enough” over optimal, is the structural condition of legacy insurance. As I explored in Why Good Enough Is Killing Insurance for Carrier Management, carriers optimize for acceptable profits plus internal stability rather than pursuing the harder path of genuine transformation. AI does not change this dynamic. AI accelerates it.
Here is the structural reason. Large language models are trained on existing data. They are extraordinary at pattern recognition, finding regularities in the corpus they were trained on and applying those patterns at scale. Researchers describe them as “statistical models of language” that “rely on memorized patterns over genuine abstraction.” A landmark study on LLM idea generation found that “no evaluations have shown that LLM systems can take the very first step of producing novel, expert-level ideas.” The novelty LLMs produce is bounded recombination of known elements; impressive, but not the same as generating approaches that do not exist in the training data.
Deploy Claude at a legacy carrier and it does exactly what it was designed to do: optimize existing patterns. It processes the same submissions through the same underwriting model, faster. It extracts data from the same document formats with fewer errors. It matches risks against the same historical loss experience more efficiently. All real improvements. None of them transformative. AIG CEO Peter Zaffino described Anthropic’s impact in a TIME interview as allowing underwriters to get “quality insights from data in a fraction of the time” – and that is precisely the point. The insights come faster. The paradigm does not change.
A taxonomy of real-world LLM business model transformations found that actual deployments fall into “efficiency gains, service enhancement, and product extension,” not creation of fundamentally new business models. This is optimized satisficing: using the most powerful pattern recognition technology ever created to get better at the thing you already do, when the question is whether the thing you already do is the right thing to be doing at all.
The AIG Case Study
AIG’s technology partnership history, documented entirely from public record, illustrates this pattern across three consecutive eras.
September 2016: AIG, Hamilton Insurance Group, and Two Sigma launched Attune, a “data-enabled platform” targeting the $80 billion U.S. SME commercial insurance market. AIG CEO Peter Hancock called it “an important way forward for the insurance industry as it adapts to the disruptive forces of data analytics.” Hamilton CEO Brian Duperreault said Attune had “the potential to transform underwriting.” In February 2017, they hired OnDeck’s COO as Attune CEO. Duperreault called it “a perfect fit.” By May 2017, Duperreault, now AIG’s CEO, having moved from Hamilton, expanded the partnership and declared that “cross-industry partnerships, what’s now called insurtech, are the way to go” as the companies worked “to transform our industry.”
October 2021: AIG, Hamilton, and Two Sigma sold Attune to Coalition, a cyber insurance startup. All-cash transaction, terms undisclosed. Five years of press releases about “transforming the industry,” and they divested the entire thing.
April 2025: AIG’s Investor Day featured Anthropic CEO Dario Amodei and Palantir CEO Alex Karp alongside Zaffino. Zaffino, as quoted on Anthropic’s website, now declared their Anthropic partnership “will fundamentally transform how we approach underwriting at scale.” In December 2025, Karp praised AIG’s “forward-thinking approach to revolutionizing the insurance sector.”
The language was recycled almost verbatim. “Transform underwriting” in 2016 and 2017. “Fundamentally transform how we approach underwriting” in 2025. “Revolutionizing the insurance sector” from a new partner in the same breath. Meanwhile, AIG lost more than $30 billion in underwriting from 2008 to 2018 and invested more than $1 billion in data technology over five years.
The logic is simple. If the Two Sigma partnership had delivered transformative results, there would be no Palantir partnership. If the Palantir partnership were sufficient, Anthropic would not need to be on stage at Investor Day. The technology changes. The organizational constraints do not. That is satisficing: cycling through technology partnerships because each one is organizationally acceptable, even though none of them addresses the structural question.
The Cost Structure Tells the Story
The enterprise AI ROI data is convergent and more nuanced than the headlines suggest. MIT’s Project NANDA, a July 2025 study reviewing more than 300 AI initiatives across 52 organizations, found that 95% of enterprises are seeing zero measurable bottom-line impact from AI investment. McKinsey’s 2025 Global Survey found more than 80% of organizations report no tangible EBIT impact. These are midterm grades, not final verdicts. The adoption curve is early, and the results will evolve. But the pattern in the data already points somewhere specific.
BCG’s October 2025 research is the finding that matters most. Sixty percent of firms are generating “hardly any material value” from AI. But 5%, BCG calls them “future-built,” are producing 5x the revenue increases and 3x the cost reductions of everyone else. AI is not failing. AI is failing at organizations that deploy it as optimization of existing processes. The 5% that built their architecture around what AI makes possible are pulling away. The 95% bolting AI onto legacy workflows are not going to close that gap by doing more of the same thing: more consultants, more middleware, another partnership announcement. The deployment model is the variable, not the technology.
The MIT finding reinforces this. The 95% figure represents zero measurable P&L impact, not that the tools do not work. The report explicitly notes that AI tools “primarily enhance individual productivity.” The failure is in converting individual productivity gains into enterprise-level financial impact. This is the optimized satisficing thesis stated in research terms: the technology works exactly as designed. The organization cannot translate operational improvement into strategic transformation. And more spending will not solve an architecture problem.
For a carrier like Allianz or AIG, “deploying AI” means enterprise LLM licensing at 156,000-employee scale, consulting intermediaries billing at $300–$500+ per hour over multi-year engagements, analytics platforms like Palantir, cloud infrastructure, internal training and change management, governance and compliance systems logging every interaction, and ongoing legacy system integration through middleware. Seventy-six percent of AI use cases are now purchased rather than built internally, up from 53% in 2024, and 71% of U.S. insurers cite difficulties integrating AI with legacy systems. The consulting-mediated, bolt-on deployment model is not incidental. It is the primary delivery mechanism and the ROI data measures its results.
What This Actually Means
This analysis is not a critique of AI technology. Claude, GPT, and the models behind them are genuinely powerful. It is not a critique of Anthropic, Palantir, or Accenture. They are selling valuable products to willing buyers, exactly as Sun sold servers to newspapers and startups alike. And it is not an argument that these carriers are foolish. Regulatory barriers, capital requirements, and the fundamental complexity of underwriting risk have prevented pure technology companies from disrupting insurance from the outside, which is precisely why many companies in the technology ecosystem are selling to carriers rather than competing with them.
The argument is narrower and more specific. Deploying AI through the legacy carrier ecosystem, however impressive the technology, however large the investment, produces optimization of existing patterns, not transformation of the business model. The press releases say “transform.” The fine print says “faster processing of the same submissions through the same model.” The gap between the rhetoric and the reality is the story. And the early ROI data, BCG’s 5% pulling away from the other 95%, suggests the gap will widen, not close. The differentiator is not who spent the most on AI. It is who built the architecture that lets AI do what it is actually capable of.
The carriers deploying Claude to 156,000 employees, hiring Accenture to manage it, running it through Palantir, and training staff through vendor-led workshops believe they are transforming. They are optimizing. And as anyone who watched newspapers build websites in 1998 can tell you, confusing the two has consequences that take years to become visible – and by then, the organizations that built differently have already won.
This article was originally published on the Indenseo blog at indenseo.com/blog.
Author Note: This analysis draws on publicly available academic research, industry data, and regulatory filings. Statistics are cited to primary sources where available.
AI Disclosure: Research compilation utilized AI tools to discover and verify publicly available data sources and citations. All analysis, interpretation, and conclusions are original work.