AI & Business Purpose: Rethinking Value Creation
This article addresses the Business Purpose hurdle in my Moving Forward with AI article.
What do we adopt AI to do?
Allianz set out to address a specific problem: how to scale AI across a regulated insurer without losing trust. It chose Anthropic as the strategic partner for its focus on safety and transparency that complements Allianz's dedication to customer excellence and stakeholder trust. The choice was made on business purpose, not technology.
Since the ChatGPT launch in 2022, model capability has advanced in reasoning quality and multi-step reliability. AIOps and tooling are maturing, and data infrastructure is catching up. However, reacting to competitive anxiety and investor pressure, many organizations moved quickly to launch AI programs and pilots without a well-defined business purpose. The real advantages of AI: data-to-information integration, judgment at scale, and mass personalization – remain largely under-harnessed.
The primary barriers to scale are now structural and organizational, not technical.
Leading organizations do not simply deploy AI to reduce costs from existing structures. Rather, they are asking: Given what AI now makes possible, what should our products be, who do we serve, and how do we deliver?
This primer* invites business and technology leaders to come together to redirect AI conversations back to business purpose and value creation as the foundation for strategic decisions that shape AI adoption. We will explore:
I. How AI creates value – promises and realities
II. Structural constraints and strategic choices
III. Rethinking ROI for AI investment
IV. Purpose and strategic choices for leaders
The examples here are mainly drawn from insurance where the structural constraints are especially visible, but banks and asset managers face the same choices.
* Many thanks to Jamie Cattell, Paul DeSilva, Sid Dixit, Maxine Goddard, Andy Jones, Jae Kang, Bill Lewis, Che-Yuan Li, Timo Loescher, Bill Martin, Russell Page, Robert Pick, Dominique Roudaut, Darren Sharp, James Shepherd, Bill Snider, Andy Watts, and several anonymous contributors for generously sharing their perspectives. Contributors participated in a personal capacity. Their views do not necessarily reflect those of the organizations with which they are affiliated. Any errors are my own.
I. How AI creates value – promises and realities
Artificial intelligence (AI), as a suite of technologies including predictive models and generative AI with multimodal capability across natural language, vision, and voice, can create value beyond traditional computing. It integrates information faster, applies judgment more consistently across a volume of decisions that once forced a tradeoff between speed and accuracy, and enables mass personalization that was previously uneconomical (though consumer reception is still evolving). These capabilities make new products and propositions that serve customers better without lengthy and costly core system rebuilds.
“What makes these things possible is the same thing that makes any startup a success — an idea executed on and brought to market that delights the customer so much, that the company offering it makes more money than those who don't.” – Bill Martin
While most AI projects focus on productivity and cost reduction at the tool level, optimizing inside the existing structure, the differentiating value is often realized at the structural layer of operating models, data, and organizational topology (i.e., how information moves and how decisions are made). In claims, for example, the ability to ingest large volumes of incoming data and detect anomalies, patterns, and exceptions beyond what humans alone can identify represents a genuinely different category of capability, not just augmentation of existing work. A hybrid human-machine model, with AI handling high-volume data processing and pattern recognition while underwriters and adjusters concentrate on complex risk selection and claim adjudication, can result in loss ratios that legacy structures would not be able to achieve.
Value above productivity can only be captured when three conditions exist.
The firm has proprietary data others cannot replicate, such as the submission and claims history a carrier has accumulated over decades.
It has a business purpose broad enough to use that data across internal boundaries, so that what underwriting knows informs claims, and what claims learns informs pricing.
It has the architectural discipline to build data and AI as a shared capability layer rather than a bolt-on to applications, since connecting AI to systems is not the same as making that data usable.
Absent any one, AI merely runs the existing workflow faster.
Adopting AI, however, is more than introducing technological capability. Attempting to recreate processes and systems that are already working by introducing some “magic AI platform” without articulating how the new process differentiates the business leads to failed pilots, and worse, opens regulatory gaps and safety and governance concerns.
Early adopters of point solutions have paid the first-mover’s tax, discovering impediments related to governance, cost, maintainability, and a lack of skills and experience, but they gained experience before the rest of the market. Most firms adopt AI for productivity and cost, which is a reasonable place to start if it is a deliberate choice for measurable, near-term return.
For the insurance industry, AI offers new tools to address the protection gap. Swiss Re puts the global natural catastrophe protection gap at USD 424 billion in 2025 and the global mortality protection gap at a record USD 432 billion. In the U.S., Hiscox finds 77% of small businesses underinsured. The gap persists because the structures and distribution built for historical constraints cannot cover these risks. However, using AI only to price more precisely may end up declining more risks. Using AI to lower the cost of serving a customer, on the other hand, can cover risks that the existing structure ignored, through new distribution and service models such as parametric coverage for extreme climate. This is a choice of strategy and purpose, enabled by technology.
The most underappreciated value AI delivers is cognitive. Beyond deterministic AI improving accuracy and consistency and generative AI expanding reasoning and simplifying institutional knowledge transfer, AI changes how people think about problems. This strengthens workforce capability over time but does not show up in efficiency metrics.ROI does not capture the analyst who asked a better question because AI pushed the inquiry further, or the underwriter who identified a risk pattern they would not have seen without AI.
A fundamental source of future value is harder to predict. It comes not from employees using AI as a productivity tool, but from what they build once they become fluent: reusable assets built through internal communities, and proprietary applications unique to the firm that invests in AI fluency.
Ultimately, given the structural changes required to realize differentiating value, clarity on the organization’s structures and business strategy, and its real capability to execute, should guide AI adoption decisions rather than popular use cases or best practices.
Let’s take a closer look at structural constraints and the choices leaders should consider as they shape AI strategy for their organizations.
II. Structural constraints and strategic choices
Today’s operations were shaped by technology waves from paper to rule-based computing, to relational databases and enterprise software. Each wave expanded what was possible, but also accumulated legacy constraints that sit in operating models, processes, workflows, handoffs, and role boundaries on the org charts. In insurance, they are found in centralized underwriting authority, submission queues, segment-level products, and distribution built for high-premium policies – the right way to organize work when information, judgment, and personalization at scale were expensive. Now that recent technological advances, particularly AI, can perform those activities cheaper and faster, keeping the existing structures becomes a strategic choice.
For example, carriers have been automating submission intake for years, and it continues to be a bottleneck. An underwriter can analyze and process only so many submissions within the current structure. When a broker cannot wait and moves a good risk to another carrier, the revenue is lost. To address this, leadership can adopt AI to integrate information for underwriters to analyze submissions faster. They can also choose to ask the harder questions: should submissions continue to be handled through a single underwriter at all? Should AI triage and price, with humans auditing autonomous decisions? Who owns the decision when an AI-triaged submission binds a bad risk?
Most incumbents default to making the existing structure faster. That delivers a respectable return and protects current margins without changing authority. Clayton Christensen named this asymmetric motivation: organizational gravity keeps incumbents from redirecting resources toward structural questions and rarely rewards those who explore the fringes. A defensible ROI from the faster process prevents the questioning of the existing structure, especially when the executive who could authorize the structural move has to personally absorb the downside (think “the model decided” in front of a regulator) but captures none of the upside. This has allowed leaner entrants with lower cost-to-serve, such as digital-first MGAs in insurance and neobanks in banking, to grow business where the incumbent’s structure cannot deliver profitably.
Questioning structure means questioning budgets and authority, which makes changing the existing structure a leadership challenge involving hard choices. Let’s explore some of those related to AI adoption.
One choice is architecture strategy. As general-purpose computing capabilities are increasingly commoditized, the choice is about control of proprietary data, knowledge, and workflows that differentiate. What is the core to build for competitive advantage? When outsourcing, what should the terms on ownership of data and workflows look like? Should the organization train a domain-specific model on its proprietary data, or rely on a general-purpose model fitted with retrieval and context tools such as RAG for a specific purpose, which can be as powerful at more sustainable cost? Some may even ask, will databases become the core?
In reality, AI-native core platforms barely exist yet. Oftentimes, enterprise data platforms are so embedded in existing systems that AI solutions end up bolted on rather than integrated into a foundation designed for AI. Any build either strengthens a reusable architecture or creates another silo that later needs to be integrated, governed, and maintained.
In practice, a strangler approach works well: route new AI-enabled value-creating flows around the legacy core incrementally until the core can be retired rather than attempting a wholesale rebuild, which is far higher risk and politically harder to sustain.
Another is strategy ownership. Should a chief AI officer be added to the org chart to centralize AI strategy on how decisions get made, who owns what questions, and how AI capability is built and distributed across the organization? Or, should the C-suite jointly own the AI strategy and execution outcome?
The third is IT’s role. The traditional software development cycle, from lengthy cost/benefit analysis, estimate, approval, to requirements, QA testing, and UAT, is being compressed by AI. Should underwriting and operations continue to rely on IT for workflow changes, or experiment with using AI to build decision logic to initiate, analyze, score, and automate routine judgment as much as possible without waiting for IT?
These are strategic choices that have consequences, and may play out more quickly now that the infrastructure and capabilities have been democratized. Nokia is a cautionary case. It saw the smartphone coming and had the talent to build one, but could not change a structure built around hardware and release cycles measured in years, and failed to adapt its human capital model quickly enough.
Allianz made the choice to change. It saw AI’s potential and chose to scale AI through its partnership with Anthropic rather than defending its existing structure. It chose to enable employees through tools like Claude Code to build agentic automation of multi-step processes such as claims handling in motor and health insurance with employees still overseeing nuanced cases, and to document each AI-driven decision and its basis to meet regulatory expectations. Claude Code is already used by thousands of Allianz developers worldwide.
"The organizations that win will be the ones that treat AI as a reason to redesign the system, not as a faster way to run the one they already have." – Russell Page
Every choice is also an investment decision. How is the existing ROI framework serving AI investments?
III. Rethinking ROI for AI investment
AI spending is climbing across the industry, but few AI initiatives have focused on new growth. In the **Evident AI Index for Insurance 2026, across 65 disclosed use cases with measured outcomes, productivity gains appear in 75% of them while revenue uplift appears in just 2%.
How is the conventional ROI framework serving AI investment?
Traditional ROI rewards value that is measurable and self-contained, with clean cash flows. It is useful to approve a point solution that can show a strong return, especially when it can pay for itself, but it is biased toward optimizing efficiency within the existing structure.
For AI investment, an enabler platform and enriched data foundation can show no standalone ROI while structurally improving the cost-to-serve curve for every workflow that later runs on it.
ROI can also undervalue gains that typically do not show up in near-term returns, such as risk reduction from fraud detection, pricing accuracy, and the value of a relationship built through better service.
Double-counting is also more acute for AI. The data foundation project and the AI-agent project can both claim the same saving, and both are required to capture the benefit.
A pricing gain through a solution provider can easily be accessed across the industry and becomes the new baseline, eroding the initial advantage. The more durable return is one built on proprietary data and workflow that the organization has chosen to own and can reuse as shared capability.
These challenges are not unique to AI. Conventional ROI has always struggled with investments whose benefits span functions or whose advantage erodes as competitors catch up. AI’s characteristics, however, challenge the assumptions the traditional framework is built on: cost is not fixed at approval, and value shows up downstream rather than in the project that funded it.
How should organizations overcome these limitations?
Start with where value is. For AI, it is more useful to fund enabler platforms as infrastructure, evaluated against unit-economics improvement across the P&L area and against optionality, and fund the use cases that run on top of them with conventional ROI. Unit economics asks whether AI moves the levers a P&L owner already manages: contribution margin per policy, cost to serve per customer, cost to acquire per new relationship, across functions rather than within a single use case.
Invest in knowing the true cost that measuring unit economics depends on. Many firms lack a robust total cost of ownership technology management process and an accounting structure that captures AI’s true direct and indirect costs, especially when involving HR costs and allocation across functions, which is often negotiated.
AI adds to the cost transparency challenge. Token and inference charges are difficult to estimate at approval, and vendor pricing is also shifting from flat subscriptions to usage-based billing.
Because of these costing challenges, there has been a shift toward high-level portfolio management with measures such as tech cost, ops cost, or total cost over revenue, and return on equity (ROE) along P&L areas. This should operate in parallel with conventional ROI where a clear payback is required for a single project, and a “kill switch” should be triggered for a deeper review when cost, timeline, or outcome change materially from the approved plan.
"AI is a process, not a goal.” – Robert Pick
Given AI adoption is still early, an investment in a program of pilots can systematically allow experimenting and learning to inform long-term strategic direction. Each pilot can be constructed to test different uses of the technology and different types of use cases. The goal is to identify the real capability and limitations of the technology and of the organization. This program should be built with a planned failure rate that allows failures to reveal those limitations while keeping the program viable.
Funding the right investment is the start. The return comes only when it is delivered. Approving what the organization's internal and external capability can actually execute is the discipline that turns a funded decision into a realized return.
IV. Purpose and strategic choices for leaders
Purpose is the result of decisions leaders make and own, and most AI adoption decisions are hard choices that question historical decisions and challenge existing authority. They are harder still because the leaders who must drive those changes often do not have the bandwidth to think and act strategically.
AI adoption needs a reliable and scalable data foundation, strong governance, and knowledgeable and engaged employees. Each involves hard choices and complex work. When the work gets tough, a clear purpose that creates genuine value is what can keep everyone focused and moving forward.
Leaders make these hard choices.
Do we optimize the structures we have built and defend our margins, or pursue growth that requires us to remove current constraints? Are we willing to take the risk and pursue that new growth?
Are we building AI capabilities that strengthen our advantage and buying those that do not? What data and workflows give us a unique advantage, and are we harnessing those?
Are we making choices around what AI can do today, or planning for what it will enable next?
Do we fund the foundation as infrastructure, or expect the same return as the features?
Can we execute what we approve?
These questions are best answered through deliberate work with clarity and focus rather than under pressure and best driven by business outcomes rather than technology decisions.
Given what AI now makes possible, what should our products be, who do we serve, and how do we deliver?
AI has changed work. The leaders who choose deliberately, with purpose, will shape what their business becomes.
**Source: Evident AI Index for Insurance 2026, Key Findings Report. Evident - [link]
* Ichun Lai founded Propel Global Advisory LLC to accelerate purposeful and responsible AI adoption in financial services. She serves as a trusted advisor for leaders who understand AI matters but don’t have the bandwidth to dive in and work through cross-functional and cross-cultural complexity strategically to drive tangible outcomes.