Saturday, 5 September 2026
Artificial intelligence can perform an expanding range of insurance tasks, but that does not mean insurers should use it for everything. As carriers move from experimentation towards large-scale deployment, the more important question is becoming whether AI is actually the cheapest, most reliable and most defensible technology for the job. That was the central argument of a presentation at AI4 2026 in Las Vegas examining the economics of AI adoption in insurance. Rather than focusing on what the latest models can technically accomplish, the presentation challenged insurers to distinguish between work that should remain conventional software, work where AI can create genuine value and work where human judgement and relationships continue to provide the greatest competitive advantage.
The starting point was deliberately simple. The speaker compared the cost of asking leading AI models to solve a basic arithmetic problem with the cost of performing the same calculation using conventional software. All produced the correct answer, but the AI models were considerably more expensive. The exact cost multiple depends heavily on model choice, token usage, infrastructure and implementation, but the broader point is more important than the individual calculation: using a large language model for a deterministic task can be economically irrational even when the model performs it perfectly.
That distinction becomes significant in insurance because carriers process enormous numbers of transactions. Eligibility checks, coverage validations, rating calculations and many other routine processes operate at volumes where a small difference in the cost of an individual transaction can become material when multiplied across millions of policies or claims. The implication is that the industry’s growing enthusiasm for autonomous agents needs to be balanced against unit economics. A system that is inexpensive to demonstrate can become considerably more costly when deployed across a large insurance operation. The correct question is therefore not simply whether an AI agent can perform a task, but whether it should.
This represents an important change in the enterprise AI debate. During the initial generative AI boom, organisations concentrated heavily on technical capability and employee adoption. Companies measured access to AI assistants, token consumption and the number of pilots underway. By 2026, management teams are placing greater emphasis on measurable financial returns. A widely discussed MIT study published in 2025 found that the large majority of organisations examined were not yet generating measurable returns from their generative AI investments. The finding was sometimes interpreted as evidence that AI itself was failing, but subsequent discussion has focused more heavily on the mismatch between technology and business processes. The problem is frequently not whether AI works, but whether it has been applied to the correct problem and integrated into the organisation in a way that creates economic value.
Insurance provides an unusually clear environment in which to examine that problem because the industry’s work ranges from highly deterministic calculations to decisions that depend heavily on professional judgement. At one end are tasks where a given input should consistently produce the same output. At the other are activities such as complex commercial underwriting, difficult claims negotiations and broker relationships, where experience and interpretation can materially influence the outcome.
The presentation argued that insurers should examine work along this spectrum before deciding which technology to deploy. Deterministic processes that can be expressed through clear rules are generally better suited to traditional software, particularly when they operate at significant scale. Conventional code is predictable, comparatively inexpensive to execute and easier to audit because the same input should produce the same result. The problem is that much of this codifiable work has never been fully automated. Insurance companies have spent decades building core systems, yet employees still perform substantial amounts of rules-based work manually because information arrives in formats those systems were never designed to process.
Broker submissions may arrive as PDFs, spreadsheets, emails, scanned documents and attachments using inconsistent structures. Employees therefore become the bridge between unstructured information and highly structured insurance systems. This is one of the areas where modern AI can have an important role without necessarily becoming the system that ultimately makes the decision. AI can read and interpret unstructured material and convert it into a format traditional software can process. It can also help software engineers build conventional applications considerably faster.
The resulting architecture can therefore combine both technologies. AI handles ambiguity at the edge of the process, while deterministic software executes rules once the information has been structured. That approach can provide many of the advantages of generative AI without requiring an expensive probabilistic model to perform calculations that established software can complete reliably. AI-assisted software development could also accelerate this transition. Insurance companies historically avoided automating some smaller or specialised processes because the development cost could not be justified. If AI coding tools substantially reduce the cost and time required to build conventional applications, some of these previously uneconomic automation projects could become viable.
This creates a different way of thinking about generative AI investment. Instead of replacing traditional software, AI may help companies create more of it. The second category consists of cognitive work that conventional systems have struggled to perform. Commercial insurance submissions illustrate the opportunity. An underwriter may receive dozens of pages containing financial statements, previous losses, property information, policy details and correspondence containing important context. Reading and interpreting the submission can require a significant amount of professional time before the underwriter begins making the actual risk decision.
Large language models are well suited to this type of task because they can extract information, summarise documents, compare different sources and identify relevant passages from unstructured material. AI can therefore operate as an assistant that gives underwriters more time for the work where their judgement creates value. The potential productivity improvement can be substantial. An underwriter capable of reviewing considerably more submissions could increase the insurer’s capacity without proportionately increasing administrative resources. Alternatively, the underwriter could maintain the same workload while spending more time understanding difficult risks, negotiating terms or developing better pricing.
The distinction matters because reducing processing time is not necessarily the same as creating competitive advantage. If every insurer eventually uses similar document-processing tools, faster submission handling simply becomes the expected market standard. The more durable advantage comes from what professionals do with the time and information the technology provides. That could include more sophisticated underwriting, development of new insurance products or the ability to enter risk categories competitors find difficult to price. In that scenario, AI does not generate the competitive advantage by itself. It releases human capacity that the insurer can redeploy into activities that create differentiation.
This becomes particularly relevant as insurers accumulate larger volumes of historical information. Many companies possess decades of claims files, underwriting documents and risk data that have never been fully analysed because much of the information remains trapped in unstructured records. AI can potentially make that material accessible to actuaries and underwriters, allowing them to identify patterns that were previously difficult to extract. Better use of historical data could lead to improved risk models and more precise pricing. Even a relatively short-lived advantage in actuarial understanding can have significant consequences because underwriting decisions compound through future renewal cycles and claims experience.
However, the same characteristics that make large language models powerful also create problems when they are used in deterministic processes. AI models are probabilistic. Their outputs can vary, and performance can change when the underlying model is updated. An application that produces acceptable results using one model version cannot automatically be assumed to behave identically after an upgrade. This creates a reliability problem for insurers. A consumer application can often tolerate occasional variation or an imperfect response. Insurance systems supporting policy decisions, premium calculations or claims require considerably greater consistency.
Carriers therefore need strong testing, monitoring and control environments whenever AI influences consequential processes. Model changes need to be tested against previous behaviour, while systems require clear rules determining when an AI output can proceed automatically and when human review is necessary. Auditability creates another challenge. Conventional software can generally provide a clear record of which rules were executed and which inputs produced an outcome. Generative AI does not naturally provide the same form of deterministic explanation.
This does not mean AI systems cannot be monitored or audited at all. Organisations can record prompts, inputs, model versions, outputs, retrieval sources, intermediate actions and other telemetry. But achieving the level of traceability required for regulated financial processes generally requires additional architecture and governance rather than emerging automatically from the model. That infrastructure also costs money. Logging, evaluation systems, additional model calls, human reviews and testing all increase the real expense of operating AI. The cost of the model itself therefore represents only one component of the total economics.
This is why insurers need to evaluate AI at the level of the complete operating process rather than the price of an individual model call. A system that appears inexpensive during experimentation may require substantial additional controls before it is suitable for production. The third category identified in the presentation is work where human relationships remain central. Claims provide the clearest example. When a policyholder has lost a home, suffered a serious accident or experienced another major event, the interaction with the insurer is not merely an information-processing exercise.
Customers may need explanation, reassurance and someone capable of responding to circumstances that do not fit neatly into a predefined workflow. A technically correct automated answer does not necessarily provide the same experience as a skilled claims professional who can understand the emotional and practical circumstances surrounding the loss. This human element can have commercial value. Insurance is unusual because customers pay for a product they hope never to use. When a major claim finally occurs, the quality of that interaction can determine whether the customer renews, recommends the insurer or leaves.
Automating that relationship purely to reduce operating expenses can therefore destroy value elsewhere in the business. The lowest-cost claims process is not necessarily the most profitable one if it damages retention, reputation or distribution relationships. The more effective model may be to automate everything surrounding the human interaction. AI can gather policy information, summarise claims history, prepare documentation and recommend next steps while allowing the claims professional to concentrate on the policyholder.
The same principle applies to complex underwriting and broker relationships. Technology can provide better information and remove administrative work, but experienced professionals may remain responsible for interpreting unusual risks, negotiating coverage and understanding the commercial circumstances surrounding a client. This leads to a more practical framework for insurers considering AI investment. Instead of beginning with a technology and searching for places to deploy it, companies can begin with an existing business process and break it into individual tasks.
Each task can then be assessed according to its inputs, expected outcome and degree of judgement involved. If the task is deterministic and high-volume, conventional software may remain the best solution. If it involves understanding ambiguous or unstructured information, AI may be appropriate. If success depends heavily on empathy, negotiation or relationships, technology should probably support the employee rather than replace them.
The exercise also requires insurers to involve employees who actually perform the work. Operational processes frequently contain exceptions and informal practices that are poorly understood by senior management or technology teams. Frontline employees know which tasks consume time, which information is difficult to obtain and where judgement genuinely matters. This bottom-up analysis can reveal AI opportunities more effectively than starting with a list of fashionable technologies. It also reduces the risk of automating a process that should have been redesigned or eliminated altogether.
The unit economics become particularly important as agentic AI develops. Autonomous agents can theoretically execute increasingly complex sequences of tasks, but every additional model call consumes computing resources. At modest volume this may appear insignificant. Across millions of insurance transactions, the economics can look very different. Insurers therefore need to distinguish between technical possibility and economic scalability.
Some tasks will justify relatively expensive AI because the value of the decision is high. Spending significantly more computing resources to help an underwriter evaluate a multi-million-dollar commercial risk may be economically reasonable. Using the same architecture to perform elementary calculations millions of times is considerably harder to justify. This is also likely to influence how insurers evaluate technology vendors. Start-ups offering complete AI replacements for existing workflows need to demonstrate not only better performance but sustainable economics once usage reaches enterprise scale.
Carriers will increasingly examine model costs, contractual exposure, reliability, portability and the additional governance required before allowing vendors to become embedded in core processes. The insurance industry’s AI opportunity therefore appears less universal than the early hype suggested, but potentially more valuable. The objective is not to maximise the number of processes touched by AI. It is to identify the places where AI creates an advantage that another technology cannot deliver as effectively.
In some parts of the organisation, that means using AI to turn unstructured information into structured data. Elsewhere, it means helping professionals analyse complex material more quickly. In software development, AI can reduce the cost of building deterministic systems. And in customer-facing situations where trust and empathy matter, the best use of AI may be almost invisible to the policyholder.
This also changes the definition of return on investment. Cutting administrative costs remains useful, but it is only one possible outcome. Better underwriting decisions, more accurate actuarial models, additional capacity, new products and access to previously difficult risks can create considerably greater value. For insurers, the strategic opportunity is therefore not simply greater efficiency. It is the possibility of using technology to redirect skilled employees away from low-value administration and towards areas where judgement generates competitive advantage.
The companies that benefit most from AI may consequently be neither the fastest adopters nor those with the highest model usage. They will be the insurers that understand which parts of their operations should become software, which should become AI-assisted and which should remain fundamentally human. As the technology becomes more capable, making that distinction may become harder rather than easier. AI will increasingly be able to perform tasks that do not necessarily make economic or operational sense for it to perform. The discipline will lie in knowing when not to use it.
For an industry built around evaluating risk, that may ultimately become one of the most important AI decisions of all.
Source: CIJ.World Research & Analysis Team
