Updated On: September 8, 2026. This article was substantially revised following a source review. Unsupported performance claims and incorrectly attributed regulatory references in the original September 2025 version were removed or replaced with verifiable primary and institutional sources.
Key Takeaways
- Artificial intelligence is producing measurable gains in insurance, but the evidence does not support a universal promise of a specific companywide expense-ratio reduction.
- The strongest results are appearing inside defined business domains such as claims, underwriting, distribution, customer onboarding, and servicing.
- A productivity improvement inside a workflow is not the same thing as an improvement in the statutory expense ratio. CFOs should measure both operational gains and the portion that actually reaches the income statement.
- The first 90 days should establish whether a bounded AI use case creates measurable economic value, operates within defined authority, and can be governed well enough to scale.
- Insurance regulators are not prescribing one AI architecture, but expectations around governance, validation, data, third-party oversight, monitoring, and documentation are becoming increasingly explicit.
The Number a CFO Should Not Start With
There is a seductive way to build the financial case for artificial intelligence in insurance: choose a large efficiency percentage, apply it to an expense base, calculate the resulting combined-ratio improvement, and present the board with an apparently inevitable return on investment.
The arithmetic can be flawless while the investment case is wrong.
AI does have the potential to change insurance economics. There is now enough production experience to move that discussion beyond hypothetical use cases. But there is not a credible industry benchmark saying that an insurer should expect 10 percent, 20 percent, or 25 percent companywide expense-ratio relief simply because it adopts AI.
In fact, the broader expense trend makes that framing even less useful. AM Best reported in January 2026 that the U.S. property and casualty industry’s underwriting expense ratio declined 2.4 percentage points between 2014 and 2024. Most of that improvement came from lower acquisition expenses, while commissions and brokerage costs remained relatively stable as a share of net written premium.
AI therefore should not be presented as a rescue plan for an industry whose expense ratio has been moving inexorably in the wrong direction. The more useful question is narrower:
Can AI change the unit economics of a specific insurance process enough to justify the cost, risk, and organizational effort required to deploy it?
That is a question a CFO can begin answering in 90 days.
The Evidence Is Getting Better, but It Needs to Be Read Carefully
The distinction between potential value and realized financial value matters.
In a 2024 McKinsey survey of more than 50 leaders from large European insurance groups, more than half of respondents believed generative AI could eventually produce productivity gains of 10 to 20 percent, premium growth of 1.5 to 3 percent, and improvement in technical results of 1.5 to 3 percentage points.
Those are executive expectations, not audited industry outcomes.
More recent operating evidence is more useful. In 2025, McKinsey reported observed results from domain-level AI transformations that included 20 to 40 percent reductions in customer onboarding costs, 10 to 20 percent improvements in new-agent success and sales conversion measures, 10 to 15 percent premium growth, and 3 to 5 percent improvements in claims accuracy.
Those numbers are important, but so is the level at which they are measured. A 30 percent improvement in the productivity of one workflow does not mean a carrier’s total expense ratio falls 30 percent. Savings can be absorbed by implementation costs, additional controls, higher service levels, increased volume, redeployed employees, model and computing costs, or investments elsewhere in the operation.
That is one reason PwC has cautioned insurers about measuring generative AI returns too quickly. PwC found that many insurer initiatives remained experimental and that measurable return on investment had been slow to materialize. Small pilots can establish technical viability without producing enough financial impact to demonstrate enterprise economics.
The lesson is not that AI lacks value. It is that the CFO needs to distinguish between technical performance, operational productivity, and captured financial value.
Where the Economic Case Is Becoming Visible
The emerging evidence suggests that the most useful unit of AI transformation is often not the individual model or prompt. It is the business domain in which the technology operates.
Claims provides a useful example. In his 2026 annual general meeting remarks, Allianz CEO Oliver Bäte said the group was operating more than 600 scalable AI applications across areas including pricing, risk assessment, sales, customer communication, and claims. He also reported productivity improvements of approximately 30 percent in certain areas of claims management. More than half of claims in one German pet-health operation were already being processed automatically.
Those are carrier-reported results, not an independent industry benchmark. But they demonstrate something more useful than a theoretical estimate: meaningful automation is possible inside bounded insurance processes, and some large insurers are now operating it at scale.
The same distinction applies in underwriting, distribution, servicing, and customer onboarding. AI can reduce document-handling time, accelerate information retrieval, improve routing, assist judgment, automate low-complexity work, and make expertise available earlier in a workflow.
The CFO’s job is to determine which of those operating improvements become economic improvements.
A carrier that processes twice as many submissions with the same underwriting staff may create substantial value without reducing headcount. A claims organization that shortens cycle time may improve customer retention even if loss adjustment expense initially changes very little. An underwriting tool that increases consistency may create value through better risk selection rather than administrative savings.
Expense reduction is only one value pool.
Governance Is Part of the Investment Case
The regulatory environment is also more mature than it was when many insurers began experimenting with generative AI.
The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, does not prescribe a single technical architecture or a four-item compliance checklist. Instead, it establishes an expectation that insurers using AI in regulated insurance decisions maintain a written AI Systems program appropriate to their use of the technology and the potential harm to consumers.
Importantly, the bulletin uses a risk-based approach. It says an insurer’s AI Systems program should be tailored and proportionate to its use and reliance on AI and to the degree of potential consumer harm.
It also places responsibility where insurance governance normally places responsibility. Senior management should be responsible for development, implementation, monitoring, and oversight of the program, with accountability to the board or an appropriate board committee.
The bulletin addresses data quality and lineage, model inventories, validation, testing, monitoring, model drift, documentation, internal controls, and the use of third-party systems and data. When an insurer relies on an outside AI provider, the insurer does not outsource its regulatory obligations with the technology. The bulletin contemplates due diligence, appropriate contractual provisions, audit rights where available, and cooperation with regulatory inquiries.
That fits a broader principle I have discussed elsewhere on InsuranceIndustry.ai: the carrier owns the decision architecture. A vendor can supply technology. It cannot assume the insurer’s authority or accountability for the resulting insurance decision.
The regulatory machinery is also becoming more operational. According to the NAIC’s current artificial intelligence resource page, 12 states were piloting an AI Systems Evaluation Tool as of March 2026 for use in market conduct, financial analysis, and financial examination contexts. The tool is intended to help regulators gather information about insurers’ AI use, governance, risk mitigation, higher-risk models, and data.
For a CFO, that changes the economics of governance. Documentation, monitoring, validation, and vendor oversight should not be treated as costs that appear after the business case has been approved. They belong inside the business case from the beginning.
Days 0-30: Define the Economic Unit
The first month should not begin with a search for the most impressive AI technology. It should begin with a process whose economics can already be measured.
The strongest candidate is usually a bounded domain with meaningful volume, a recognizable cost base, sufficient historical data, and an outcome the organization already monitors. Claims triage, submission intake, underwriting research, document classification, customer servicing, or a defined portion of policy administration may qualify. The right choice will differ by carrier.
Before introducing AI, establish the baseline. How many transactions are processed? How much employee time is required? What is the current error or rework rate? How long does the process take? Where are referrals created? What portion of the work requires actual judgment, and what portion consists primarily of gathering, organizing, comparing, or transferring information?
The baseline should also identify the business outcome. Saving five minutes on an underwriting submission has little meaning by itself. Saving five minutes across enough submissions to expand underwriting capacity, reduce backlog, improve response time, or allow underwriters to spend more time on risk selection has an economic consequence.
This is also the point at which ownership and authority need to be defined. Who owns the business outcome? What information may the AI access? What may it recommend? What, if anything, may it decide? When must the system refer the work to a person?
By Day 30, the CFO should have something more useful than an AI use case. The organization should have a measurable economic hypothesis.
Days 31-60: Measure the Workflow, Not the Demo
The second month is where many attractive AI demonstrations should fail.
A controlled test should compare the AI-assisted workflow with the existing process using actual operating measures. Depending on the use case, those measures might include handling time, throughput, accuracy, rework, referral rates, cycle time, conversion, claim leakage, customer outcomes, or some combination of them.
The test also needs to include the costs that disappear from technology demonstrations: model usage, software licensing, integration, data preparation, monitoring, human review, security, governance, training, change management, and vendor management.
This is particularly important with generative and agentic systems. AI is not free labor. A 2026 McKinsey analysis of insurance economics notes that AI can lower the marginal cost of underwriting, servicing, and claims tasks, but also carries real costs for models, computing, and orchestration.
The difference between gross productivity and captured value should therefore remain visible throughout the test.
If an employee completes a task 25 percent faster but spends the recovered time waiting for another system, the economic benefit is small. If the same improvement allows an underwriting team to absorb 20 percent more submissions without adding staff, the value may be substantial even though payroll does not decline.
The measurement should follow the economics of the actual workflow.
Days 61-90: Decide Whether the Evidence Earns the Next Dollar
By the final month, the question should no longer be whether the AI works.
The question is whether it works well enough, economically and operationally, to justify scaling.
The financial model should separate one-time implementation costs from recurring costs and distinguish gross productivity from benefits the organization can realistically capture. It should test what happens if transaction volume changes, model costs rise, accuracy deteriorates, additional human review becomes necessary, or the anticipated productivity improvement proves smaller than the pilot suggested.
The governance evidence should be reviewed at the same time. The carrier should know which data and models are being used, how outputs are validated, what decisions remain with people, how exceptions are handled, what records are retained, and what monitoring will occur after deployment.
There are three perfectly acceptable outcomes at Day 90.
The organization may have enough evidence to scale. It may have found value but need to modify the workflow or controls before expanding it. Or it may conclude that the economics do not justify further investment.
The third result is not a failed AI strategy. Killing an uneconomic project after 90 days is substantially cheaper than discovering the same thing after an enterprise rollout.
What the CFO Should Actually Put in Front of the Board
The board does not need another presentation explaining that artificial intelligence will transform insurance.
It needs evidence.
The useful presentation begins with the old process and the measured baseline. It then shows what changed under AI assistance, how much of the improvement is economically capturable, what the technology costs at production volume, what risks and controls accompany the deployment, and what assumptions determine the projected return.
The most important number may not be an expense-ratio target at all.
For one carrier it may be cost per claim. For another it may be submissions processed per underwriter. For another it may be quote turnaround time, new-business conversion, retention, leakage, rework, or the percentage of low-complexity transactions that can be completed without manual intervention.
Those operational measures can eventually influence financial ratios. But jumping directly from a workflow improvement to a companywide expense-ratio promise hides more than it explains.
AI Needs to Earn the Right to Scale
The case for AI in insurance is stronger today than it was when generative AI first arrived in boardrooms. There are real production deployments, measurable operating results, more mature implementation practices, and a clearer regulatory framework.
There is also enough evidence now to stop pretending that every pilot will create material financial value.
That is why a 90-day roadmap still makes sense, but for a different reason than the familiar technology sprint.
Ninety days is not a credible timetable for promising a predetermined reduction in a carrier’s expense ratio. It is enough time to establish a baseline, test a bounded workflow, measure the economics, define the decision authority, document the controls, and determine whether the evidence justifies the next investment.
The objective of the first 90 days is not to prove that AI works. It is to determine whether this AI, in this process, under these controls, creates enough value to earn the right to scale.
Sources
- AM Best, “Lower U.S. Property/Casualty Insurer Expenses Boost Segment’s Underwriting Results,” January 6, 2026
- McKinsey & Company, “The Potential of Generative AI in Insurance,” November 20, 2024
- McKinsey & Company, “The Future of AI in the Insurance Industry,” July 15, 2025
- McKinsey & Company, “How AI Is Reshaping the Economics of Insurance,” 2026
- PwC, “Building Sustainable GenAI ROI in Insurance”
- National Association of Insurance Commissioners, “Model Bulletin: Use of Artificial Intelligence Systems by Insurers,” adopted December 4, 2023
- National Association of Insurance Commissioners, “Artificial Intelligence,” current regulatory activity and AI Systems Evaluation Tool information
- Allianz SE, CEO Report at the 2026 Annual General Meeting
AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.
