The measured size of it
- 44%
- Insurers in top AI quartile
- 11%
- Insurers with clear AI return view
- 11%
- Insurers with foundations to scale
On the desk is a single page that names one claims process, its pre-AI minutes per file and error rate, the fully loaded annual cost of the tool and the staff who check its output, the month cumulative benefit overtakes that cost, and the executive who signs the comparison. Anyone who was not in the meeting can verify all five. The rest of this page works back from that sheet, because the ability to produce it is what separates an AI strategy from a subscription list.
A claims operation makes the shape concrete. Suppose triage of incoming motor claims takes a fixed number of minutes a file and misroutes a share of cases. An AI triage tool might improve both. Whether it did is a question about the minutes, the misroutes, the cost of the tool, the cost of the people who now check its output, and the date the comparison is made. Nothing about the model answers it. That is the whole difficulty, and it is an accounting difficulty before it is a technology one.
Confidence versus meaningful business transformation
Insurers rate their own transformation higher than their process changes justify. Reinsurance News reported that 44 per cent of insurance executives place themselves in the top quartile for AI transformation, and that not one respondent considers their firm significantly behind. The same Reinsurance News report found that no surveyed firm had fully rebuilt sales and distribution or underwriting around AI, and that 3 per cent had achieved full redesign in claims management and policy servicing.
Urgency is not the missing ingredient. The same Reinsurance News report found that 77 per cent of insurers expect to lose competitiveness within five years without an enterprise architecture redesign, and that 68 per cent see slow adoption as the greater risk. What the technology is used for is narrower. Content generation and routine automation are the common applications, while only a minority run end-to-end processes through AI agents.
Measurement has not kept pace. Reinsurance News reported that only 11 per cent of insurers describe their view of AI return as very clear, and that 21 per cent describe themselves as only partially ready on data. The data foundation is the binding constraint: Reinsurance News reported that only 11 per cent possess the strong foundations and governance needed to scale AI beyond pilots. That is not a technology gap in the model. It is a gap in ownership, quality and access.
Workforce capability follows the same pattern. Reinsurance News reported that 8 per cent rate their workforce as highly proficient with AI tools, even though 54 per cent say they provide effective training. Training is an input. Proficiency is an output. The survey suggests that many firms are counting the first and assuming the second.
The employment picture in the same report is not distant. Reinsurance News reported that 72 per cent expect underwriting to shift to a hybrid model with fewer roles by 2029, while 36 per cent anticipate significant role elimination in claims management and 33 per cent in policy servicing. A board that accepts those projections and still measures AI only by licence count is measuring the wrong thing. The Reinsurance News report also records the KPMG view that "the gap between activity and transformation is where the real opportunity and risk now sit." That sentence is the argument of this page in one line.
Prioritise the process over the tool
The first discipline is to choose the process, not the tool. A tool has users, prompts and a licence fee. A process has a baseline, a cycle time, an error rate, a cost per case and a customer outcome. The same model can be a success in one process and a distraction in another because the surrounding workflow decides whether the output is used. MIT Sloan Management Review's research on three approaches to AI ROI found that two companies making nearly identical investments may define success in entirely different ways. That is not a failure of arithmetic. It is a failure to agree on the unit being measured.
The second discipline is to separate efficiency from effectiveness. Efficiency reduces the time, cost and effort of executing a known task. Effectiveness improves the quality of a decision, the priority of a case, the personalisation of an offer or the accuracy of a forecast. The insurance survey is full of efficiency claims. Reinsurance News reported that 92 per cent say AI enhances productivity and lowers operating expenses, but only a minority use it for growth through new capabilities and services. A cost saving is real, but it has a ceiling. A better decision can change revenue, loss ratio and retention. The measurement should show which one is being claimed.
The third discipline is attribution. AI is rarely introduced alone. It arrives with a data clean-up, a team reorganisation, a new case management system or a change in policy wording. Deloitte's 2025 survey found that most respondents achieve satisfactory ROI on a typical AI use case within two to four years, and that only 13 per cent of the most successful projects saw returns within 12 months. The same research notes that AI is entangled with broader transformation, which makes it difficult to isolate its contribution. That is not a reason to avoid measurement. It is a reason to build a control group, a before-and-after baseline or a difference-in-differences estimate into the business case.
How to build the one-page record
The baseline is the first column and the most commonly skipped. It should record the current minutes per file, error rate, rework rate, cost per case, volume, customer satisfaction and any regulatory or compliance measure that the process touches. The baseline must be dated. A baseline without a date becomes a moving target once the project starts.
The full cost is the second column. It includes the software licence, cloud usage, integration, data preparation, testing, security review, training, change management, human review, monitoring and maintenance. It also includes the cost of the people who check the output. A tool that drafts a claims decision still needs a skilled adjuster to sign it. If that check is not costed, the return is overstated. The full cost should be annualised and stated in the same unit as the benefit.
Attribution is the third column. The question is not whether the process improved after the tool arrived. The question is how much of the improvement belongs to the tool rather than to the data clean-up, the new team structure or the market. A control group is the cleanest answer. A staged rollout is the practical answer. A before-and-after comparison with a clear list of concurrent changes is the minimum answer.
Payback is the fourth column. The arithmetic is net benefit divided by total cost, with net benefit measured against the baseline. The date matters because a project can show a positive lifetime return and still fail the cash test in year one. The board should see the cumulative benefit line cross the cumulative cost line, and it should see who owns the number.
Data readiness sets the ceiling on any return
Data readiness is not a preliminary IT task. It is the ceiling on what the AI can do. Reinsurance News reported that only 11 per cent of insurers possess the strong data foundations and governance needed to scale AI beyond pilots, and that 21 per cent describe themselves as only partially ready. The reported causes are fragmented data, poor quality, unclear ownership and legacy systems. Each of those is a management problem before it is a model problem. If the claims data does not agree on what a closed case means, a better model will produce a better answer to the wrong question.
IBM's work on AI ROI makes the same point from the organisational side. IBM reported that only 29 per cent say they can measure ROI confidently, and that around 25 per cent of AI initiatives deliver expected ROI. The constraints it names are culture, governance, workflow design and data strategy. That list should be read as a budget line. A programme that funds the model but not the data ownership, the workflow redesign and the governance will produce a pilot that cannot scale.
Assigning ownership and risk management
Accountability is not a steering committee. It is a named manager who can stop the project, change the workflow and defend the number. KPMG UK's AI Pulse found that 41 per cent cite risk management, including cybersecurity and data privacy, as a key challenge to AI strategy. The same research found that 19 per cent are actively scaling AI agents across multiple functions, while 34 per cent remain in exploration or pilot. The gap between those two groups is not a gap in model capability. It is a gap in operating model, controls and decision rights.
The guardrails should be proportionate to the process. A content drafting tool needs review and provenance. A claims triage tool needs a human-in-the-loop, an audit trail and a route to appeal. An agent that can change a customer record needs access control, a stop condition and a monitoring owner. The governance question is not whether AI is allowed. It is which decisions the system can take, which decisions a person must take, and how the organisation knows the difference after the event.
KPMG UK's AI Pulse also found that 71 per cent are confident measuring ROI on performance. That confidence is useful, but it is not the same as clarity on enterprise AI return. A team can measure the performance of a tool inside a function and still miss the effect on the customer journey, the cost of rework or the risk that the tool has shifted elsewhere. The measurement has to match the boundary of the process.
Workforce proficiency is the adoption test
A trained workforce is not a proficient workforce. Reinsurance News reported that 8 per cent rate their workforce as highly proficient with AI tools, even though 54 per cent say they provide effective training. The difference matters because adoption is the point where a model becomes a return. If the adjuster does not trust the triage recommendation, the recommendation is not used. If the underwriter cannot explain the model output to a broker, the output does not enter the decision. If the compliance team cannot evidence the control, the process cannot scale.
The practical test is task-level. Can the person complete the new task without help, to the required quality, within the new cycle time? If the answer is no, the training has not landed. The business case should include the cost of getting that answer to yes, and the time it takes. That cost belongs in the full cost column, not in a separate transformation budget that nobody reconciles.
What the board should ask before approving the next tranche
The first question is which process, and which baseline, is being changed. The second is what the full annual cost includes, especially human review and data work. The third is how attribution will be tested, and what concurrent changes could explain the result. The fourth is when cumulative benefit overtakes cumulative cost, and who owns the comparison. The fifth is what would stop the project, and who has the authority to stop it.
A strategy that can answer those five questions is investable. A strategy that cannot is a portfolio of experiments with a communications plan. The insurance survey suggests that many firms have the ambition and the activity. The one-page record is the test of whether they have the return. The work back from that sheet is not glamorous. It is the work that turns an AI strategy into a business result.
Sources
- Insurance sector faces major gap between AI confidence and meaningful business transformation – KPMG
- KPMG Quarterly Pulse: Making AI scale across the Enterprise | KPMG UK
- AI ROI: The paradox of rising investment and elusive returns
- How to maximize AI ROI in 2026 | IBM
- Three Approaches to Measuring and Managing AI ROI
Source: Insurance sector faces major gap between AI confidence and meaningful business transformation: KPMG, Reinsurance News





