"The Brittle Workforce" -- Insurance, Entry-Level Automation, and What Happens Next

Twin Ladder Casebook Series | Twin Ladder | February 2026

The Hook

A claims adjuster retires after thirty years at a mid-sized European insurer. In her final week, she processes a motor claim that the system has flagged as straightforward: a rear-end collision, moderate damage, clean police report. She reads the file for ninety seconds and picks up the phone. The claimant's repair estimate includes parts for a vehicle model two years newer than the one on the policy. The garage is padding the invoice. She has seen this pattern four thousand times. She catches it without thinking, because thinking is not what she is doing. She is recognizing.

Her replacement starts the following Monday. He is twenty-four, university-educated, proficient with every digital tool the company deploys. He has never processed a claim manually. The AI does it. The system ingests the documentation, cross-references the policy, runs the estimate against historical benchmarks, and produces a recommendation. The new adjuster reviews the recommendation, confirms it, and moves on. He processes three times the volume his predecessor handled. His throughput numbers are exceptional.

Six months later, a regional fraud ring begins submitting inflated repair estimates across the insurer's portfolio. The pattern is subtle -- not the kind that triggers a rule-based flag, but the kind that a human who had spent years reading invoices would notice in the texture of the numbers. Nobody notices. The AI does not catch it because it was not trained on this specific pattern. The junior adjuster does not catch it because he has never learned what a legitimate invoice feels like. The expertise retired. The training ground was automated. And the organisation is brittle in a way it cannot yet see.

The Story

The Race to Automate

The European insurance industry has embraced artificial intelligence with a speed and conviction that few sectors can match. AXA and Allianz -- the two largest insurers on the continent -- have emerged as global leaders in AI deployment, and the scale of their ambitions is not incremental. It is structural.

AXA launched its proprietary AXA Secure GPT platform in July 2023, built on Microsoft Azure OpenAI Services. The tool was initially made available to one thousand employees at AXA Group Operations, with the stated objective of reaching all 140,000 employees globally. By 2025, the company reported approximately four hundred use cases spanning predictive, generative, and agentic AI -- deployed across underwriting, claims, customer service, compliance, and internal operations. AXA's data engine fuses traditional actuarial data with real-time streams from IoT devices, satellite imagery, and advanced simulations. The integration is not peripheral. AI is, in the company's own framing, "the connective tissue of the entire enterprise."

Allianz has followed a parallel trajectory. As of December 2024, the company reported close to four hundred generative AI use cases alone, with the number continuing to grow. AllianzGPT, the company's internal chatbot, launched on September 1, 2023, and now serves over 60,000 active users, with the ambition of reaching all 158,000 Allianz employees worldwide. The platform handles document summarisation, translation, complex comparisons, and customer communication drafting -- tasks that previously required hours or days and now complete in minutes. In early February 2025, Allianz added a ring-fenced version of DeepSeek to AllianzGPT, expanding the range of models available within its secure environment.

Aviva, the UK insurer, has taken a different but equally aggressive approach. The company deployed more than eighty AI models across its claims function, assembling a team of over fifty data scientists and engineers to build and embed the technology. The results were substantial: liability assessment time for complex cases fell by twenty-three days, claims routing accuracy improved by thirty percent, and customer complaints dropped by sixty-five percent. Aviva told investors that the transformation saved the company more than sixty million pounds -- approximately eighty-two million dollars -- in 2024 alone.

The Question Nobody Is Answering

The deployment numbers are impressive. The return-on-investment numbers are not -- because they largely do not exist.

In June 2025, Evident Insights published its first-ever ranking of the thirty largest insurers in North America and Europe. The study assessed seventy-six individual metrics across four pillars: talent, innovation, leadership, and responsible AI transparency. AXA and Allianz dominated, ranking as the only two insurers to place in the top five across all four categories. But a different finding in the report carried far greater implications.

Of the thirty insurers evaluated, only twelve had disclosed at least one AI use case with a tangible business outcome. And only three -- Intact Financial, Zurich Insurance Group, and Aviva -- had publicly disclosed a monetary return from their AI efforts. Intact Financial, the Canadian property and casualty insurer, reported a five hundred million dollar technology investment that had deployed five hundred AI models and generated one hundred and fifty million dollars in measurable benefit. Three out of thirty. Ninety percent of the largest insurers in the Western world could not -- or would not -- demonstrate that their AI investments had produced financial returns.

Meanwhile, the workforce question remains unanswered. More than half of insurance companies report skills gaps and recruitment challenges around AI, according to a Gallagher survey published in early 2026. Only one in three insurers has a formal AI training programme in place. PwC's 2025 Global AI Jobs Barometer found that skill requirements for roles exposed to AI are evolving sixty-six percent faster than those in other fields, and that jobs requiring AI skills command a wage premium of fifty-six percent -- up from twenty-five percent the previous year.

The industry is automating at speed. It is not building the workforce to govern what it has automated.

The Demographic Cliff

The timing could not be worse. The insurance industry is facing a retirement wave of historic proportions. The number of insurance professionals aged fifty-five and older has increased seventy-four percent in the last decade. Over the next fifteen years, fifty percent of the current workforce will retire, leaving more than four hundred thousand positions unfilled. The ratio of retirement-age employees to young entrants stands at six to one: 1.37 million workers aged fifty-five or older, against just 214,000 between the ages of twenty and twenty-four.

These departing professionals carry something that no AI model currently replicates: the tacit knowledge built through years of processing claims by hand, evaluating risks in ambiguous situations, and recognising patterns that no training dataset has captured. When they leave, that knowledge leaves with them. And the entry-level roles that once served as the pipeline for rebuilding that knowledge -- the claims intake positions, the data entry functions, the junior underwriting support roles -- are precisely the roles being automated first.

Through the Twin Ladder Lens

The insurance industry's automation pattern maps directly onto a Level 1 erosion in the Twin Ladder framework -- and it illustrates the most dangerous variant of the Competence Paradox: the destruction of the training ground.

In the Twin Ladder, Level 1 -- the Professional Twin -- exists to mirror individual roles with AI agents so that human professionals can compare their judgment against AI output, learn from the differences, and develop deeper domain expertise through deliberate engagement. The design principle is explicit: the twin must preserve domain competence, not erode it. The human remains actively engaged with the domain, questioning AI output, correcting it, and understanding why it produced the results it did.

Entry-level claims processing was never recognised as a Level 1 asset, but it functioned as one. It was the training ground. A junior claims processor spent two to three years reading documentation, cross-referencing policies, spotting inconsistencies, and developing the pattern-recognition instincts that would eventually qualify them for underwriting, fraud investigation, or complex claims adjudication. The work was repetitive. It was also formative. The repetition was the mechanism through which durable competence formed -- precisely the dynamic that Robert Bjork's research on desirable difficulties predicts.

When insurers automate claims intake, they do not merely remove a task. They sever the pipeline that produces the professionals who will govern the AI systems performing that task. The junior processor who never manually reconciles an invoice against a policy cannot recognise when the AI's reconciliation is wrong. The trainee underwriter who has never assessed a risk without algorithmic support cannot override the algorithm when the algorithm encounters a situation outside its training data. The expertise that the retiring generation carries was built through the very work now being automated. Remove the work, and you remove the mechanism that creates the expertise.

This is the "brittle workforce" -- an organisation staffed by people who can operate AI tools but cannot evaluate their output, who can confirm a recommendation but cannot challenge it, who can process volume but cannot exercise judgment. The workforce is not incompetent. It is untested. And in an industry where a single undetected fraud pattern or a misassessed catastrophe risk can cost hundreds of millions, untested is indistinguishable from brittle.

The European Insurance and Occupational Pensions Authority recognised this risk in its August 2025 Opinion on AI Governance and Risk Management, which emphasised human oversight, explainability, and the requirement that AI outputs be "meaningfully explainable" to allow identification of potential bias. The regulatory assumption is that humans will provide meaningful oversight. But meaningful oversight requires domain competence that the current automation trajectory is systematically eliminating. The regulation presumes a workforce that the industry is choosing not to build.

The Pattern

Insurance is not an isolated case. It is the latest instance of a pattern that is now visible across every knowledge-intensive industry: automate the training ground, and you lose the future experts.

Manufacturing recognised this first. A study by Deloitte and The Manufacturing Institute projected that 2.1 million manufacturing jobs in the United States could go unfilled by 2030, at a cost of one trillion dollars in that year alone. The root cause is not a shortage of machines. It is a shortage of people who understand what the machines are doing. Executives reported that they cannot fill even higher-paying entry-level production positions, let alone specialised roles. The pipeline that once moved workers from the shop floor to the engineering office has been disrupted by automation at the entry level.

Financial services exhibits the same dynamic. Klarna, the Swedish buy-now-pay-later company, replaced approximately seven hundred customer service agents with an AI chatbot in 2023 and reported dramatic efficiency gains -- only to announce in May 2025 that it was rehiring human agents after customer satisfaction scores dropped by twenty-two percent. The agents who were replaced carried tacit knowledge about customer behaviour, fraud patterns, and escalation triggers that no dataset had captured. When the AI made errors, nobody remained who could recognise them.

Forrester's Predictions 2026 report quantified the broader trend: fifty-five percent of employers who laid off workers for AI now report regretting the decision. Gartner predicted in February 2026 that by 2027, half of all companies that reduced customer service headcount due to AI will rehire staff to perform similar functions. The cycle -- automate, discover the gap, rehire at higher cost -- is becoming a recognisable pattern across sectors.

The common thread is not that automation is wrong. It is that automation without a competence strategy is self-defeating. Every industry that eliminates entry-level roles without redesigning them is making the same bet: that AI will advance fast enough to close the gap before the absence of trained humans becomes critical. It is a bet on a future capability to compensate for a present destruction, and the historical record offers no precedent for that bet paying off. Electricity took thirty years to deliver its promised productivity gains, as Paul David demonstrated, because organisations substituted new technology into old structures without rebuilding the human capabilities to govern it.

The insurance industry, with its six-to-one retirement ratio and its four-hundred-thousand-position shortfall, is making that bet with the worst demographic odds of any sector in the Western economy. And unlike manufacturing, where the unfilled roles are visible on the factory floor, the insurance knowledge gap is invisible until a claim goes wrong, a risk is mispriced, or a fraud pattern runs unchecked for months. The brittleness does not announce itself. It compounds silently until the moment it breaks.

The Lesson

The lesson is not to stop automating claims intake. It is to redesign what entry-level insurance professionals do, not eliminate them.

The Twin Ladder approach is direct. Junior staff should not process claims manually because the manual work is valuable in itself. They should evaluate AI-processed claims because the evaluation builds the judgment that the organisation will need in five, ten, and twenty years. A trainee who reviews fifty AI-adjudicated claims per day -- flagging anomalies, questioning assessments, comparing AI output against policy language -- develops pattern recognition faster than one who processes ten claims manually. The AI handles the volume. The human handles the judgment. Both become more capable.

This is a Level 1 implementation applied to workforce development. The Professional Twin does not replace the junior professional. It becomes the instrument through which the junior professional develops expertise. The trainee learns not by doing the work the AI does, but by critically assessing whether the AI did it correctly. This is harder than manual processing. It requires higher-order thinking. It is, in Bjork's framework, a desirable difficulty -- and the evidence shows that desirable difficulties produce dramatically better retention and transfer than easy repetition.

The insurers that will lead the next decade are not the ones deploying the most AI models. They are the ones ensuring that their people can govern those models with genuine understanding. AXA has four hundred use cases. Allianz has four hundred generative AI deployments. These are substantial achievements. But the question that will determine whether those investments produce lasting value is not how many AI models the company has deployed. It is whether, ten years from now, the company still employs people who can tell when those models are wrong.

The demographic clock is running. Four hundred thousand positions will go unfilled. Fifty percent of the current workforce will retire. The knowledge they carry was built over decades. It cannot be rebuilt in months. And it will not be rebuilt at all if the industry automates the only mechanism through which it was ever created.

Monday Morning Question: If you removed AI from your claims processing function tomorrow, how many of your current staff could evaluate a claim from first principles -- and if the answer is fewer than five years ago, what are you doing about it?

Sources

AXA -- "AXA Offers Secure Generative AI to Employees" (2023): https://www.axa.com/en/press/press-releases/axa-offers-securegenerative-ai-to-employees
Allianz -- "AI at Allianz: The Impact of AllianzGPT" (February 2025): https://www.allianz.com/en/mediacenter/news/articles/250218-ai-at-allianz-the-impact-of-allianzgpt.html
Fortune -- "The Insurance Sector Is Getting Onboard with AI. But, Like Other Industries, They Are Quiet About ROI" (June 2025): https://fortune.com/2025/06/24/insurance-companies-ai-ranking-allianz-axa-evident-insights-eye-on-ai/
McKinsey -- "Aviva: Rewiring the Insurance Claims Journey with AI" (2024): https://www.mckinsey.com/capabilities/tech-and-ai/how-we-help-clients/rewired-in-action/aviva-rewiring-the-insurance-claims-journey-with-ai
Insurance Business -- "When the Algorithm Knocks: How AI Is Quietly Rewriting Insurance Jobs" (2025): https://www.insurancebusinessmag.com/us/news/breaking-news/when-the-algorithm-knocks-how-ai-is-quietly-rewriting-insurance-jobs-554187.aspx
Insurance Business / Gallagher -- "AI Goes Mainstream, But Insurers Face Skills, Risk and Coverage Gaps" (2026): https://www.insurancebusinessmag.com/us/news/technology/survey-ai-goes-mainstream-but-insurers-face-skills-risk-and-coverage-gaps-566336.aspx
PwC -- "2025 Global AI Jobs Barometer": https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2025/report.pdf
Deloitte / The Manufacturing Institute -- "2.1 Million Manufacturing Jobs Could Go Unfilled by 2030": https://nam.org/2-1-million-manufacturing-jobs-could-go-unfilled-by-2030-13743/
EIOPA -- "Opinion on AI Governance and Risk Management" (August 2025): https://www.eiopa.europa.eu/eiopa-publishes-opinion-ai-governance-and-risk-management-2025-08-06_en
Jonus Group -- "Insurance Talent: Why 1.4 Million Retirements Will Reshape the Industry" (2025): https://www.jonusgroup.com/blog/2025/10/insurance-talent-why-1-dot-4-million-retirements-will-reshape-the-industry