AI Hiring by U.S. Insurers: Adoption, Diffusion and Outcomes

Figure 1

Insurers observed in 2026 and their AI hiring status

Figure 1 decomposes the 4,480 NAIC insurers present in 2026 according to what the postings data record about them. Of these, 2,352 (52%) are matched to Linkup and 1,905 post at least one job; among insurers that post, 83% post at least one AI role. Insurers not matched to Linkup are shown separately and treated as unobserved rather than as non-adopters. The second table reports the corresponding counts for every insurer present at any time in 2010–2026: of 5,900 such insurers, 24% had merged, entered run-off or exited by 2026.

Figure 2

Insurers with AI hiring by year

Figure 2 reports, for each year, the number of insurers observed in Linkup and the number with AI hiring. The observable population is stable at roughly 2,100–2,400 insurers, whereas the number with AI hiring rises from zero in 2010 to 1,580 in 2026. Because both series are measured in the same year, insurers that merge or enter run-off leave the numerator and the denominator together.

Figure 3

AI job postings by quarter

Figure 3 plots the number of AI job postings per quarter. AI postings are negligible before 2014 — six in total over 2010–2013 — and reach 4,092 in 2026Q2. The final quarter, 2026Q3, contains two months of data and is shown in a lighter shade.

Figure 4

AI postings as a share of all insurer postings

Figure 4 expresses AI postings as a share of all insurer postings. The share rises from about 0.4% in 2019 to 2.47% in 2026Q2. The final observation, marked with a hollow point, is affected by truncation: AI postings are truncated more severely than total postings in the partial quarter.

Figure 5

AI postings and total postings by year

Figure 5 compares annual AI postings with total insurer postings. Total postings show no trend over the period, so the rise in the AI share in Figure 4 reflects growth in AI hiring rather than a shrinking denominator. Lower total postings in 2010–2013 partly reflect thinner source coverage in those years; AI postings are close to zero in those years in either case.

Figure 6

Share of hiring units posting at least one AI job

Figure 6 measures diffusion at the extensive margin: the share of hiring units that post at least one AI job in a year. The share rises from 28% in 2016 to 65% in 2026.

Figure 7

Distribution of AI intensity across hiring units

Figure 7 shows the distribution of AI intensity, defined as AI postings as a share of a unit's postings, across hiring units. The median unit's intensity is below 0.03% in every year before 2025 and reaches 0.82% in 2026, while the 90th percentile rises to 6.67%. The distribution has thus both shifted upward and widened.

Figure 8

Year of first AI posting by hiring unit

Figure 8 reports the year in which each hiring unit first posts an AI job. First adoptions are spread across the period rather than concentrated in a single wave: between 14 and 30 units post their first AI job in each year from 2014 onward, and 287 units had posted at least one AI job by 2026.

Figure 9

AI postings by business function

Figure 9 reports AI postings by business function over the full period, with the share posted in 2025–2026 shown within each bar. Recent growth is uneven across functions: 79% of all cybersecurity AI postings and 78% of compliance AI postings were posted in 2025–2026, compared with 25% for data science. Functions are not mutually exclusive, so a posting may be counted in more than one.

Figure 10

Change in the functional composition of AI postings

Figure 10 compares each function's share of AI postings in 2016–2019 and 2025–2026. General machine learning gains the most (+17.0 percentage points), followed by cybersecurity (+7.3) and compliance (+4.2), while data science loses the most (−29.8). Actuarial and underwriting functions lose share (−2.9 and −0.2) even though their posting counts reach record levels in 2026 (Figure 11); a declining share and a rising count are compatible when total AI hiring grows faster.

Figure 11

AI hiring in six supervised functions

Figure 11 reports average quarterly AI postings in six functions of direct supervisory interest: actuarial, underwriting, claims, compliance, fraud and cybersecurity. Postings in all six reach their highest level in 2026. The increases in compliance and cybersecurity do not appear to be an artifact of classifier drift, since the underlying job titles are new rather than relabelled. Panels are drawn on different scales.

Figure 11A

Years from the first AI posting to the first AI posting in each function

Figure 11A reports the sequence. Data science comes first: 60% of adopters post a data-science role in the year of their first AI posting and 93% within eight years. General machine learning, actuarial and pricing, underwriting, and marketing and sales follow with a median lag of two years, each reaching between 40% and 66% of adopters by year 8. Claims comes later and is rarer, 2% in the first year and 21% by year 8, with a median lag of 3.5 years; most other functions have median lags of three to 4.5 years. Compliance and cybersecurity are last among the functions with more than a handful of adopters, with median lags of five and six years, and their shares accrue mostly in 2024–2026.

Figure 11B

AI postings by function and year

Figure 11B reports the same functions at the posting level. Data science dominates until 2024; general machine learning overtakes it in 2025 and reaches 2,920 postings in the first three quarters of 2026. Actuarial and pricing, underwriting, and marketing and sales grow through the period, and claims postings reach 101 in 2024 and 206 in 2026. Compliance and cybersecurity are small until 2024 and jump in 2025, to 321 and 542. The 2024 customer-service count comes almost entirely from one unit, Berkshire Hathaway.

Figure 11C

AI postings by function, one panel per function

Figure 11C plots each function on its own scale, grouped into three patterns: functions that grow through the 2010s (data science, actuarial and pricing, underwriting, marketing and sales); functions that grow gradually and accelerate in 2025–2026 (general machine learning, claims, finance, product development, policy administration); and functions with little history before 2025 (compliance, cybersecurity, risk control, catastrophe modeling). Three health insurers account for 36% of the 2025–2026 compliance labels and 45% of the cybersecurity labels.

Figure 12

AI hiring by line of business

Figure 12 reports AI hiring by the segments in which a unit writes business. A unit writing in more than one segment is counted in each, so the three series overlap and do not sum to the total. In 2026, AI roles account for 3.68% of postings by property-casualty units, 3.40% by life units and 1.31% by health units.

Figure 13

Concentration of AI postings across hiring units

Figure 13 shows the cumulative share of recent AI postings accounted for by the largest hiring units. The five largest account for 33% and the twenty-five largest for 80%. Industry-wide averages therefore largely reflect the hiring of a small number of units.

Figure 14

The twenty largest AI recruiters

Figure 14 lists the twenty units with the most AI postings. Rankings by volume and by intensity differ: UnitedHealth posts the most AI jobs, but these are 2.0% of its postings, whereas AI roles account for 8.6% of postings at Berkshire Hathaway and 7.2% at USAA.

Figure 15

AI hiring by insurer size

Figure 15 relates AI hiring to size for property-casualty units grouped into quartiles of 2024 admitted assets. The share of units posting any AI job rises monotonically with size, from 51% in the smallest quartile to 93% in the largest, whereas the AI share of postings varies much less (2.0% to 4.2%). Size thus predicts whether a unit hires AI talent more strongly than how much it hires. The figure is limited to property-casualty units because the NAIC financial data used here do not cover life or health insurers.

Figure 16

AI postings by state

Figure 16 reports AI postings by state. All postings in the sample are located in the United States. States with large concentrations of insurance employment — Connecticut, Iowa and Minnesota — account for more AI postings than their populations alone would suggest.

Figure 17

AI postings by occupation

Figure 17 reports average quarterly AI postings by occupation. Management occupations grow fastest, from 61 postings per quarter in 2019 to 769 in 2026, compared with an increase from 390 to 1,834 in computer and mathematical occupations. Insurers are therefore hiring for the management and oversight of AI as well as for its development.

Figure 18

Seniority of AI postings

Figure 18 classifies AI postings by seniority, inferred from job-title keywords. Manager, lead and principal roles, together with director-level roles and above, account for 51% of AI postings in 2026, compared with 25% in 2020. The 2026 figure covers a partial year.

Figure 26

Hiring shares and AI intensity by function

Sections 1–6 describe the volume and composition of AI hiring, and Section 9 its association with insurer outcomes. Section 7 asks whether AI hiring comes at the expense of other hiring within the same business function. For each function, Figure 26 plots its share of all insurer postings over 2012–2026 (upper panel) and AI's share of the function's postings from 2016 (lower panel); the dashed segment and hollow point mark 2026, which contains only 2026Q1–Q2. If AI hiring displaced other work, the functions losing share of total postings would be those with the highest AI intensity. The data show the reverse. The five functions that lost share — marketing (−6.00 percentage points), finance (−2.59), customer service (−2.21), policy administration (−1.04) and risk control (−0.49) — have AI intensities between 0.09% and 0.77%, whereas the most AI-intensive functions — enterprise risk management (6.4%), catastrophe modelling (5.5%), fraud (4.5%), cybersecurity (4.4%) and actuarial (3.4%) — all gained share.

Measuring the correlation. Because each posting is either AI or non-AI, the AI and non-AI shares of a function sum to one and are perfectly negatively correlated (−1.000) by construction. Any measure with the function's own total in both denominators inherits part of this mechanical relationship: a fall in non-AI hiring raises AI intensity whether or not AI caused it. We therefore normalize both series by a denominator that neither controls, namely all insurer postings in the year. On this basis, the correlation between AI and non-AI hiring is +0.06 across function-years, +0.16 in 2016–2026 changes across the sixteen functions, +0.18 across functions in 2026, and +0.09 in the firm-level panel after removing insurer-by-year and function-by-year effects. None is negative.

The influence of 2026. Estimated function by function over eleven years, the median correlation is +0.32, but 6 of the 15 functions with any AI hiring have negative correlations, compared with 4 when 2026 is excluded; these include customer service (−0.65), finance (−0.50), policy administration (−0.40) and cybersecurity (−0.30). Both sets of estimates are reported in the table. The function-level correlations should be interpreted with caution: each rests on eleven annual observations, the last of which covers half a year, and the finance estimate moves from +0.29 to −0.50 with the addition of that observation. The pooled and firm-level estimates, which no single year can move materially, remain positive.

Regression evidence. A regression of non-AI postings on the firm's accumulated AI postings in the same function, estimated on the full panel through 2026Q2 with unit, function-by-quarter and unit-by-quarter fixed effects and standard errors clustered by insurer group, yields a coefficient of −0.18 non-AI postings per AI posting (s.e. 0.27; 95% confidence interval −0.70 to +0.36) across 454,650 unit–function–quarters and 391 groups. The fixed effects leave 76% of the variation in AI hiring, so the estimate is informative rather than a consequence of absorbed variation. By period, the 2023–2026 coefficient is −0.24 (p = 0.43), also indistinguishable from zero. The periods differ significantly (p = 0.008) because the coefficient for 2016–2019 was positive (+4.16; p = 0.007). The evidence therefore points to a weakening of the earlier complementarity between AI and non-AI hiring rather than to displacement.

Figure 19

AI hiring by organizational form, property-casualty insurers

Figure 19 compares AI hiring across organizational forms among property-casualty insurers. By count, 52% of mutual insurers hire for AI, compared with 71% of stock insurers. Weighted by premium written, the ranking reverses: 74% of mutual premium and 70% of stock premium are written by insurers that hire for AI. The difference in counts therefore reflects size: small mutuals are largely absent both from the postings data and from AI hiring, whereas the large mutuals that write most mutual premium are present in both. Adoption among mutuals also grew fastest, from 31% of observed companies in 2024 to 52% in 2026.

Figure 20

Coverage of the property-casualty industry

Figure 20 measures how much of the property-casualty industry the matched sample covers. Insurers matched to Linkup are 48% of property-casualty companies but write 84% of net premiums and hold 73% of admitted assets; the unmatched half of companies writes 16% of premium. Insurers with AI hiring alone account for 66% of premium.

Figure 21

Any AI hire and the loss ratio, all lines

Figure 21 reports event-study estimates of the effect of a first AI hire of any kind on the loss ratio, defined as losses incurred plus DCC per $100 of direct earned premium. The imputation estimate of the average post-adoption effect is −0.64 (s.e. 1.36; p = 0.637), against a sample mean of 63.9; the two-way fixed-effects estimate is 0.97 (s.e. 1.04; p = 0.349). Neither is distinguishable from zero, and the pre-adoption coefficients are jointly insignificant (p = 0.217). The confidence interval rules out reductions of more than 3.3 per $100 of premium. We therefore find no evidence that AI hiring, in the aggregate, has altered insurers' loss experience in the markets they write.

Figure 22

Claims AI and the paid loss ratio

Figure 22 considers first hires of AI for claims handling and fraud detection, with the paid loss ratio as the outcome. Across all lines (left panel), the imputation estimate is −2.64 (s.e. 0.89; p = 0.003), or 5% of the sample mean of 48.7. In personal auto and homeowners (right panel), it is −4.04 (s.e. 1.46; p = 0.006). The dynamic coefficients show an effect that builds gradually, reaching about −6.1 four to five years after adoption, while the pre-adoption coefficients are small and jointly insignificant (p = 0.542 and p = 0.964, respectively). The two-way fixed-effects estimates have the same sign but are smaller: −1.61 (s.e. 1.11; p = 0.151) across all lines and −2.26 (s.e. 1.05; p = 0.033) in personal lines. A lower paid loss ratio is consistent with more effective fraud detection, firmer settlement or reduced payment on valid claims; the statutory data do not distinguish among these channels.

Figure 23

Distribution AI and the loss ratio

Figure 23 examines first hires of AI in distribution — marketing, customer service and policy administration. Across all lines (left panel) we find no effect: the imputation estimate is −0.32 (s.e. 1.70; p = 0.852) and the two-way fixed-effects estimate 0.66 (s.e. 1.60; p = 0.678). In personal auto and homeowners (right panel), by contrast, the estimated effect on the loss ratio is −3.38 (s.e. 1.45; p = 0.020) under imputation and −3.17 (s.e. 0.95; p < 0.001) under two-way fixed effects, and the pre-adoption test does not reject (p = 0.157). Estimated line by line, the decline appears larger in homeowners (−4.93) than in private passenger auto (−2.14), but the pre-adoption test rejects within each line (p = 0.043 and p = 0.016), so we do not interpret this split. One candidate mechanism is AI-assisted targeting and retention of lower-risk policyholders, which would bear on access to coverage; the data do not identify the mechanism.

Figure 24

Summary of estimates by estimator

Figure 24 collects the average post-adoption effects from both estimators. We regard a result as robust when the two estimators agree in sign and approximate magnitude and the pre-adoption test does not reject. The effects of claims AI on paid losses and of distribution AI on the loss ratio in personal lines meet this standard. The claims-AI effect across all lines has the same sign under both estimators but is statistically significant under imputation only. The effect of underwriting and pricing AI on the personal-lines loss ratio is not significant under either estimator, and the settlement-speed effect (Figure 25) is significant under imputation only.

Figure 25

Claims AI and settlement speed, personal lines

Figure 25 reports imputation estimates for settlement speed in personal auto and homeowners, defined as losses paid during the year as a percentage of losses paid plus losses unpaid at year end (sample mean 62.2%). The estimated average change after a first claims-AI hire is −1.96 (s.e. 0.69; p = 0.005) percentage points. The pre-adoption coefficients are small and jointly insignificant (p = 0.154), and the decline widens over time, reaching −3.6 percentage points by the fifth year after adoption. The effect is not specific to claims AI, however: first hires in underwriting and pricing and in distribution are followed by comparable declines of −1.94 (s.e. 0.72; p = 0.007) and −2.20 (s.e. 0.89; p = 0.014) percentage points, respectively, which suggests that the pattern may reflect broader organizational change coinciding with AI adoption. The two-way fixed-effects benchmark for this outcome is reported in Figure 24. Because slower claim payment is relevant to market-conduct oversight, we regard this result as warranting monitoring.