What mid-size insurers and manufacturers disclose about their AI

Insurers and manufacturers with $50M-$2B revenue: how many name an AI initiative in public, and how few credit a cash result to AI. November 2026.

How many mid-size insurers disclose their AI initiatives?

Mid-size US insurers talk about their AI work readily, but rarely say what it returned. Of 48 US-listed insurers with $50 million to $2 billion in 2025 revenue, 39 (81%) named at least one specific AI initiative in public between January 2024 and early November 2026. Only 2 (4%) gave a cash figure they credit directly to AI, and 1 of those 2 is an insurtech built around AI. None of the figures was audited or measured against a control group.

Insurers Share
Named at least one specific AI initiative 39 81%
Gave any figure for an AI initiative (cash, hours or operational) 17 35%
Gave a cash figure they credit directly to AI 2 4%
…among the 42 that aren't AI-native insurtechs or MGA platforms 1 2%
Mentioned AI only as a risk factor, anywhere we looked 6 13%

The insurer half of the study is the research behind the Roiva AI in Practice Index, Mid-Size Insurers 2026, which names eight leaders.

Do manufacturers disclose AI work as often as insurers?

Manufacturers name AI work less often than insurers, and none credits cash it has already saved or earned directly to AI. Of 49 US manufacturers with $50 million to $2 billion in 2025 revenue, 25 (51%) named a specific AI initiative in their own operations. Two (4%) credited a cash figure directly to AI, and both were forecasts: Westrock Coffee's CEO expects tens of millions of dollars of benefit a year from its Palantir rollout, and iRhythm expects more than $100 million in cumulative savings over five years from an ECG algorithm launching in 2027.

The comparison below leaves out the six AI-native insurtechs and MGA platforms, which have no counterpart among the manufacturers.

Insurers (42) Manufacturers (49)
Named a specific AI initiative 33 (79%) 25 (51%)
Gave any figure for one 11 (26%) 8 (16%)
Gave a cash figure they credit directly to AI 1 (2%) 2 (4%), both forecasts
Nothing found beyond risk factors, or no AI mention at all 6 (14%) 12 (24%)

Two things set manufacturers apart. Much of what they say about AI is about selling to it, not using it: 13 of the 49 talk about AI as customer demand, such as chips for AI data centers, which we didn't count as an initiative. And AI on the factory floor is rarely disclosed: production, quality and maintenance account for 9 of the 55 internal initiatives named. Supply chain is the most common use (10), followed by company-wide AI programs (7), sales (6), and customer service and marketing (5 each).

Which companies were included?

The study covers 48 insurers that file annual reports with the SEC, are classed by the SEC as insurance carriers, agencies or brokers, and reported between $50 million and $2 billion in revenue for calendar 2025 in the SEC's public financial data. 62 companies were in that band. We left out foreign filers, subsidiaries of other listed insurers, a reinsurer and a shell holding company. Mutual insurers don't file with the SEC, so none is here, and private companies aren't either.

The manufacturers are a random sample of 50 drawn from the 420 companies that file 10-Ks, are classed by the SEC as manufacturers, and reported $50 million to $2 billion in 2025 revenue. Pharmaceutical and biotech companies, publishers and foreign filers were left out before the draw. One sampled company, a used-boat broker the SEC classes as a shipbuilder, is left out of the figures, which cover 49. Eight of the 49 make medical devices. An AI feature in a product a manufacturer sells, such as an AI-enabled device, counts as the product, not as an initiative in its own operations.

Do SEC filings show a company's AI initiatives?

For insurers, mostly yes; for manufacturers and for figures, mostly not. 26 of the 39 insurers that name an AI initiative do so somewhere in their SEC filings. But 16 of the 25 manufacturers that name one do so only outside their filings, and 13 of the 17 insurers that give a figure give it only outside them. So we looked at every public source we could find for each company. For each company we searched its 10-K and 10-Qs, its earnings releases and shareholder letters, its earnings call and investor conference transcripts, its press releases and investor presentations, its own website and investor site, case studies by its vendors, and interviews its executives gave to the trade press. Investor presentations weren't checked for 16 of the manufacturers; the limits below say why that is unlikely to change the counts.

Found in SEC filings Found only outside SEC filings
Insurers naming an AI initiative 26 13
Insurers giving a figure 4 13
Manufacturers naming an AI initiative 9 16
Manufacturers giving a figure 1 7

Earnings calls carried the most figures: 36 of the insurers' rows with a figure come from calls. The richest SEC documents were the glossy annual reports to shareholders and proxy letters, not the 10-K itself.

How were the disclosures graded?

Each row, one source describing one initiative, got one grade for the strongest outcome that source states. The row counts below are the insurers' 768; the manufacturers' rows were graded the same way.

Grade What the source gives Rows
A A cash figure the source sets beside the initiative: dollars, a loss or expense ratio, or revenue. How directly it credits the figure to AI is graded separately, below 19
B Hours, staff or headcount, with a figure 5
C An operational figure: speed, volume, accuracy or share automated 65
D A specific initiative, with no figure 450
E A general statement about AI, with no specific initiative 229

Premium written through an AI tool is graded C, not A: it is volume that went through the tool, not money the tool made or saved. A count of people using a tool, with no result, is graded D: it shows adoption, not an outcome. A company's own statement that it uses AI counts wherever it appears, risk factors included, but a risk factor that mentions the company's use of AI without naming any use doesn't. Every figure here is a claim as published, by the company or, for some figures, by its vendor or the press, and we report it as one.

How often do insurers report a cash return from AI?

Rarely. 6 of the 48 insurers set a cash figure beside an AI initiative, and only 2 credit one directly to AI. Counting each insurer once, by its most direct figure:

How the insurer's most direct cash figure ties to AI Insurers
Credited directly to AI 2
AI named as one of several causes, such as growth, automation, a wider cost program or an accounting change 1
Only figures that can't show an AI effect: a correlation, or a level with nothing to compare it to 3

We count insurers rather than figures because a few insurtechs say most of it: the six AI-native insurtechs and MGA platforms give 17 of the 19 cash figures we found, and Lemonade alone gives 8. Two of the 6 insurers with a cash figure aren't insurtechs, and only one of them, Porch, credits a figure directly to AI.

The two that credit a figure directly:

What do mid-size insurers use AI for?

Underwriting is the most common use, followed by claims, customer service and AI coding assistants. Commercial and specialty carriers most often describe AI that reads broker submissions and ranks them for underwriters.

Use Initiatives
Underwriting and risk selection 27
Claims triage and handling 15
Customer service 15
AI coding assistants 14
Workplace AI assistants 13
Pricing and rating models 8
Online quoting and binding 7
Claim file summaries and letters 7
Submission intake and triage 6

Can a company benchmark its AI return against peers' disclosures?

Not yet: an insurer or manufacturer whose AI work matches its peers' still has almost no public outcome to compare against. The 4% of insurers that credit a cash figure directly to AI each measure it differently, over different periods, and none against a control. A peer's figure shows what the peer chose to say, and little about what its AI returned. Large companies' annual reports are no better: none of 498 S&P 500 annual reports states an achieved, quantified return from the company's own AI use. The comparison that holds up is a company's own before-and-after, measured the same way on both sides.

How was the research done?

AI research agents (Claude, made by Anthropic), working under the direction of Roiva staff to a written protocol, searched the sources above for each company in the first week of November 2026, and recorded each initiative with a quote of 40 words or fewer and a link to its source. On November 10 and 11 they went back over the insurers: they re-read in full every earnings call from July 2025 on, because some transcripts in the first pass were incomplete, searched each source type again, and extended the search back to January 2024, adding investor conference talks, releases published on other sites, and letters on investor websites. On November 13 they re-read the AI sentences they had set aside as not describing the company's own use, under the risk-factor rule above, which added five rows to initiatives already found. A separate AI agent then re-researched a sample of insurers in three blind audits, and what the audits found had been missed was added. Reviews removed rows whose source doesn't say the work or the figure came from AI, such as rules-based automation or tools never called AI, and regraded some cash figures as shared. Every quote was checked against its source page: 731 of the insurers' 768 by software that matched the quote's text on the page, 11 from PDF text, and 26 by a separate AI agent that hadn't done the research, which corrected the wording of one. The manufacturers were researched the same way on November 7, 2026, and their earnings calls were re-read in full on November 10, which added 50 rows. All 167 of their quotes were found on their source pages, 151 by software and 16 by a separate AI agent, which corrected the wording of one. The grades and the mapping to uses follow written rules and haven't yet been reviewed by a person. A few rows may describe the same deployment from both the company's side and its vendor's.

What are the limits?

This is a pilot sample of 48 US-listed insurers, not a census, and the shares above have wide margins: 4% of 48 is two companies. Six of the 48 are AI-native insurtechs or MGA platforms, and they supply 17 of the 19 cash figures. Mutual insurers, which file no SEC reports, would likely disclose less. Nearly half of the 257 initiatives (120) were first disclosed in 2026, which is less than a third of the period searched, so disclosure is rising quickly, and a later look will find more.

The manufacturing sample is 49 companies, so 4% is two. iRhythm's algorithm is part of a product it sells, but we counted it as an internal initiative because the savings it expects are its own technicians' time. Investor presentations weren't checked for 16 of the manufacturers. Among the insurers, a presentation was never the only place a company named an initiative.

Has this study been corrected?

Seven times, all on November 17, 2026. The figures above are the corrected ones.

  • A figure credited to AI in general isn't a figure for an AI initiative (November 17). After the AI in Practice Index launched, checking agents re-read the figures at its leaders that the earlier re-reads hadn't covered. A figure a company credits to AI in general, rather than to a named use of AI, now counts as a general statement about AI. Three cash figures no longer count: Ategrity's expense ratio, of which it said "Technology and AI are a big part of the story"; Employers Holdings' underwriting expense ratio, credited to "investments in automation, customer self-service capabilities, and artificial intelligence"; and Baldwin's forecast savings from a program with AI as one of several levers. Baldwin's productivity figure credited to "the tools we are deploying internally" no longer counts either, and three trade-press accounts of eHealth's and Palomar's figures were dropped, because trade press counts only an executive's or spokesperson's own words. Insurers that set a cash figure beside an AI initiative went from 12 to 9: those crediting AI as one of several causes from 6 to 3, and those that aren't insurtechs from 7 to 4. Insurers that give a figure in their SEC filings went from 6 to 5, and those that give one only outside them from 12 to 13. Rows went from 838 to 836 (grade A from 29 to 26, C from 73 to 69, D from 526 to 528 and E from 202 to 205), rows with a figure from earnings calls from 45 to 42, and initiatives from 290 to 288, of which those first disclosed in 2026 went from 139 to 137 and company-wide AI programs from 13 to 11. No insurer moved in or out of naming an initiative or giving a figure, and the shares in the first table didn't change. Ategrity no longer meets the index's leader rule, so the Mid-Size Insurers 2026 edition now names eight leaders.
  • General statements and governance practices read on their sources (November 17). After the AI in Practice Index launched, checking agents read on their sources the governance practices and general statements about AI that the earlier re-reads hadn't covered. A general statement now counts only when it says what the company is doing with AI now: plans, hopes, hypotheticals and AI as an industry trend no longer count, and neither do risk sentences that name no use, trade-press writers' words or quotes that don't name AI. No insurer moved in or out of naming an initiative or giving a figure, and no share in this note changed. Rows went from 836 to 816 (grade B from 8 to 7, C from 69 to 70, D from 528 to 518 and E from 205 to 195), and initiatives from 288 to 286, of which those first disclosed in 2026 went from 137 to 136. The index's leader list didn't change.
  • Every leader item read a second time (November 17). After the AI in Practice Index launched, a second reader, working without the first reading, read again every item behind its leaders that could be opened. Plans and hopes were dropped, programs and training recorded as uses became general statements, and quotes, statuses and dates were brought to the sources' words. No insurer moved in or out of naming an initiative or giving a figure, and no share in this note changed. Rows went from 816 to 800 (grade D from 518 to 505 and E from 195 to 192), and initiatives from 286 to 283, of which company-wide AI programs went from 11 to 9; with dates corrected, those first disclosed in 2026 went from 136 to 137. The index's leader list for insurers didn't change.
  • A blind sample read again (November 17). A separate reader read again a random 10% sample of every item the AI in Practice Index counts, 199 items, without seeing earlier readings, and the errors it found were corrected. Among the insurers, SelectQuote's forecast of $30 million credited to "the AI technology efforts" names no specific use of AI, and eHealth's AI "within product management, software development, and UX design" names functions, not the work AI does; both are now general statements. Neptune's revenue per employee, set beside its "AI-first approach", which names no AI tool or work, no longer counts, and a Hippo statement recorded twice from the same call is now counted once. No insurer moved in or out of naming an initiative, giving a figure or giving a cash figure, and no share in this note changed. Rows went from 800 to 793 (grade A from 26 to 24, C from 70 to 68, D from 505 to 503 and E from 192 to 191), and rows with a figure from earnings calls from 42 to 38. Cash figures from AI-native insurtechs and MGA platforms are now 20 of 24. Initiatives went from 283 to 281: AI coding assistants from 19 to 18 and claims triage from 17 to 16; with dates corrected, those first disclosed in 2026 went from 137 to 138. The index's leader list for insurers didn't change.
  • Leaders' uses checked against a sharper test (November 17). After the AI in Practice Index launched, a reader checked every use behind its leaders against a sharper test of what a deployment is. That test applies from the next edition; for these editions we corrected only what the rules they were published under already rule out: plans and hopes, AI as an industry trend, and AI in a product a company sells. Among the insurers, rows went from 793 to 789 (grade D from 503 to 499). No insurer moved in or out of naming an initiative or giving a figure, and no share or other figure in this note changed.
  • The other companies read a second time (November 17). After the AI in Practice Index launched, a second blind sample was read again, and then a second reader, working without the first reading, read again every use and governance practice at the companies that aren't among its leaders. Writers' words, figures credited to AI in general and plans were corrected or dropped. Among the insurers, Lemonade's pet cost per claim of $14 is a level with no before-and-after in the words quoted, so it no longer counts as a figure credited directly to AI; Accelerant's "up to 300%" recoveries rested on a trade-press writer's narration, and its members' 2 to 3 point loss-ratio improvement is a whole-book ratio with other causes; and figures at Hippo, Trupanion, SelectQuote and Clover were set in general statements or dropped. Insurers that give any figure went from 18 (38%) to 17 (35%), and those that credit a cash figure directly to AI from 3 (6%) to 2 (4%). Insurers with a cash figure of any kind went from 9 to 6 (crediting AI as one of several causes from 3 to 1), and among the 42 that aren't AI-native, any figure from 12 (29%) to 11 (26%). Insurers giving a figure in SEC filings went from 5 to 4. Rows went from 789 to 768 (grade A from 24 to 19, B from 7 to 5, D from 499 to 449 and E from 191 to 227), rows with a figure from earnings calls from 38 to 36, and initiatives from 281 to 258, of which those first disclosed in 2026 from 138 to 120. The index's leader list for insurers didn't change.
  • A third blind sample (November 17). A separate reader read again a random sample of the items at the companies that aren't among the index's leaders, after their second reading. Two of Trupanion's rows became general statements, and operational figures that show only a correlation no longer count as figures. No insurer moved in or out of naming an initiative or giving a figure, and no share in this note changed. Grade C rows went from 68 to 65, D from 449 to 450 and E from 227 to 229, and initiatives from 258 to 257, of which AI coding assistants from 15 to 14.

Questions about the data?

Ask us for the rows, the sources or the grading rules.

Get the next one

New field notes by email as they're published. Or subscribe to the RSS feed.