By the end of the 2026 interim reporting season, the five A-share listed insurers had, for the first time, translated AI from a "strategic narrative" into quantitative entries in their financial statements: token consumption, agent counts, AI call volumes, and AI seat coverage.

In the past, AI disclosures from insurers were largely directional statements from executives about "enabling" or "transformation." This reporting cycle marks a clear departure. Ping An Insurance (Group) Company of China, Ltd. (SSE: 601318) disclosed its average daily token consumption, which surpassed 120 billion in June 2026, up from roughly 30 billion in December 2025—a fourfold increase in six months—before exceeding 210 billion in August. China Life Insurance Co., Ltd. (SSE: 601628) reported its agent count, PICC Group (SSE: 601319) detailed call volumes, and New China Life Insurance Co., Ltd. (SSE: 601336) listed the volume of auto-generated plans.

At a results briefing, Qin Hongbo, vice president of New China Life, shared a telling anecdote: when asked what resources young employees needed most, the unanimous answer was "more token allocations." Amidst all these figures sitting alongside income statements, a critical question lingers: after burning through all these tokens, what has actually been gained?

How the "AI Balance Sheets" Differ Across the Five Insurers

The disclosure density and metrics vary significantly among the five insurers. Ping An presents the most complete picture: average daily token consumption grew fourfold in six months; AI seats handled approximately 939 million service interactions, covering 81% of total customer service volume; AI agents facilitated RMB 57.313 billion in sales; intelligent anti-fraud claims interception saved RMB 7.11 billion in losses; and AI-generated code for internal staff rose from 30% to 62% since the start of the year. These numbers span five business lines—customer service, marketing, claims, risk control, and R&D—demonstrating that AI at Ping An has evolved beyond a tool for specific functions into foundational infrastructure across its entire operational chain.

China Life, PICC, and New China Life each use different measurement approaches. China Life has built over 90 digital application scenarios and deployed more than 500 agents, reducing claims processing time to 0.36 days. However, unlike Ping An, China Life does not break down where these 500 agents operate in terms of customer service coverage or sales contribution. PICC rolled out 246 AI application scenarios, with over 2.3 billion AI capability calls in the first half. New China Life's marketing assistant auto-generated over 4 million proposals, while its digital employee for group insurance bidding improved recommendation accuracy from 60% to 90%.

China Pacific Insurance (Group) Co., Ltd. (SSE: 601601) stands out as the only one of the five without specific quantitative metrics. It lists "AI+" as one of its three core strategies, but the interim report contains only qualitative statements, such as "building an enterprise-grade knowledge system and management platform," with no token counts, agent numbers, or call volumes. President Zhao Yonggang's description—“playing a positive role in enhancing marketing staff productivity and comprehensive operational capability”—remains qualitative, not quantitative.

Four insurers provided numbers; one spoke only of strategy. AI has moved from auxiliary functions like customer service and office work into the most money-sensitive areas: underwriting, claims assessment, and anti-fraud interception. Yet each company uses a different yardstick, making it difficult for the market to compare or assess which insurer is truly converting AI into profits.

Tokens Prove Penetration, Not Profit Quality

Comparing token metrics side-by-side with income statements reveals a counterintuitive result: Ping An, with the most aggressive token consumption and most complete AI narrative, ranked fourth among the five insurers in first-half net profit growth, ahead of only China Pacific Insurance. Meanwhile, New China Life, which raised its equity and fund allocation to 25.6%, posted the highest annualized total investment return at 6.7%.

The profits came from a different source. In the first half of 2026, the five listed insurers reported combined net profit attributable to shareholders of RMB 317.387 billion, up 78.12% year-on-year, with total investment income reaching RMB 641.32 billion, up 74.57%. China Life led with total investment income of RMB 314.504 billion (up 146.7%), corresponding to net profit of RMB 134.489 billion (up 228.6%). Ping An, PICC, China Pacific Insurance, and New China Life reported investment income of RMB 136.942 billion, RMB 66.327 billion, RMB 66.022 billion, and RMB 57.525 billion, respectively. The trajectory of investment income gains aligns closely with profit growth—the primary driver of this earnings surge is the rising equity market, amplified by new accounting standards that place more equity assets under fair value changes.

The money earned this half-year came primarily from the market, not from AI.

Consider Ping An's most closely watched figure: the RMB 57.313 billion in sales "assisted" by AI agents. The operative word here is "assisted." This metric captures the sales scale facilitated by AI involvement in certain processes—a business volume measure, not newly recognized incremental revenue or profit. It does not subtract token, computing, or labor costs. Ping An's first-half net profit of RMB 92.585 billion sits in a wholly different accounting category.

The denominator is also missing. Ping An does not itemize token spending as an auditable financial line item. Chief Technology Officer Wang Xiaohang's statement—“Ping An invests over RMB 10 billion annually in technology”—positions AI as one of three strategic tracks (comprehensive finance, healthcare, and internal efficiency), with token costs distributed across R&D expenses, operating expenses, and computing depreciation. The closest available reference to “cost” is the claim that through context compression, model inference optimization, heterogeneous multi-GPU clusters, and peak/off-peak scheduling, per-token costs have been reduced by more than half.

Yet even with unit costs halved, absolute spending remains undisclosed—how much it costs per 100 million tokens consumed, and what returns it generates, cannot be calculated. In response to “burning cash” concerns, Chief Financial Officer Fu Xin stated that while AI investment is increasing, the group's overall cost-to-income ratio continues to improve year-on-year. On the marketing front, Ping An's AI tool AskBob is used daily by 39% of agents, but whether AI's contribution shortens a policy inquiry or closes an additional policy is not separately disclosed.

The industry's only disclosure tied directly to a profit metric comes from Ping An Healthcare and Technology (SEHK: 1833): in the first half of 2026, AI contributed 4.6% of the company's gross profit. Only when contributions flow into the gross profit line does AI's ledger become verifiable. The other insurers' AI metrics largely stop at “how much was used” without reaching “how much was earned.”

From Zero-Sum Cost Reduction to Positive-Sum Repricing

The AI ledger from this reporting season can be broken into three categories. The first is labor replacement. AI seats handled 939 million interactions covering 81% of customer service, AI-generated code reached 62%, and “AI inquiry-handling” covers 88% of business scenarios. These metrics translate into reduced work hours and headcount—the easiest area for AI to generate cost savings.

The second is loss reduction. Ping An's property and casualty arm intercepted RMB 7.11 billion in claims through intelligent anti-fraud, with reduced payouts flowing directly into underwriting profits.

The third, sales, is the hardest to verify. With RMB 57.313 billion in agent-assisted sales, whether AI shortened a single policy inquiry or closed an additional policy is not disclosed by the company, making external verification impossible.

These three categories correspond to two distinct types of AI value. Labor replacement is a zero-sum cost reduction—the savings merely redistribute funds within the insurer from expenses to profits, creating no new value for the industry, often at the cost of layoffs. Loss reduction and novel risk pricing, by contrast, are positive-sum creations—the RMB 7.11 billion saved is real underwriting profit, and using AI to capture climate, tail, and individualized risks represents genuine value creation for the insurance sector.

Currently, the vast majority of tokens are allocated to the first category. The second and third are where AI's true potential lies.

None of the four metrics—token consumption, agent counts, AI call volumes, or assisted sales—constitute audited financial data, nor does regulation require their disclosure. By recasting AI in quantitative terms in the same reporting season, the five insurers are signaling to a market where AI narratives have become a core valuation variable, vying for the premium of being the “AI transformation benchmark.” The absence of unified standards has led each company to create its own metrics, often packaging input metrics as output metrics—this “metric arbitrage” is the genuine reason the calibers are incomparable.

Three developments warrant close attention going forward. First, whether metrics will be standardized: with four distinct algorithms in play, whether annual reports consolidate AI spending and returns into clearer accounting categories matters more than the metrics themselves. Second, the cost-to-income ratio: the substitution effects in customer service, claims, data entry, and code generation must ultimately appear in the expense line. Whether this ratio continues to improve in the 2026 annual report is the first hard constraint on whether the “AI-replaces-labor” equation balances. Third, whether evaluation standards shift: according to a representative from Nuanwa Technology, insurers have moved from treating AI as “optional” to “mandatory.” Once leading institutions complete their infrastructure build-outs, assessment criteria will evolve from “how many scenarios were deployed” to “how much was saved, how much less was paid out, and how many more conversions were generated.”

Zhang Xinyu, assistant president of China Life, noted at the results briefing that the company's long business cycles and comprehensive operations have accumulated long-cycle data assets, “which will play an even greater role in the AI era.” Technology providers serving insurers are also upgrading rapidly—Nuanwa Technology recently completed model upgrades across three product lines: AI claims, AI customer operations, and AI risk control. Investment in AI within the insurance sector is expanding from insurers themselves to the broader ecosystem.

Zhang Ning, director of the China Fintech Research Center at Central University of Finance and Economics, offers a deeper perspective: digital intelligence will first impact insurance's core foundation—the actuarial pricing system. Traditional statistical models cannot cover all risk forms; AI gives the industry the ability to capture and manage novel risks. These pricing-side changes will not appear in token numbers or on first-half income statements.

AI's true positioning at insurers today is “defensive,” not “offensive.” Profit surges driven by equity markets on the asset side are the offense—provided by the market. AI serves as a cost-side defensive tool, helping insurers lose less, pay out less, and employ fewer people in a low-rate, spread-loss environment, but it does not generate additional earnings. Tokens in financial reports appear as technology narratives but are, in substance, defensive cost management.

When insurers put tokens into their financial statements, it signals that AI usage has grown significant enough to warrant disclosure. The high growth on income statements indicates that the half-year profits were primarily market-driven. The gap between these two realities is the true distance this AI race must cover. And when that gap is measured, companies that dressed up input metrics as output metrics will be the first to be exposed.