Ping An AI Agents in Insurance, Banking, and Healthcare

Ping An did in fact report that in the first half of 2026, AI agents helped generate 57.313 billion yuan in sales, and AI operators handled about 939 million service interactions — 81% of customer service volume. These are AI-assisted sales, not proven incremental revenue from fully autonomous transactions. The source is the group’s official disclosure dated August 20, 2026. Ping An’s First-Half 2026 Results.

Ping An Insurance (Group) Company of China, Ltd. is a Chinese financial and insurance group. Those two figures tell only part of a much broader story. Ping An combines insurance, banking, investments, medical services, and senior care. That is why AI here should be viewed at the same time as a customer interface, an employee assistant, a risk analytics tool, and a participant in document processing.

Prepared by AIrassvet using publicly available sources, verified on September 30, 2026. It is intended for business leaders, insurance and financial companies, product teams, and automation specialists. Below, we break down publicly described use cases; the author does not have access to Ping An’s internal systems. This is an analysis of corporate disclosures, not medical advice or an assessment of the stock’s attractiveness. The AIrassvet section describes the publisher’s services.

Contents

• Is the 57.3 billion yuan and 81% service claim true?
• How to verify Ping An’s figures
• AI use map across the group
• Technology foundation
• Sales and agent operations
• Express Service and customer support
• Underwriting and policy issuance
• Claims, loss assessment, and fraud
• EagleX and disaster warning
• AI at Ping An Bank
• Healthcare and AI Doctor
• In-home senior care
• Development, documents, and recruiting
• Agent management and personal data
• AI economics and KPI limitations
• What businesses can learn from Ping An
• Frequently asked questions
• Primary sources and verification method

Is the 57.3 billion yuan and 81% service claim true?

Yes, those metrics were published by the company itself. Clarifying the wording changes the meaning: “AI helped generate sales” describes technology’s role in the commercial process. By itself, it does not show how many purchases would not have happened without AI or how much of the work was done by employees.

A note in the full report states that the service metric includes inbound and outbound voice and text bot interactions in credit cards and insurance. Ping An report, page 37.

The service share refers to the number of interactions. One customer may call several times, send messages, receive an automated reply, and then speak with a human. So interaction count cannot be substituted for unique customer count. Likewise, handling a request does not always mean the entire issue was finally resolved.

A correct news headline would be: from January to June 2026, AI participated in Ping An’s sales at the stated amount and handled most of the measured service volume. These data confirm the scale of adoption, but they do not allow us to calculate AI’s causal contribution to profit. Ping An’s official first-half responses.

Is it really more than $8 billion?

For a reproducible conversion, we should use a specific date rather than an undefined “current exchange rate.” As of September 4, 2026, the official central parity rate was 6.7787 yuan per dollar. This value was published on the website of China’s Ministry of Commerce with a reference to the People’s Bank of China. Dividing 57.313 billion by 6.7787 gives approximately $8.455 billion. Exchange-rate publication for September 4, 2026.

This is an equivalent-value calculation, not U.S. dollar revenue recognized by Ping An. It does not account for transaction-specific exchange rates and does not use the average rate for the reporting half-year. For the claim “more than $8 billion,” the chosen rate is sufficient; an accounting comparison would require a different methodology.

Why you should not say “AI earned the company 57.3 billion”

That wording mixes turnover, revenue, profit, and incremental effect. Even if an assistant was involved in every sale, humans may have developed the product, other channels may have acquired the customer, and an employee may have closed the deal. AI may have helped choose an offer or explain the terms while remaining only one part of the chain.

To evaluate the incremental result, you would need comparable customer groups or another sound way to separate AI’s effect from changes in products, pricing, advertising spend, and sales rep quality. The public disclosure of these two KPIs does not provide that proof. That is why precise attribution matters more than a flashy headline.

How to verify Ping An’s figures

Every figure should have four labels: who reported it, for what period, what was measured, and compared against what baseline. If even one label is missing, readers risk comparing different things. In a large corporate AI case, this is especially easy: service messages, completed tasks, tokens, insurance premiums, and prevented claims all appear side by side.

Metric	Value and period	What it describes
AI-assisted sales	57.313 billion yuan, first half of 2026	Sales volume involving AI agents
Service handled by AI operators	About 939 million interactions, 81%, first half of 2026	Service volume, not unique customers
Fast life insurance claims payouts	59% of claims, first half of 2026	The share of cases in the fast-claims service
Intelligent auto policy issuance	94% in the dealer channel, averaging within one minute, first half of 2026	A specific channel and issuance operation
Antifraud in Ping An P&C	7.11 billion yuan in claims savings, first half of 2026	The reported reduction in payouts due to fraud detection
Model consumption	More than 120 billion tokens per day in June 2026	The scale of model activity, without measuring the value of each call

The source for all table rows is Ping An’s first-half 2026 disclosure. The table reflects the company’s statements, not independent testing results for each AI product.

Which comparisons make sense

The share of automated policy issuance is useful if we know which products and channels it applies to. Processing time should be compared for identical operations: document reading, preliminary review, contract issuance, and the full customer journey are different tasks. Even the word “automatically” needs explanation: a system may handle standard operations while keeping humans for exceptions.

Financial benefit metrics also differ. Saved payouts are not product sales revenue. Prevented losses do not necessarily equal lower insurer payouts: part of the loss might have been borne by the customer. Lower token cost describes computing infrastructure, but not the full service cost.

Why the Number of Agents Does Not Equal the Number of Autonomous Workers

An internal AI application can be a small assistant for a single document. Another agent may coordinate multiple systems and participants. You cannot determine the level of autonomy, task complexity, or the share of use in industrial operations from one overall application count.

It is useful to ask what permissions an agent has and where its work ends. Did it only prepare a response? Fill out a form? Review documents? Record a decision in the core system? Send money? These actions have completely different economic value and error costs.

AI Use Map Across the Group

To avoid mixing the functions of different business units, a navigation map is provided below. It brings together the areas described in the source materials cited later. The detailed action sequences in the following sections are an analytical explanation of the business logic, not a published technical specification from Ping An.

Area	Publicly Described Role of AI	Where Separate Oversight Is Needed
Insurance Sales	Sales support, needs analysis, recommendations	Product accuracy and explanation of terms
Customer Service	Voice and text responses, service initiation	Identification and escalation of complex requests
Underwriting	Risk analysis and faster review	Exceptions and material decisions
Policy Issuance	Document handling and processing	Data mismatches and nonstandard documents
Claims Handling	Faster claims submission and review	Disputed circumstances and incomplete materials
Antifraud	Detection of suspicious cases	Review of suspicion and customer entitlement
Disaster Prevention	Risk models and alerts	Forecast quality and practical guidance
Banking	Marketing, credit processes, operational checks	Financial risks, approvals, and compliance
Healthcare	Consultations, screening, physician support	Clinical assessment and patient safety
Elder Care	Sensors, alerts, concierge support	Availability of help and alert verification
Internal Processes	Code, documents, recruiting	Verification of output, data, and evaluation criteria

There is no single universal “robot” in this map. In some tasks, language and dialogue matter; in others, statistical forecasting, image recognition, or rule execution matter. The word “AI” combines technologies that operate at different stages of the same process.

Technology Foundation

Data and Internal Platform

According to Ping An's 2025 disclosures, more than 230,000 employees used its internal AI agent platform, on which more than 70,000 AI applications were created. The company also reported 3.65 billion model calls during the year, a corpus of more than 3.2 trillion text tokens, 500,000 hours of labeled speech, and more than 8.5 billion images. Ping An 2025 Results.

These figures describe the infrastructure base. They do not show that every employee created an autonomous agent, that every application underwent the same testing, or that all images were used to train one model. The reasonable conclusion is much narrower: AI is available to a large number of business units and is used through a shared platform.

For business, the value of this setup lies in component reuse. If several teams work with similar documents, they do not need to build their own search, identification, and activity log each time. At the same time, a unified platform must preserve access boundaries: a shared technical base does not mean shared access to all data.

PingAnGPT-Qwen3-32B Financial Model

In a March 2026 announcement, Ping An named its financial LLM PingAnGPT-Qwen3-32B and said it ranked first in CNFinBench as of that date. The benchmark covers financial knowledge, business understanding, computation, compliance, and application safety. The company cited 97 implemented business scenarios, including auto insurance claims notifications, customer service, expense audits, and smart telephony. Announcement on the financial model.

LLM stands for large language model, which processes text and generates responses. Financial specialization helps it work with industry concepts and documents. However, the model itself is not a complete banking or insurance system: it needs tools, access rights, and verifiable rules to take action.

The ranking is useful as a limited evaluation of a specific task set. It does not prove leadership across all financial operations or guarantee the quality of any particular implementation. It is more important to separately check how the model works with real documents, incomplete data, and access to transactions.

Open Models and Proprietary Industry Expertise

Ping An described using DeepSeek in Ping An Health and combining DeepSeek, Qwen, and OneConnect's proprietary agent platform. For external financial organizations, it mentioned private cloud, model compression, and distillation. Ping An material on generative AI.

The practical principle that follows is this: a company can build application-level expertise on top of an available base model. Distillation — transferring part of a more complex model's capabilities into another model; it can help reduce compute requirements. But the final suitability depends on which tasks and constraints are preserved after that transfer.

In a corporate environment, it is important to separate the model, knowledge, rules, and operations. The model interprets the request; up-to-date documents provide the factual basis; rules define the permitted action; and the business system executes it. This framework is an analytical explanation of how such solutions are structured, not a claim that every Ping An product is implemented the same way.

Insurance Sales and Agent Work

AskBob: AI Beside the Sales Rep

In a corporate podcast, Ping An describes AskBob as a support tool for its agent network. The published transcript mentions task management, intelligent responses, and training; for 2024, it reported more than 13 million monthly queries and use by more than 95% of insurance agents. These are historical figures for AskBob specifically and should not be confused with the group's overall results for 2026. Insurance podcast transcript.

The practical role of such an assistant is easy to understand: a sales rep works with a large number of products, rules, and customer situations. They need quick access to current information, the ability to clarify terms, and preparation for the conversation. Part of the value arises before contact with the customer; another part comes when explaining a complex product.

Consider a typical hypothetical example. A customer asks whether coverage applies to an already diagnosed condition. The assistant should find the relevant terms and exclusions, not invent a sales-friendly answer. If the information is insufficient, the next step is clarification or escalation to a specialist.

Needs Analysis and Personalized Offers

Ping An describes building a customer profile, using CRM, and helping the sales rep based on data. This approach connects information about a person's needs with offers from the insurer and other group businesses. Ping An podcast on digital transformation in insurance.

In the analytical model, the process consists of several parts: identify the need, check suitable products, prepare an explanation, hold the conversation, and record the outcome. AI can take part in each part with different degrees of independence. That is why the overall AI-assisted sales metric covers heterogeneous activities.

Personalization is useful only when the underlying data is accurate. If the system gets the client’s life situation wrong, good language alone will not fix a bad recommendation. To implement it, you need to separately verify profile quality, the relevance of product rules, and the basis for using the data.

Reinstating lapsed policies

In its 2025 results, the company reported a 30% increase in the number of reinstated policies when using intelligent task allocation between AI and humans. This metric refers to coverage reinstatement, not overall growth in insurance sales. Ping An annual results.

This example is interesting as a way to organize work. Reinstating a policy may require different actions: send a reminder, explain the process, find out why coverage ended, and help with documents. Simple tasks can be routed to an automated channel, while ambiguous ones go to an employee.

For another company, we should transfer not the percentage, but the queue management approach. The selection rules for who handles the task, the quality of context transfer, and task completion all matter. If an employee has to reread the entire conversation from scratch, part of the expected savings disappears.

Why assistive selling and autonomous selling are different products

An employee assistant usually prepares information that a person reviews before sending it to the client. An autonomous seller communicates directly and can influence the decision without that intermediate review. For it, recommendation limits, consent checks, and full disclosure of terms are especially important.

When designing the system, it is worth separately defining which words the agent is allowed to say as a verified fact. Coverage terms, exclusions, and claims procedures should be based on the current document. Recommendations can be evaluated for accuracy and suitability for the client, not just conversion.

Express Service and customer service

One request instead of searching across apps

In its May article, Ping An describes Express Service as a conversational entry point to the group’s different services. The client states a need, and the system connects actions across apps. Examples include moving to medical help when feeling unwell and coordinating assistance, insurance, and car replacement after an accident. Express Service description.

The main change is shifting navigation to the system. The client does not need to know first which subsidiary owns the needed service. For the business, this means linking not only interfaces but also task state: what has already been done, what information has been obtained, and who is responsible for the next step.

A hypothetical customer journey after a car accident can be described as follows: check whether the people are safe, identify the need for help, find the insurance policy, register the claim, and send the materials for review. This is an instructional sequence, not a guarantee that Express Service automatically performs every listed step in every case.

How an agent moves from conversation to action

In a 2026 podcast, Ping An lays out the sequence: intent recognition, identity and entitlement checks, system coordination, compliance control, and task progression. It also lists four levels of control: identification, controlled access to tools, human involvement in risky actions, and audit. Podcast on Express Service technology.

An AI agent in this context is a system that uses a model to choose steps and calls approved tools. The defining trait of agency is the ability to move a task toward completion. But autonomy always has limits: the system may ask for confirmation, stop, or hand the work to a person.

For design purposes, it is important to distinguish between “understood the request” and “performed the operation.” Between them are user verification, data retrieval, condition checks, and result recording. Each stage should have a clear error, a retry option, and a way to know whether the operation finished.

Express Service scale and the meaning of a completed task

In a September statement by the CTO, the company said it supported more than 300 services, handled over 100 million interactions in the first four months, achieved 92% end-to-end resolution, and cut average handling time by more than half. Ping An CTO statement.

Here, the difference between an interaction and a completed task is important. Several clarifications within one request can create multiple messages. A successful answer to a question is not the same as a completed transfer or a resolved claim. That is why the completion metric needs to be read within the company’s methodology, which is not fully disclosed in the brief statement.

In your own project, it is useful to define a specific criterion: data changed in the core system, an application created, a document received, or the client received a confirmed status. This makes success measurable. For such processes, the metric “the bot responded” is not enough.

Voice and text agents

For 2025, Ping An reported about 1.702 billion AI agent interactions and a 80% share of service volume. Annual disclosure. The figures for the following half-year are given at the beginning of the article; you cannot compare annual volume with half-year volume without adjusting for the time window.

Analytically, this type of channel work includes request recognition, answer retrieval, collection of missing data, and routing. A voice interface adds another layer: it has to recognize speech correctly and understand what belongs to the current question. An error in a name or policy number can affect the downstream handling.

Good human handoff preserves the history and the reason for escalation. If the client has to explain the situation again, automation makes the journey worse. That is why service quality should also be measured by repeat contacts, complaints, and completeness of resolution, not just the share of automated answers.

Underwriting and policy issuance

What underwriting actually evaluates

Underwriting is the assessment of risk before a contract is issued: whether to accept it and on what terms. Ping An reports digital underwriting using big data, AI, and machine learning; in 2025, 94% of life insurance policies were processed in seconds. Ping An technology overview.

In a typical insurance process, the system matches the application, product rules, and the information needed for evaluation. Some cases fit predetermined conditions. Others require more data or expert review. Fast standard processing does not mean all applications are equally simple.

It is important to separate risk prediction from the authority to make the decision. A model can estimate the likelihood of an event, but the allowable contract terms are also determined by the product and company requirements. For complex situations, there must be a way to explain the basis of the decision, correct inaccurate data, and review the result.

Intelligent underwriting in P&C

In its July WAIC 2026 material, Ping An P&C said average initial review time had fallen to about 1.5 hours, daily processing volume per underwriter had doubled, and the risk interception rate had increased by 16%. WAIC solutions coverage.

These are separate metrics for a specific business unit. They cannot be combined with life insurance second-level underwriting into one overall Ping An speed figure. Also, the wording “increased by 16%” does not justify writing “by 16 percentage points.”

For a practical assessment, you need details on the composition of applications and the measurement stages. A preliminary review may end with a request for additional materials. If you count only model processing time and exclude document waiting time, the metric will not describe the full timeline for the customer.

Documents and intelligent auto policy issuance

Ping An describes the use of multimodal perception and reasoning to process unstructured and nonstandard documents. In its 2025 technology overview, it says 93% of dealer auto policies were intelligently issued in an average of under one minute. Technology overview.

A multimodal system works with multiple types of data, such as a document image and text. In a training setup, it recognizes fields, matches them across documents, and checks whether there is enough information to complete issuance. The actual issuance must then follow the rules of the insurance system.

This kind of solution has clear exceptions: blurry images, mismatched numbers, unusual document formats, or missing required fields. That is why the interface needs to allow corrections, and the log should show which data were extracted automatically. A confident model response does not replace review of the source document.

Why do one minute and 1.2 minutes both appear?

The July disclosure describes an auto policy issuance robot for new cars: 93% of cases processed automatically, with average time reduced from six minutes to 1.2 minutes. WAIC 2026. In the semiannual release, the wording used for the dealer channel is the one shown in the table above.

These messages differ in sample description and in how the operation is phrased. Until the methodologies are aligned, you cannot say one number is a correction of the other. For readers, it is more useful to keep both captions than to pick the prettier speed.

Claims, damage assessment, and fraud

Fast payouts and the 111 Quick Claims service

Ping An describes 111 Quick Claims as a one-sentence case report, one-step document upload, and minute-level review. This is a customer service workflow, not a universal promise to pay out money within a minute for any event. Ping An technology page.

To understand the process, it helps to think of four separate outcomes: the case is registered, documents are received, eligibility is checked, and payment is made. AI can speed up the first stages, but the final result depends on the completeness of the materials and the coverage terms. The more complex the situation, the more important specialist involvement becomes.

In claims automation, it is especially useful to check whether the documents match the reported event and the active policy. If information is incomplete, the system should request a specific addition. A “denied” message without a clear reason makes fast service difficult for the customer to verify.

Non-auto claims and automated review

On its technology page, Ping An says that in 2025 it processed nearly one million non-auto cases with end-to-end automation. The same page also states that 70% of personal injury claims were settled automatically and that the fastest case took about 51 seconds. Ping An overview.

The fastest result is not the average. That distinction matters because an uncommon simple case can move much faster than a typical one. To compare services, you need the distribution of processing times and the share of cases that required additional documents.

When designing the system, it is useful to collect the reasons for manual intervention separately. Some are related to photo quality, others to event verification, and others to policy terms. This breakdown shows where a stronger model is actually needed and where it is enough to fix the form or data access.

Computer vision and vehicle damage

Ping An describes AI-based damage assessment from images as a way to reduce processing from hours to minutes. Material on generative AI and applied technologies. This confirms the direction of use; this source does not support a universal accuracy claim for assessing any damage.

The explanatory workflow looks like this: receive images, identify visible damage, link it to vehicle parts, and pass the results into pricing or review. It is useful to separate recognition of the visual feature from estimation of repair cost. The second result also depends on prices, parts availability, and hidden damage.

The full first-half 2026 report adds a specific metric: for 57% of auto insurance claims, AI-enabled automated inspection was performed after the owner uploaded photos of the incident. This is the share of such reviews, not of all fully automated payouts. Ping An report, page 36.

A photo does not show everything. Poor lighting, reflections, covered body panels, and internal damage can limit assessment. That is why photo-taking instructions and rules for escalating to a specialist inspection are needed. These limitations apply to the technology as a whole and are not a statement about the error rate of any specific Ping An product.

Antifraud: a risk signal is not proven fraud

Antifraud means identifying signs of possible fraud. For 2025, Ping An P&C reported 10.51 billion yuan in claims savings from intelligent fraud detection. Annual results. The 2026 half-year figure is shown in the overall table.

Saved claims payments cannot be called “stolen money recovered by AI” without verification. This is a company methodology measure of prevented or reduced payouts. An independent assessment requires calculation rules, decision reviews, and information on false suspicions.

Analytically, possible signals include mismatched documents, unusual event sequences, or links between repeated claims. But the public reference to smart fraud detection does not reveal the exact set of features in a specific model. So the list should be understood as an explanation of the task category.

From a practical standpoint, the system should generate a verifiable signal and preserve evidence. If a suspicion turns into a denial without further review, the cost of a false positive rises. Oversight requires not only prevented payouts but also the quality of decisions for honest customers.

EagleX and disaster warnings

The insurer starts acting before a loss occurs

EagleX is described as a platform that combines big data, satellite observation, and machine learning to assess nine types of natural disasters. During the Beijing hailstorm on May 13, 2025, the company said it detected the risk two hours in advance and sent alerts to 3.17 million individual customers and more than 20,000 corporate customers. Ping An’s Hail Alert.

Here, the value of a forecast depends on the action that is available. A “hail possible” alert is more useful if the customer still has time to move the car and the company can protect property or adjust operations. The forecast itself may not reduce losses if the warning is not delivered in time.

This changes the insurer’s task: beyond assessing and transferring risk, it also takes on work to reduce that risk. Analytically, the chain consists of forecasting the hazard, identifying affected objects, sending the alert, and carrying out the recommendation. The last step often depends on the customer.

How the warning is tied to claims handling

After that same hailstorm, Ping An allowed photo or video evidence for damage under 10,000 yuan without an on-site inspection or separate proof of the weather event. This is a special measure for a specific incident, not a universal rule for all claims. Post on hail response.

If the event is confirmed by general data, some documents may be redundant. However, such simplification should be a process rule, not an arbitrary model decision. Otherwise, identical cases will be handled differently.

Estimating prevented losses requires a counterfactual calculation: how much damage would have occurred without the warning. That cannot be observed directly in each individual case. So the claimed savings from risk prevention require a more careful interpretation than the mere fact that notifications were sent.

AI at Ping An Bank

What the bank discloses directly

In Section 3.3.4 of its first-half 2026 report, the bank describes AIGC content with quality and compliance checks, a combination of AI, remote, and offline service, intelligent marketing, analysis of the reasons for changes in indicators, and the use of agents in lending processes. It separately discloses 140 operational review agents in 55 scenarios, 35 consumer protection review models with coverage above 90%, more than 450 large-model use cases, and more than 5.3 billion tokens per day. Ping An Wealth Management names channel servicing, investment information collection, and risk control. Ping An Bank report, pages 43–45.

The business logic behind these use cases is explained below. This is an analytical interpretation of the disclosed areas: the report does not contain a full list of tools, all prompts, decision thresholds, or the architecture of each agent. Therefore, the specific conditional sequences should not be treated as instructions for how the bank’s production systems are built.

Generating Marketing Content and Controlling Publication

The task of an AIGC platform is different from that of a standard ad copy generator. In a banking process, the product, audience, current terms, and required disclosures all need to be aligned. A polished message with the wrong rate or an incomplete fee description creates a problem for both the customer and the company.

One possible operational setup is to prepare material from approved data, check it against a set of requirements, and send it for publication after the necessary approvals. AI helps create and analyze text; control should not rely solely on a second free-form generation. For numbers and terms, checks against a structured source are useful.

There is a separate quality metric here: how many materials had to be corrected, and why. If the team only measures how fast text is created, errors may be discovered after publication. It is much more useful to measure time to an approved draft and the share of correct outputs.

AI, Remote Employee, and Physical Channel

Combining channels makes it possible to distribute tasks by complexity. An automated assistant can explain the next step, a remote employee can clarify an unusual situation, and a physical channel can complete an operation that requires in-person presence. This is an explanation of an interaction model, not a promise that every Ping An customer goes through exactly this sequence.

Quality depends on context transfer. The customer should not have to repeat all details when moving between handlers. At the same time, only permitted data can be transferred: convenience does not override access restrictions.

For a banking product, the result for each channel should be tracked separately. A lower burden on the call center may be accompanied by longer lines in branches if automation fails to recognize complex requests well. Total servicing cost reveals this better than a local metric for one department.

Intelligent Marketing and KPI Analysis

Marketing support includes two different kinds of tasks: who should be offered the product, and why a business metric changed. The first task is about audience selection. The second is about explaining changes and preparing decisions for the team.

In analytics, it is dangerous to confuse correlation with causation. A model may notice that sales changed after a campaign, but that does not prove the campaign’s contribution: rates, seasonality, or the customer mix may also have changed at the same time. A good process separates observed relationships, verified causes, and hypotheses for further analysis.

Implementation requires not only natural-language answers, but also access to metric definitions. If departments calculate an active customer or an application differently, a language model will not resolve the conflict. The data and calculation rules must be aligned first.

Lending: From Review to Servicing

The bank explicitly states the use of AI in due diligence, approval, disbursement, and collections. Digital Financial Business Section. These stages have different goals, so a single generic “lending agent” is not enough to explain them.

In a hypothetical system, a review assistant gathers documents and flags discrepancies. A decision assistant prepares risk information. A disbursement assistant checks that conditions have been met. In the servicing stage, the system helps determine who needs to be contacted and how.

But the right to make a credit decision cannot be inferred from the fact that an LLM is used. It depends on process rules and the model’s allowed role. Similarly, automation of debt-related communication must preserve the accuracy of information and respect limits on contacting the customer.

AML and Operational Reviews

AML stands for anti-money laundering. In the report, the tasks assigned to operational agents include AML due diligence, account review, clearing, and cross-border transfers. The bank’s semiannual report.

In such processes, AI can help read documents, look for discrepancies, and prepare materials for review. However, a match or an unusual transaction is not yet proof of a violation. Rules for review, explainable grounds, and result monitoring are all needed.

The practical value of an agent often lies in preparing a complete package for the employee. If a specialist can see the needed documents, history, and reason for the alert more quickly, part of the routine work disappears. That may be more useful than trying to give the model the final decision without intermediate review.

Consumer Protection and Wording Review

In financial services, the system must check not only whether a client is suspicious, but also whether the company’s own actions are correct. An ad message, a product description, or an operator’s answer may contain ambiguity or missing conditions. Automated review helps catch these issues before they reach the client.

Training example: the text promises “free service,” but the original terms include a fee if a requirement is not met. The review must compare the promise and the source, then point out the exact mismatch. A vague conclusion like “the text looks fine” is not enough here.

The coverage metric for checks is not the same as the metric for errors found. You can automatically review most materials while still missing rare but significant violations. That is why separate quality checks are needed for difficult and borderline examples.

Investment Analysis and Trading Infrastructure

In the bank’s report, the volume of transactions through iDeal structured inquiry/quotation over the half-year is listed at about CNY 2.9 trillion. This is the platform’s volume, not proof that AI independently traded that amount. Ping An Bank report, digital finance section.

For investment analysis, it is useful to distinguish between information gathering, structuring, and decision-making. A model can prepare a summary of events or extract the relevant section from a filing. The analyst then verifies the source and assesses the meaning of the data in a specific context.

The problem arises when polished text hides a weak evidence base. That is why an investment assistant should preserve links and dates, separate fact from assumption, and show where the data is insufficient. Generating a confident forecast does not make it justified.

Medicine and AI Doctor

Medical LLM: What Has Been Disclosed About the Approach

Ping An Medical LLM 3.5 is described as an industry-specific model using data from PKU Healthcare and Ping An Health. The June article names clinical knowledge graphs, digital patient twins, adaptive evaluation, multi-step reinforcement learning, distillation, and hallucination control. The company also reported more than 90 early screening models in 1,500 institutions, 1.5 million examinations, and the use of multiple agents in a community of roughly 2 million patients with chronic diseases. Description of the medical model.

These details explain the development direction. They do not justify claiming that the system reproduces a doctor’s full work in any clinical situation. The public material also does not provide all the details of safety testing across different patient groups.

Model hallucination — a generated statement that is not based on a reliable foundation. In medicine, it is important to evaluate not only how often such responses appear, but also how severe they are. A rare incorrect statement about a dangerous condition may matter more than many errors in low-impact explanations.

Initial consultation and triage

Ping An describes AI Family Doctor within the “AI + doctor” model: AI handles common requests, helps with case sorting, reminders, and interpretation of test reports. Doctors are involved in complex conditions, chronic disease management, and online prescriptions. Description of medical services at WAIC.

Triage — the initial sorting of cases by the need and urgency of further care. For a system, the task is to direct the person correctly, not just list possible diseases. It requires rules for when the dialogue must stop and a medical professional must step in.

In an explanatory flow, the assistant gathers symptoms, clarifies context, and prepares information for the next step. It can make handoff to a doctor more useful. But that is not a reason to use an article’s text to choose treatment on your own: an individual decision requires clinical evaluation.

What the reported accuracy means

In the first-half 2026 results, Ping An reported more than 9.7 million AI Doctor users, support for over 11,300 diseases, and 96% accuracy in AI-aided diagnosis/treatment. Half-year disclosure.

This should be read as a company metric for the stated class of tasks. The brief release does not provide enough information to determine accuracy for any new patient: you need the sample, the correctness criteria, the conditions for physician involvement, and the error distribution. Also, coverage of diseases does not mean the same depth of validation for each one.

In your own medical AI deployment, quality should be evaluated by scenario. Explaining a term from an analysis, collecting complaints, and proposing a treatment plan are different levels of responsibility. A single overall percentage can hide weak performance in a small but important task group.

Oncology and multidisciplinary team support

In the June description, the model is used to support an MDT — a multidisciplinary team. The company reported that physicians included 85% of the generated material in final plans and that recommendations matched senior specialists in more than 92.5% of the serious diseases described. Ping An material on AI-enabled MDT.

These indicators measure different things. A doctor using the text is not the same as the correctness of the entire plan or an improvement in treatment outcomes. Agreement with specialists depends on the cases being compared and the criteria used, so it cannot be generalized to all clinical practice.

The role of AI here can be understood as helping collect and align information. Different disciplines use different data, and preparing a complete package takes time. If the system makes information more accessible, specialists can spend more time checking and discussing the decision. This is an analytical explanation of a possible benefit, not a proven result for every patient.

Screening and chronic disease management

Screening identifies people who may need additional examination. A risk model result is not the same as a diagnosis: it serves as a basis for the next step. For program quality, not only the number of screenings matters, but also the accuracy of referrals and the availability of follow-up care.

In chronic disease management, the task is different: help the person follow the agreed plan and notice changes. A typical system can remind, collect information, and pass it to the doctor. It must also distinguish between no response, a technical failure, and a worsening condition.

A claim about improving adherence — compliance with the treatment plan — requires defining exactly what counted as adherence. Responding to a reminder, taking a medication, and seeing a doctor have different ways of being verified. Public high-level metrics do not provide a full methodology for independent replication.

Insurance and medicine in one workflow

At WAIC, Ping An presented a product line that combines disease risk prediction, specialized insurance coverage, and an AI Disease Manager for ongoing support, including connections to doctors, hospitals, and rehabilitation. Description of the presented product line.

The strategic idea is to connect financial protection with access to care. But for the customer, the boundaries matter: medical assessment, eligibility for service, and eligibility for an insurance payout are not the same decision. Similar-looking interfaces should not obscure different rules.

This kind of integration also raises the data requirements. Information a doctor needs may not be needed by an insurance product salesperson. That is why personalization should be paired with clear access controls, not with the assumption that any data is useful to share with every ecosystem participant.

Home Care for Seniors

Sensors and Concierge

In the Ping An podcast, 26 types of smart devices are described, including fall-risk detection radar, sleep-monitoring mattresses, and watches. It outlines a three-party model: an AI concierge, a doctor, and an everyday services assistant. When an unusual event is detected, the system initiates a human check-in and, if needed, connects the person to help. Podcast on Home Care.

The value here comes from combining measurement and response. A sensor that detected an event has not yet provided help. The signal has to be delivered, contact established, and the next action determined. For someone living alone, these steps are especially important.

In a simplified workflow, an alert is checked: what was detected, whether the client responds, and who can help. Such a setup requires backup communication channels and a clearly assigned owner. AI cannot replace physical access to a person when help is needed on site.

Personal Profile and Household Functions

The same source also lists a profile built from medical records, wearable-device data, and assessments, voice communication, video calls, calendar, notifications, companionship chat, and a memory generator. Description of Home Service Functions.

It would be wrong to assume that all 26 device types are installed in every home. The mix depends on the service and the situation. Likewise, service eligibility, registered users, and the number of households actually served are different measures of scale.

Household functions are useful if they are easy to use without complicated navigation. At the same time, a voice interface has to account for misrecognition and the ability to undo an action. For sensitive tasks, such as sharing health information, convenience has to be paired with recipient verification.

Communication with an AI companion and memory recording belong to a different class of tasks than safety monitoring. They should not be evaluated with a single “accuracy” metric. For these tasks, usability, user control over content, and clarity that the other party is an automated system matter more.

Development, Documents, and Recruiting

AI writes code, but the share of generation does not equal the share of responsibility

The 2026 half-year report states 62% AI code penetration and defines it as the share of AI-generated code in new code. In the September CTO update, it says approximately 80% total code output and about 20% of routine development tasks are performed by AI end to end. These are different metric definitions. Half-year report, page 37, CTO update.

That share cannot be translated into “AI wrote 80% of all active systems.” New code may include tests, standard constructs, changes to small functions, and suggestions that an engineer then reviews. The number of lines produced does not show architectural quality or the reliability of the result.

Performance assessment needs metrics for accepted and working changes: time to release, post-release errors, review workload, and maintenance complexity. If code appears faster but requires more fixes, the gross output of generation creates a misleading impression.

Document Processing and Expense Audits

Expense auditing and servicing are named among the financial model scenarios. PingAnGPT Description. This is an example of “invisible AI”: the customer does not interact with the system directly, but its output affects internal processing.

A simplified document workflow includes classification, field extraction, rule matching, and generating a basis for review. Here it is useful to keep a link to the original excerpt. Otherwise, the employee gets a polished result that is hard to verify.

Special attention is needed for numbers and required fields. A language model can help interpret a document, but arithmetic and formal conditions are best checked separately. Readability is not a substitute for accurate verification.

Ping An Talent: Resumes, Interviews, and Compensation Explanations

In a March 2025 update, Ping An described semantic matching of resumes and job postings, AI interviews, and intelligent calls to discuss salary information. According to the company, the interviewer considers the logic of answers, speaking speed and tone, and facial expressions. Ping An Talent Update.

The mere use of these features does not prove that they reliably measure ability or personality traits. Validity and error data across different groups are needed for evaluation. Extra caution is necessary when the findings affect hiring decisions.

A clearer benefit of semantic search is finding resumes where experience is described differently than in the job posting. But even good search does not replace content review. AI can prepare a shortlist and questions, but the final process should provide explainable criteria and a way to correct inaccurate information.

OneConnect and Offers to Other Organizations

OneConnect is described as combining open models with a proprietary agent platform for financial institutions; wealth management, lending, digital banking, and office support are named. Ping An Material on GenAI.

This shows the potential to turn internal capabilities into technology products. However, Ping An’s internal results do not guarantee the same outcome for an external client. The data, processes, integrations, and constraints are different.

When purchasing a platform, it is important to separate the ready-made component from implementation work. The model may be available, but data, tool connections, permission checks, and exception-handling rules are still needed to achieve results. This part usually determines how well the solution fits a specific organization.

Agent Management and Personal Data

What the Agent Is Allowed to Do

The four levels of control that Ping An lists for agentic AI are discussed in the Express Service section. In practical terms, that means verifying the user, restricting tools, involving a human in risky cases, and keeping an activity history. Agent Management Description.

For business, this means authority should be defined in advance. Drafting an email and sending it are different permissions. Reading a contract and changing the terms are different operations. The model’s ability to formulate an action does not mean it is allowed to carry it out.

It is useful to tie controls to consequences. Viewing a public document, changing contact details, filing an insurance claim, and making a financial transaction require different levels of verification. One rule that says “the agent can do everything” or “everything must be confirmed by a human” rarely fits the whole workflow.

Action Audit and Human Review

A log is useful when it lets you reconstruct the basis for a decision. An entry should make it possible to understand what data was available, which tool was used, and what the outcome was. A log of a long conversation without operation state may not answer the key question: did the action occur in the core system?

Human review also needs to be substantive. If an employee is shown only the model output, they may miss an error in the source data. You need supporting materials, constraints, and the reason the case was escalated so the decision does not become a mechanical confirmation.

In a reviewable workflow, the employee can correct the data, reject the suggested action, and state the reason. These details then help analyze common errors. But feedback must be controlled before the model or rules are changed.

Privacy: data purpose matters more than data volume

Ping An publishes AI governance principles including human-centered design, user autonomy, safety, fairness, and transparency. Its corporate governance page also describes data protection and privacy-preserving computing. Ping An corporate governance.

Its 2025 sustainability report describes data minimization, explicit consent, user rights management, assessment of impacts on personal information, and limits on transfers to third parties. Sustainability report, pages 135–137.

This does not mean that every employee or agent has the right to combine any medical and financial information. Permission depends on the purpose, role, and specific service. Knowing the customer better does not automatically justify any use of the information.

For implementation, it is worth mapping the data flow before model selection: source, purpose, recipient, retention period, and permitted action. Then you can determine which information the agent actually needs. That often reduces both risk and integration complexity.

Why a general policy does not prove every service is safe

Governance principles set the framework, but having them is not a technical test. For a specific agent, you need to verify that the constraints work when the request is malformed, the data is incomplete, or someone tries to use another person's information. It is also important to evaluate actions after a failure.

Public materials do not provide a full independent audit of every Ping An service. So the correct conclusion is that the company discloses its governance approach and mechanisms, not that “all risks have been solved.” For readers, that boundary is more useful than a blanket safety promise.

The economics of AI and the limits of metrics

Four different sources of value

In a case analysis, it is useful to separate sales growth, lower operating costs, reduced losses, and improved customer relationships. One agent can affect several areas, but they should not be added together automatically. Otherwise, the same benefit gets counted multiple times.

Mechanism	What to measure in your own project	Typical interpretation error
Sales support	Additional verified sales relative to a comparable baseline	Counting all revenue where AI was involved as its contribution
Work automation	Cost of a completed task, including review	Counting only the model cost
Loss management	Verified change in payouts or risk	Calling a prevented payout revenue
Convenience and retention	Repeat contacts, usage, complaints, customer retention	Attributing all cross-product retention to AI

This table is an analytical framework for interpretation, not a financial model from Ping An. Each row requires its own data set and a way to verify the result.

Token cost and service cost

In its September update, Ping An said daily consumption increased from 30 billion tokens at the end of 2025 to more than 300 billion, computing power rose from 800 to 1,500 PF, token output per unit of power improved by 368%, and token cost fell by 76%. CTO disclosure.

A token is a unit of text processed by a model; it should not be treated as a customer, a request, or a finished service. Multiple model calls may be required for one task. So token growth describes infrastructure activity, not automatically better business results.

Task cost includes more than compute: there are integrations, storage, oversight, employee time, fixes, and support. Even a cheap answer can be expensive if the operation has to be redone afterward. For business, it is more useful to calculate the cost of the accepted result.

Illustrative formula: cost of completed tasks = models + tools + infrastructure + review + fixes + support. This is a list of calculation items, without assuming their size at Ping An.

Retention and multi-product usage

In a May Express Service article, the group links the ecosystem to deeper customer relationships and reports high retention across users of multiple product lines. Ping An ecosystem article.

Causality is not established here by a single comparison. Customers who have trusted the company for a long time may choose more products on their own. Using multiple services also creates a different level of connection with the organization. AI may make access easier, but its contribution needs special analysis to isolate.

For your own project, it is worth evaluating the specific barrier: the customer cannot find the service, does not finish the form, or has to switch channels. If that barrier is removed, you can check whether task completion changes. That makes the conclusion more precise than saying “the agent increases loyalty.”

Why you cannot apply Ping An's percentages to another business

Customer base size, product mix, and accumulated infrastructure all affect which processes can be automated. A smaller company may have a different document set and far fewer repetitive tasks. So the same model will produce different results.

Quality requirements also differ. A knowledge base search assistant and a financial transaction executor have different acceptance criteria. For transferring the lesson, it matters more to choose the right process than to reproduce the highest automation percentage.

What businesses can learn from Ping An

Start with a process that can be verified

The first step is to choose a repetitive task with a clear input and output. For example, preparing a response based on current terms, classifying a request, or extracting data from a specific document. The choice depends on available data, current workload, and the consequences of error.

It is worth documenting how the task is done today: who receives it, where they look for information, what they check, and where they record the result. Often the bottleneck turns out to be access to data or coordination. In that case, adding a model without fixing the process will not solve the problem.

The verifiable criterion should be framed in terms of outcome: the request is routed to the correct queue, the information matches the document, the response includes the current basis. The phrase “the agent works intelligently” does not let you accept or reject the decision. For more on choosing measurements, see the article AI agent performance metrics.

Choose the level of autonomy based on risk

For a new system, it makes sense to consider several roles: information search, draft preparation, recommendation, execution with confirmation, and limited autonomous execution. These are design options, not a mandatory implementation sequence. The right choice depends on the task.

If an error can be easily detected before action is taken, the assistant can be useful even without full automation. If the operation is irreversible or affects significant customer rights, controls should be stricter. A high degree of automation is not a goal in itself.

Task Type	Possible Starting AI Role	What to Check
Product Terms Search	Answer with Source	Timeliness and completeness
Document Preparation	Draft for an Employee	Data, numbers, and required fields
Case Routing	Suggested queue assignment or limited sending	Incorrect routing and loss of context
Record Update	Action after the required confirmation	Identity, permissions, and record outcome
High-Risk Decision	Preparation of materials for a specialist	Grounds, exceptions, and the ability to review

These are practical options for evaluation, not recommendations to hand the listed operations over to AI without analyzing the specific organization.

Prepare Knowledge and Tools

RAG — retrieving relevant materials from a knowledge base before generating an answer. This approach helps connect text with current documents, but it does not by itself guarantee correctness. You need to check the quality of the retrieved passage and how the model used it. A detailed breakdown is here: RAG accuracy and hallucinations.

For actions, another layer is needed — tools. They should return a clear operation status: completed, rejected, requires clarification, or result unknown. If the agent does not distinguish these states, it may report success after a failure.

It is useful to plan for retries and cancellation in advance wherever possible. For example, creating a request again may produce a duplicate, while requesting a status update again may be safe. These differences belong to the tool's business logic, not the model's free-form reasoning.

Check Exceptions and Human Handoff

The test set should include incomplete documents, contradictory information, unusual wording, and requests outside the authorized scope. These cases show whether the system knows how to stop. Success on simple examples does not answer that question.

The human handoff should include the information already collected and the exact reason. It is desirable to check whether the employee can quickly continue the task. If the automated stage leaves more work for them, the bot's local speed does not reflect the overall result.

It is also useful to evaluate user messages separately. They should explain the current status and the next step. The wording “everything is ready” is not acceptable if the task has only been sent for review.

Measure Results Across the Entire Workflow

For a pilot, you can choose several metrics: completion quality, time, cost, manual rework, and user satisfaction. This is a list of measurements, not expected values. Before comparing, you need to define the same task sample.

If the mix of requests changes, the result needs to be explained. For example, after automation, employees may be left with more complex cases, and their average handling time will increase. That does not necessarily mean worse performance: the workload changed. But without separate accounting, the conclusion will be wrong.

Pre-Launch Checklist

• The task input, outcome, and process owner are clear.
• There is an up-to-date knowledge source and a way to verify the answer.
• Permitted data and the rights of each tool are defined.
• Criteria for when a human should be involved are selected.
• Context is preserved between service stages.
• Numbers, documents, and completed operation status are checked.
• Complex and edge-case examples are prepared.
• Errors can be detected, corrected, and reviewed.
• Cost is calculated with verification and support included.
• The user understands when they are interacting with an automated system.

This checklist connects technology with a manageable service. It does not promise Ping An-level results: each task should be evaluated on its own data.

Frequently Asked Questions

Did Ping An really sell more than $8 billion through AI?

The company disclosed AI-supported sales in yuan. The conversion using the specific official exchange rate is given at the beginning of the article and exceeds $8 billion. This is not additional profit and not proof of fully autonomous deal closing.

Does the 81% service share mean people were replaced?

No. The metric counts interactions, not unique customers, employees, or fully completed customer journeys. The half-year report note clarifies voice and text robots, incoming and outgoing services for credit cards and insurance. Report, page 37.

How is Express Service different from a chatbot?

The company describes it as an agentic entry point into services that coordinates request execution across systems. A simple answer to a question and actually carrying out an operation are different outcomes. The difference is determined by tools and permissions, not just by the quality of the conversation.

Does Ping An use one proprietary model for all tasks?

Public materials describe industry-specific models and the use of open models in certain businesses. Based on these data, the entire architecture cannot be reduced to a single LLM. Speech, images, risk models, and rule execution solve different tasks.

Can AI Doctor be considered a replacement for a physician?

No. Ping An describes a collaboration model between AI and a doctor. The stated accuracy for a specific assessment is not a guarantee of diagnosis or treatment for a particular person. Decisions about medical care require appropriate clinical evaluation.

Why do the code figures show 62% and 80%?

The half-year report defines the first figure through a new code, while the September message uses wording about the total amount of code generated. Without a full comparable methodology, the exact increase or the share of autonomously written systems cannot be assessed.

Are prevented payouts AI profit?

This is a separate corporate loss-management metric. It is not equal to sales, revenue, or net profit. To assess the final effect, the methodology, costs, and quality of decisions need to be taken into account.

Can Ping An's approach be replicated in a small company?

You can use the process principle: choose a task, prepare the data, limit the tools, and measure the result. Copying the scale of the platform and the claimed percentages without your own verification makes no sense.

Primary Sources and Verification Method

This material is based on public disclosures available on the verification date. For the figures, reports and official publications were used; for the scenario descriptions, technology materials and corporate podcasts. Statements from the source text that lacked sufficient verified methodology were not turned into facts.

Key documents:

1. Ping An Full First-Half 2026 Report — definitions of service, new code, and AI-assisted sales.
2. Official first-half results release — initial KPIs.
3. 2025 annual results — prior period and platform scale.
4. September CTO presentation — Express Service, computing, and code.
5. Description of Express Service and a podcast on agentic AI — the interface and action management.
6. Ping An Bank’s First-Half Report — banking applications.
7. Description of Medical LLM 3.5 and WAIC 2026 solutions — healthcare and P&C.
8. Technology page, AskBob, EagleX, home care and Ping An Talent — separate business lines.

For each key metric, the period, unit, and meaning were checked. The corporate statement preserves the company attribution. Currency conversion is a separate calculation, and typical implementation patterns and takeaways are analytical commentary. The material does not claim that all of the technologies described have undergone independent audit.

Main limitation: the public data is sufficient to confirm the scale and directions, but not enough to reconstruct the full architecture, assess the errors of each product, or determine the causal contribution of AI to profit. Nor is it possible, based on the data presented, to list all of the group’s internal AI applications.

How AIrassvet helps you move from the Ping An example to your own challenge

For a company that wants to automate a similar process, it makes sense to start with one specific operation. On the AIrassvet website, which also uses the name “AI Studio of Vladimir Lomtev,” the following applicable directions are described:

• Process analysis and data preparation for the chosen automation task.
• Building an agent connected to a communication channel and a business system.
• Preparing a corporate knowledge base and RAG for responses based on current documents.
• Integrating and testing the solution with escalation of complex cases to an employee.

The first step is to define the current workflow, source data, constraints, and the criterion for an acceptable result. This makes it possible to discuss a specific implementation without applying someone else’s percentages to your own business.

Discuss the task.

What the Ping An experience shows

The original news is confirmed by official disclosures, with important clarifications about the meaning of sales and service volume. The broader takeaway is the distribution of AI across different stages: information retrieval, employee assistance, document processing, risk assessment, and action coordination.

What is especially useful in this case is the link between the model and the real process, along with the allowed tools. The scale of tokens and the number of applications describe technology usage, while the business outcome must be measured through completed service, cost, quality, and the consequences of errors.

For your own company, the next step is concrete: choose a recurring operation and describe how to know whether it was completed correctly. That is where adaptation begins — with a clear result, data, and control.