What problem does it solve?
Most contact with an insurer between purchase and claim is routine: am I covered for this, send me my certificate, I changed car, add my partner, update my card. Each request is simple but depends on the specific policy wording, endorsements and product version, so front line staff spend time looking things up and customers wait on hold at renewal peaks.
Many first generation chatbots answered generic FAQs and could not see the customer's policy or change anything, so the conversation ended in a queue anyway. Coverage answers are also regulated: a wrong "yes, you're covered" becomes a complaint or a dispute when a claim is declined.
How does it work?
- Identify and authenticate. The agent verifies the policyholder in proportion to the request: a logged in session for questions, step up checks before changes.
- Answer from the customer's own policy. Coverage questions are answered by retrieving the policy schedule, wording and endorsements that apply to this customer, with the clause cited.
- Complete routine changes. Through an allow list of policy system actions the agent updates details, adds drivers or items within set limits, issues documents and takes payments, and shows any premium change before the customer confirms.
- Know when to stop. Claims, complaints, cancellations with refunds above a threshold, vulnerability signals and ambiguous coverage questions go to a human with the conversation summary.
- Learn from the gaps. Unanswered questions and handovers are reviewed to fix content and add new intents.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Early adopters
- Channels
- Web chat, Mobile app, WhatsApp, Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Containment rate | Too few to pool | 30% to 50% | 2 | 1 organization, 1 vendor |
| First contact resolution | Too few to pool | 60% | 1 | 1 organization |
Value drivers: Lower cost to serve, Customer experience, Inclusion and access, Employee productivity.
Indicative value
A personal lines insurer with 1 million policyholders
USD 500,000 to USD 4 million
Human handled servicing cost avoided per year
How this is calculated
Formula: policyholders * contactsPerPolicyholder * containment * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Policyholders policyholders, policyholders | 1,000,000 | 1,000,000 | The reference insurer. |
| Servicing contacts per policyholder per year contactsPerPolicyholder, contacts per policyholder per year | 0.5 | 1 | Editorial assumption, excluding claims contacts. Replace with your own contact volume. |
| Share of servicing contacts the agent resolves containment, fraction of contacts | 0.25 | 0.5 | In line with the evidence on this page (Lemonade's 10-K says over half of inquiries are handled without human intervention). |
| Cost of a human handled contact costPerContact, USD per contact | 4 | 8 | Editorial assumption for a blended phone, chat and email contact. Replace with your own fully loaded cost. |
What it leaves out: Gross avoided contact cost only. It leaves out the cost of running the agent, integration with the policy system, and effects on retention and complaints, which can go either way depending on answer quality.
Who already uses it?
6 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Lemonade
United States · Insurance · 2025
Lemonade sells renters, homeowners, pet, car and life insurance through a chat with its bot AI Maya, which collects information, personalizes coverage, creates the quote and takes payment by asking a limited number of high impact questions. Its 2025 annual report says AI Maya and its APIs sell 98% of its policies. A second bot platform, CX.AI, resolves pre and post purchase requests such as coverage questions, adding a spouse, changing coverage amounts or payment methods and adding newly bought items, and handles over half of customer inquiries without human intervention.
- Containment rate: at least 50%, customer inquiries handled by CX.AI without human intervention, as reported in the 10-K for 2025
"Currently, over half of Lemonade’s customer inquiries are handled this way."
Claimed by: organization
Sun Life
Canada · Insurance · 2025
Sun Life's 2025 annual report says AI tools cut median response times for individual insurance applications in a target segment by close to half, including a 50% increase in straight through underwriting. In its Client Contact Centre, generative AI raised chatbot containment by 16 percentage points year over year, and advisors use a generative AI Notes Assistant Tool; in Hong Kong it launched Advisory Buddy, a GenAI chatbot inside its Advisor Workbench. In Malaysia almost two thirds of clients received automated underwriting decisions within two hours.
- Cycle time reduction: about 50%, median response time for individual insurance applications, target segment, 2025
"Leveraged AI tools to do more for our Clients including reduced median response times for individual insurance applications for a target segment by close to half (including increasing straight-through underwriting by 50%), increased chatbot containment rate (up 16 percentage points year-over-year) with generative AI tools in the Client Contact Centre, and enhanced advisor productivity using a generative AI-powered Notes Assistant Tool."
Claimed by: organization
Waterdrop
China · Insurance · 2025
Waterdrop, a Chinese online insurance distribution and health services platform, runs a suite of AI applications called Waterdrop Guardian that either talk to users directly or support its online insurance consultants. In its second quarter 2025 results it reported that its AI Customer Service Agent resolved 60% of inquiries on first contact, that its Life Planner Copilot handled 300,000 product consultations for consultants, and that premiums facilitated by its AI Medical Insurance Expert rose 155% from the previous quarter (the release does not say whether this tool talks to users directly or supports consultants). It also launched KEYI.AI, a real time AI underwriting assistant for consultants. Its 2023 annual report describes an earlier LLM powered AI Insurance Consultant that was tested internally in medical insurance scenarios, an internal test rather than a sales deployment.
- First contact resolution: 60%, second quarter 2025, AI Customer Service Agent
"‘AI Customer Service Agent’ resolved 60% of inquiries on first contact, enhancing user experience."
Claimed by: organization
Nsure.com
United States · Insurance · 2024
Nsure.com is a Florida based digital insurance agency that lets consumers compare home and auto quotes from more than 50 insurers and buy online. Generative AI in Power Automate cut its service representatives' manual processing time by more than 60%, for example by triaging the shared inboxes and either preparing an automated response or routing each email to an agent. It replaced a third party chatbot with a Copilot Studio copilot, Friendly John, that helps customers submit payments, review renewal offers and request discounts, with an interactive voice response option and after hours support. Its VP of AI and Automation says it handles around 60% of customer questions, and the company plans to use copilots for new policy sales and cross selling.
No outcome disclosed.
Zurich Insurance (Hong Kong)
Hong Kong SAR China · Insurance · 2024
Zurich Insurance (Hong Kong) added WhatsApp to its contact centre with Dynamics 365 Contact Center and an agent built in Copilot Studio that captures preliminary information such as policy numbers before escalating to live staff, who then already know what the customer needs. Its head of customer services management says the bot might be able to answer simple questions. A second agent automates motor claim status updates from external surveyors to customers by SMS or email. The company reports lower call and email volumes and staff handling up to two chats at once, and was piloting Copilot to search the knowledge base during live chats. No containment figure is published for the AI agent itself.
No outcome disclosed.
LAQO
Croatia · Insurance · 2023
LAQO, Croatia's first fully digital insurance provider (part of Croatian Insurance), built Pavle with Infobip on Azure OpenAI Service to answer customers 24/7 on WhatsApp in Croatian. The assistant is limited to insurance claims and general information about LAQO to reduce the risk of misleading answers, guides customers through reporting an accident after confirming cover, and transfers complex queries to a live agent. The vendor reports that Pavle handles 30% of customer queries and that 90% of queries are resolved within three to five messages; LAQO's head of digital sales and customer support says the contact centre now spends 10 percent less effort.
- Containment rate: 30%, share of customer queries handled by the assistant
"Today, LAQOs digital assistant is handling 30 percent of customer queries, freeing LAQO’s agents to focus on complex cases and customer acquisition."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Policy wordings, schedules and endorsements by product version, linked to each policy
- A catalog of servicing intents with volumes from the contact centre
- Rules for which changes can be made without underwriting review, and their limits
- Approved answers for regulated topics (cancellation rights, complaints, claims)
Systems to integrate
- Policy administration system (read and mid term adjustment APIs)
- Document generation for certificates and proof of cover
- Payment provider for premium changes
- Contact centre platform for handover with context
- Identity verification and customer portal login
Complexity: Medium
Answering from generic content is easy; answering from the customer's own wording and changing the policy is the work. Legacy policy administration systems may lack APIs for mid term adjustments, and product wordings exist in many versions.
- 1
Rank intents by volume and risk
Start with document requests, payment updates and simple coverage questions; leave cancellations with refunds and anything touching claims for a later wave.
- 2
Link answers to the right wording
Index wordings by product and version and retrieve by the customer's policy, not by keyword. If the customer is not identified, answer only in general terms and say so.
- 3
Define the change allow list
For every change write the API call, the authentication level, the underwriting limits (for example which vehicles or sums insured may be changed without review) and the confirmation the customer sees.
- 4
Design handover and vulnerability rules
Hand over on complaints, claims, bereavement, financial difficulty and repeated failure, with the summary and verified identity passed to the human.
- 5
Test with real conversations
Build a test set from transcripts, including tricky coverage questions and attempts to push the agent into confirming cover it cannot confirm, and run it on every change.
Guardrails
- Coverage answers only from the customer's own policy documents, with the clause cited
- Explicit wording that the agent does not decide claims, with handover for any claim question
- Premium changes shown and confirmed by the customer before they apply
- Step up authentication before changes to payment details or named persons
- AI disclosure at the start of the conversation and an easy route to a human
KPIs to instrument
- Containment per intent, counting repeat contacts within seven days as not contained
- Accuracy of coverage answers on a weekly audited sample
- Handover rate and reasons
- Customer satisfaction for AI handled versus human handled contacts
- Complaints that mention the assistant
Human in the loop
Humans handle claims, complaints, vulnerable customers and any change outside the allow list. A service quality team reviews a weekly sample of contained conversations, with extra focus on coverage answers, and approves every new intent before release.
Common failure modes
- Confident wrong coverage answers
- The agent answers from the current product wording when the customer holds an older version. Retrieve by policy and version, and refuse when unsure.
- Changes that should have gone to underwriting
- A mid term change alters the risk (a new driver, a higher sum insured) without review. Encode underwriting limits in the allow list.
- Containment that is really abandonment
- Customers give up and call instead. Measure repeat contacts and satisfaction per intent.
- Claims conversations handled as service
- A customer describes a loss while asking about cover. Detect claim signals and hand over to claims.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
A customer facing assistant must be designed so that people know they are interacting with AI (Article 50(1), applicable from 2 August 2026). It is not high risk as long as it does not carry out risk assessment and pricing in relation to natural persons in life and health insurance (Annex III point 5(c)).
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context.
- Opinion on Artificial Intelligence governance and risk management (European Insurance and Occupational Pensions Authority, Europe). Published 6 August 2025 and addressed to national supervisors. Sets out risk based, proportionate expectations for insurers using AI systems, including fairness, transparency and explainability, and human oversight, and mentions chatbots as an example use.
Controls to put in place
- AI disclosure and a visible route to a human in every channel
- Inventory entry with an accountable owner and a documented action allow list
- Audit trail of every policy change the agent made
- Regression tests on coverage questions for each wording release
- Monitoring of complaints and outcomes for vulnerable customers
Frequently asked questions
- What share of customer questions can an AI agent handle?
- Published figures sit between about 30% and 60%, though each covers a different scope. Infobip reports that LAQO's assistant, which covers claims and general information about LAQO rather than policy servicing, handles 30% of customer queries. Nsure.com says its copilot handles around 60% of customer questions. Lemonade's 10-K says over half of its customer inquiries are handled by its bot platform without human intervention. Waterdrop reported in its second quarter 2025 results that its AI Customer Service Agent resolved 60% of inquiries on first contact.
- Can the agent tell a customer whether they are covered?
- It can explain what the customer's own policy wording says, with the clause cited, and should hand over when the answer depends on the facts of a loss. Claims decisions stay with the claims team.
- Is a policy servicing chatbot high risk under the EU AI Act?
- Usually not; it carries the Article 50 duty to make clear that customers are talking to AI. It would become high risk under Annex III point 5(c) if it were used for risk assessment and pricing in relation to individuals in life and health insurance.
How to cite this page
Blits.ai AI Use Case Library, "AI agent for insurance policy servicing", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/insurance-policy-servicing-agent. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published