Artificial Intelligence at Allstate – Two Use Cases

Allstate is one of the largest personal lines insurers in the United States, responsible for serving millions of policyholders through its claims, service, and distribution operations. That scale creates pressure to handle high volumes of routine interactions efficiently while still supporting customers in stressful moments, such as after a loss event. 

Research from the Organisation for Economic Co-operation and Development (OECD) notes that insurers are increasingly using AI to streamline administrative work, personalize communication, and support decision-making, while regulators are paying closer attention to the risks posed by opaque or biased models. 

In the United States, the National Association of Insurance Commissioners has issued principles that encourage strong governance, human oversight, and transparency whenever insurers adopt AI in customer-facing workflows. The stakes are high: in J.D. Power’s 2023 U.S. Claims Digital Experience Study, only 41% of customers “completely agree” that insurers’ digital tools met their expectations, and just 35% said the estimation process was “very easy,” underscoring how fragile trust can be when AI-enabled experiences fall short.

Within this environment, Allstate has described several internal AI deployments. Two in particular stand out as mature, clearly scoped applications that address specific business problems rather than abstract innovation goals. Both examples show Allstate focusing on concrete workflow changes, measurable efficiency gains, and oversight practices that align with evolving expectations for responsible AI in insurance.

This article breaks down both use cases and extends on the reporting Emerj conducted in 2022 into Allstate’s AI strategy:

• Conversational AI for customer and agent support: A virtual assistant used in digital channels to resolve common service inquiries and route more complex cases to human agents. 
• Generative AI for claims communications: Generating drafts of claim-related messages directly in the adjuster’s workflow, shifting work from manual writing to review and approval.

Conversational AI for Customer and Agent Support

Contact volumes have been rising industry-wide: in a 2022 McKinsey survey, 61% of customer-care leaders reported growth in total calls and anticipated further increases. As the survey notes, an expanding customer base is a good indicator of business growth for insurance enterprises, but it also adds stress to already overburdened contact centers. 

As call volumes rise — especially with more complex issues — customers often have to reach out multiple times, which strains capacity even further and leads to a poorer overall customer experience. As customer service volumes and demands have increased for Allstate, the company has explored ways to automate common, well-understood questions so human agents can focus on nuanced situations and more value-driven conversations.

According to a joint write-up with technology partner Tietoevry, Allstate initially experimented with chatbots that tried to support too many intents at once. That broad scope made it challenging to train models effectively and to show clear business value. The team eventually “paused and pivoted,” narrowing the focus to a smaller set of high-volume questions where automation was realistic, and outcomes were easy to measure. 

The conversational AI system now relies on natural language understanding models trained on historical chat logs and labeled examples of customer intent. It pulls data from internal sources such as policy information, billing records, and claim status, and combines it with a structured knowledge base containing approved answers and procedural guidance. Business rules determine when the bot can complete a task on its own and when it must escalate the conversation.

A central performance metric for the system is containment: the percentage of digital conversations that the bot can handle from start to finish without a handoff to a human agent. Tietoevry reports that Allstate’s consumer-facing bot achieves a containment rate of roughly 38-40%. In other words, nearly four out of ten chats are fully resolved by the virtual assistant.

For customers, the experience changes from waiting to speak with an agent to receiving instant responses for straightforward questions. A typical interaction might involve checking a claim status, confirming a payment date, or updating contact information. When the intent is ambiguous, or the customer’s situation is emotionally fraught, the bot passes the conversation to a human agent along with the transcript and any relevant context it has already gathered.

• The customer states the reason for the inquiry, initiating the system to identify whether the intent matches a well-defined, repetitive use case.
• The virtual assistant provides immediate responses to routine questions by drawing on predefined workflows and available customer information.
• If the customer’s intent is ambiguous, the bot asks a clarifying question to confirm what the user needs.
• When emotional tone or complexity increases, the system escalates the interaction to a human agent.
• The agent receives the full transcript and context, allowing them to begin the conversation with relevant background rather than asking the customer to repeat information.

For agents, the virtual assistant reduces the volume of repetitive inquiries, allowing them to focus on complex cases that require judgment, negotiation, or empathy. This pattern aligns with NAIC guidance, which emphasizes that AI should assist human decision makers rather than replace them entirely in sensitive areas. 

While Allstate has not disclosed specific cost or headcount impacts, it supports a significant share of chat volume through automation. Industry research shows that at enterprise scale, AI-enabled claims transformations can meaningfully reduce manual handling and improve consistency. In one UK example, an insurer that deployed more than 80 AI models across its claims domain cut the time needed to assess complex cases by 23 days, improved routing accuracy by 30%, and saw customer complaints fall by 65%.

The following 23-second clip features Allstate’s conversational AI lead explaining that the company’s consumer-facing virtual assistant achieves 38-40% containment, meaning roughly 400,000 out of one million conversations are now handled end-to-end by the bot:

Video: Allstate’s Containment Results Explained (Source: Tietoevry)

Generative AI for Claims Communications

Across the insurance sector, inconsistent and overly technical claims communications have become a major driver of customer frustration, repeat calls, and churn.

A recent interview with Mark Garett, Director of Insurance Intelligence with J.D. Power by Reinsurance News, noted that homeowners who say their insurer is “very easy to communicate with” report an average property-claims satisfaction score of 777. Compared with just 337 among customers who find communication difficult, that number represents a gap of more than double in perceived experience. 

As catastrophic losses, higher deductibles, and premium increases put more pressure on policyholders, carriers have strong incentives to standardize and simplify adjuster messaging at scale, making clarity, empathy, and compliance in every outbound message a core business problem rather than a mere wording choice.

A Wall Street Journal interview with Allstate Chief Information Officer Zulfi Jeevanjee reports that Allstate has introduced a generative AI system that drafts many of these messages automatically. The system uses large language models trained on internal data, including historical communications, policy language, compliance templates, and preferred tone guidelines. It operates inside the claims platform, where it can access claim metadata, loss details, and policyholder information as context for each draft.

While not explicitly mentioned in the article, documentation from the hybrid consultancy Aimfluence covering the same use case (and published 15 days after the Wall Street Journal article from February of this year) claims that Allstate leveraged OpenAI models to replace “jargon-laden messages with compassionate, personalized interactions.” 

While neither the Wall Street Journal article nor Aimfluence’s use case documentation confirms the system’s capabilities, neither provides detailed operational steps either. To describe how the tool functions in practice, the following explanation draws on common workflow patterns in AI-assisted drafting across the insurance sector, as described in this Boston Consulting Group article, which notes that similar systems use claim metadata, policyholder details, and loss information as contextual inputs.

In these deployments, the system typically generates a suggested message based on the claim status and the required communication type. Industry research shows that first-draft messages are designed to present information clearly, reduce dense jargon, and maintain a more empathetic tone. The adjuster then reviews the draft, ensures accuracy and regulatory alignment, and makes any needed edits before sending it to the customer.

Before AI assistance, communications often followed a template-and-edit process in which adjusters selected a base message and modified it to fit the claim, while navigating regulatory and legal language. These manual steps consumed time and occasionally led to overly technical or unclear messages in a large organization where tone and complexity varied.

With AI-assisted drafting, the workflow becomes faster and more consistent across the organization:

• Adjusters provide the claim context with Information such as loss type, required documentation, and customer concerns being fed into the drafting tool.
• The AI generates a first-draft message using standardized templates, coverage rules, and regulatory-safe language.
• Adjusters review and refine the draft to ensure accuracy, tone, and situational sensitivity.
• Final communications are sent to the customer, preserving human oversight while reducing the amount of manual writing required.

Graphic illustrating a generative-AI workflow for drafting insurance claims communications. (Source: Boston Consulting Group)

Allstate reports in the same Wall Street Journal feature that the AI-generated messages tend to be less accusatory, less dependent on internal terminology, and more aligned with how customers actually speak. They also note that the system is used at scale, helping produce tens of thousands of claim-related messages per day rather than operating as a small pilot.

The workflow change for adjusters is significant. Instead of starting from a blank template and crafting each sentence, they begin with a context-aware draft and focus on verification and personalization. For Allstate, this likely shortens the time required to send updates, increases consistency across teams, and supports a customer experience that feels more transparent and less bureaucratic.

While the Wall Street Journal does not mention measurable business results, the Aimfluence case documentation reports the following:

Quantified Impact:

• 70% reduction in email drafting time
• 30% fewer complaints about jargon
• Improved Net Promoter Scores (NPS)

Operational Metrics:

• 250,000+ monthly conversations handled by AI
• 75% resolution on first contact (likely includes Amelia, the cognitive agent)

Expanded Capabilities:

• Emotional intelligence (detecting stress)
• Personalization based on customer history