UnitedHealthcare has expanded access to Avery
UnitedHealthcare has expanded access to Avery, its generative AI assistant, to 21 million members across commercial, Medicare, and Medicaid plans. The scale deserves attention, but the number describes members eligible to use the service, not confirmed active users. UnitedHealthcare has not disclosed how many members have used Avery or the total number of conversations.
Introduced in March 2026, Avery initially reached approximately 6.5 million members in employer-sponsored plans and 160,000 Medicare Advantage members. The insurer had previously aimed to reach 20.5 million eligible members by year-end.
The expansion illustrates a challenge facing enterprise AI assistants: making a service available to millions of people is only one part of delivering dependable assistance. The harder work involves connecting accurate information, supporting complex interactions, and knowing when automation should hand responsibility back to a person.
Avery helps members understand benefits, claims, medical costs, deductibles, out-of-pocket expenses, provider networks, and plan balances through the UnitedHealthcare mobile app and myuhc.com. Its answers draw on benefits, provider, and claims systems, making the assistant more relevant than a tool offering general insurance explanations.
Are you looking for developers?
UnitedHealthcare developed Avery partly by studying patterns in customer service conversations and voice calls, then fine-tuning AI systems to answer questions using digital information. That work points to a broader engineering reality: useful enterprise AI depends on how well existing records can be retrieved and interpreted.
In a large organisation, information may be distributed across platforms with different structures and update cycles. A personalised assistant therefore needs reliable connections to business systems and careful controls over who can access individual records. Backend integration, APIs, structured data, and identity management are essential design considerations, even when the precise technical architecture of a particular assistant has not been disclosed.
UnitedHealthcare reports that Avery resolves 90% of member interactions and transfers the remaining 10% to human advocates. That is a meaningful operational claim, but the company has not published its calculation methodology. Resolution should not be confused with independently verified answer accuracy or customer satisfaction.
The insurer says it tracks accuracy, completeness, resolution, and response time, without disclosing separate numerical results for those measures. Enterprise teams should consider these dimensions together. A conversation may appear complete while still leaving a member confused, and a fast response may be less useful if it relies on outdated coverage information.Escalation is therefore part of product design, not evidence that automation has failed. When Avery transfers a member to a human advocate, it provides a conversation summary so the member does not have to repeat the entire exchange. That handoff connects the digital assistant to a wider service workflow.
Are you looking for developers?
Avery also supports brokers, employers, and service representatives with insurance information, plan comparisons, eligibility and enrollment documents, information retrieval, and interaction documentation. UnitedHealthcare is developing functionality that could let Avery call providers and schedule appointments, but this should not be treated as a broadly available feature.
Healthcare benefits information is sensitive, and errors can have consequential effects. UnitedHealthcare acknowledges that Avery may provide incomplete or inaccurate responses. It also states that Avery offers general information rather than medical advice, formal benefit determinations, coverage guarantees, or healthcare authorisations.
The company places Avery under UnitedHealth Group’s Responsible AI governance framework, which addresses model performance, data use, fairness, privacy, and bias. The group’s AI Review Board includes technologists, clinicians, ethicists, data scientists, security and privacy specialists, legal professionals, and member advocates. According to its 2025 proxy statement, the board can approve models or require temporary or permanent removal from production.
A Responsible AI Program Office also maintains a model inventory and oversees required reviews. These organisation-wide practices matter, but they should not be mistaken for published Avery-specific findings. The available material does not detail Avery’s conversation retention periods, use of conversations for further training, or independent accuracy assessments.
Are you looking for developers?
For any large enterprise assistant, technical governance calls for careful access controls, testing, monitoring, and records that allow teams to investigate failures. Human oversight needs to remain workable when a request falls outside the assistant’s capabilities.
This is where the UnitedHealthcare example extends beyond insurance. Enterprise AI assistants rely on connected workflows, accurate business data, maintainable integrations, and systems that can be tested as information and user needs change.
For organisations building comparable applications, Square Codex can help extend engineering capacity through nearshore software development and staff augmentation. Based in Costa Rica and serving North American companies, it supports custom and enterprise software, AI integration, backend and API development, data engineering, cloud applications, DevOps, workflow automation, and QA testing. Its role is software delivery alongside internal teams, not medical advice, coverage decisions, or clinical validation.
UnitedHealthcare’s expansion shows why enterprise AI assistants should be assessed through more than reach or a single resolution figure. Long-term value depends on accurate data, responsive software, meaningful performance measures, and dependable human assistance when the model cannot adequately answer a question.