AI benefits assistants

Our agency partners have spent years building well-organized benefits sites for their clients, and we help publish much of that content. But users are users: they rarely search and hardly read. The answer an employee needed was usually on the site, a few clicks away, and they didn't find it.

Benefits questions are also specific and high-stakes. A vague or wrong answer can cost someone money or coverage, so an assistant that guesses is worse than none.

Each assistant indexes the employer's benefits site and takes in extra material written just for it. We set its tone and teach it the organization's own names for plans and programs, so a question asked in internal shorthand still finds the right answer.

For the topics that matter most, the instructions spell out what a complete answer has to cover. Ask about one of those topics and the answer includes those points every time, without reading from a fixed script. Every answer links to its sources so employees can check it for themselves, and when a question is outside what the assistant should answer, it says so and points to the right contact.

Some employees would rather talk than type. For one employer we built a chat interface that works the way people expect from ChatGPT or Claude, then added voice conversations using ElevenLabs. Voice agents usually get their own copy of the content and their own prompt, so over time the spoken and written answers drift apart. We connected the voice agent to the same knowledge base and instructions as the chat, so an employee gets the same answer whether they type the question or ask it out loud.

Since moving to our chat provider's newer engine and adding the per-topic instructions, conversations started and total questions asked have grown two to three times, with usage up across the board.

Employees ask in their own words and get an answer in the employer's own terms, with a link to the page it came from. The content our partners built is finally being found.