A caller wants a claim status, but must recite several long numbers before the assistant even searches. Each extra turn creates another chance to mishear an answer, force a retry, or lose the caller. Complexity is the burden created by the steps, decisions, and information a conversation requires—not simply the length of its responses. section 8.1.1 section 8.0
Some complexity earns its place. An FAQ bot can explain how to check an account balance; a self-service bot can provide the balance, but may need authentication and access to a backend system. The design question is not whether to remove every step. It is which steps are necessary to complete the user’s task, and whether the user or the organization bears their burden. section 8.1
Ask for a claim number only when needed
A claim-search flow initially requested five pieces of information before searching and succeeded only 40% of the time. Research showed that a member-and-date search usually returned one claim. The revised flow omitted the claim number for those callers and requested it only when multiple claims needed disambiguation. Task completion improved, and claim-number search failures were cut in half. section 8.1.1
Listing 8.2 Most callers only need four pieces of information in improved experience
CALLER: I want to check the status of a claim. VIRTUAL ASSISTANT: I can help you check the status of a claim. First, please tell me your provider ID. CALLER: <gives provider number> VIRTUAL ASSISTANT: Thank you. Next, please give me the Member ID. CALLER: <gives member number> VIRTUAL ASSISTANT: Got it. And what is the claim date? CALLER: <gives claim date> VIRTUAL ASSISTANT: I found your claim.
The improved dialogue shows the shorter path most callers take: after giving the requested identifiers and date, they hear that their claim was found without supplying a claim number. section 8.1.1
A step can also be burdensome after the lookup succeeds. When callers were asked whether to hear claim details or receive a fax, almost two-thirds requested a representative instead; only 1% chose fax. Giving the details immediately and offering fax afterward removed a decision between the caller and their goal. The relevant measures are not just turn count, but completion, retries, abandonment, and escalation. section 8.1.2 section 8.1.1
To simplify a flow without losing its purpose:
- Collect information when it is needed; reserve disambiguation for cases with multiple results. section 8.1.1
- Deliver the requested result before offering an optional follow-up. section 8.1.2
- Use generative AI to suggest or test flow improvements, then judge the proposed changes against observed user outcomes. section 10.0 section 8.1.2
Complexity can also describe a retrieval query rather than a dialogue. In a retrieval-augmented chatbot, a highly specific, multi-attribute query can strain vector search and slow retrieval. That is a distinct performance concern from making users navigate too many conversational steps. section 6.5.1 section 8.1.1
