which model are you using

Here is an example transcript from one of our calls (callid: call_01985dd5393ffe2a7e468bb577d). Note that the user spells (@ 0:21, 0:40) their name "A", "E", "R", "I", "Q" (spoken as the name "Eric", but spelled differently), yet the AI agent initially spells it (@ 0:32) "E", "R", "I", "Q". Even though we have explicitly instructed the AI via the prompt to use every spelled letter the caller gives and to ignore common spellings/etc... The agent still will inaccurately record the name. Note that when tool invocation @ 0:44 is called, the spelled name is "Eriq", despite the caller clearly spelling the name (@ 0:21, 0:40). Even the speech to text is using the "Eric" spelling as well. Agent 0:14 I can help get things started so our scheduling specialist can take over. First, may I have your first and last name, please? User 0:21 Yeah. My name is Eric Porter. That's a e r I q Last name Porter, p o r t e r. Knowledge Base Retrieval Agent 0:32 Thanks, Eric. Let me just confirm the spelling. For your first name, I have: E, R, I, Q. Is that correct? User 0:40 No. It's a e r I q. Tool Invocation: save_confirmed_name 0:44 tool_call_id: c897accf5feddab4 { "spelled_name": "Eriq", "name_type": "caller_first_name", "execution_message": "Thanks for clarifying that. And for your last name, I have P, O, R, T, E, R. Is that correct?" } Tool Result 0:44 Knowledge Base Retrieval Agent 0:44 Thanks for clarifying that. And for your last name, I have P, O, R, T, E, R. Is that correct? User 0:50 Yes. Tool Invocation: save_confirmed_name 0:54 Tool Result 0:54 Agent 0:55 Thank you, Eric. What's the best phone number for you, including area code?

Our main concern with this is that we need name collection to be 100% as when we work with our callers and record their names for our clients, incorrectly spelled names will cause issues for all involved. So we are looking to understand what we can reasonably expect from the AI agent. We have exhausted all our idea's with how we are tackling it in the prompt and our save_confirmed_name tool would work perfect in theory, but we still rely on the LLM correctly assembling the name from the spelled letters. We see from the example transcript above, the LLM did not do that correctly. There are other examples of this behavior as well. We have also isolated the name confirmation prompt to another single prompt agent to focus on just name collection, and regardless of whether we turn on accuracy mode, switch models, or reinforce the prompt to push the agent to collect names from how the caller spells them - we still find that this behavior persist. Is this solvable?