Conversation Flow

Voice agents fail at conversation flow because dialog management in voice has to solve real-time turn-taking, persona consistency, structured data capture, and business-logic compliance simultaneously, and a failure in any one layer — a mistimed interruption, a leaked internal system name, a multi-field question, an ignored “busy” signal — breaks the illusion of a coherent conversation partner. Conversation flow is the largest failure surface in voice AI because it is not one problem but at least seven: audio-level turn mechanics, agent persona and identity integrity, call lifecycle transparency, data capture mechanics, flow and compliance logic, adaptive responsiveness, and content pacing — each with its own root cause and its own fix.

Key Takeaways

• 44 distinct failure patterns make conversation-flow the largest goal in the repository, spanning everything from sub-second turn-taking timing to multi-turn business-logic compliance.
• Data-collection design choices have measurable, large effects: first-try success drops from 92% for single-field questions to 48% for three-or-more fields in one turn, and script-compliance research finds required exact phrases matched only 40-60% of the time.
• Several patterns are self-inflicted by prompt design rather than model capability: long “never say X” banlists measurably increase the banned phrase’s occurrence rate 2-3x, and oversized system prompts add roughly 200ms of time-to-first-token per 1,000 tokens, both of which are fixable without a better model.
• Roughly a third of patterns concern real-time audio mechanics (barge-in, end-of-turn detection, turn-taking, silence interpretation) that have no equivalent in text-based chatbots — real-time audio-mechanics patterns are voice-specific failure modes, not general LLM dialog failures.

Scope

• Turn-Taking & Timing Mechanics — barge-in-failures, end-of-turn-detection, interruption-mishandling, turn-taking-errors, silence-misinterpretation, backchannel-timing-errors, response-latency-issues, slow-tool-silence, opening-timing-mismatch, prompt-bloat-latency. Grouped because all ten concern the sub-second, audio-level mechanics of who speaks when — a layer with no analogue in text chatbots, governed by VAD, endpointing, and pipeline latency rather than by conversation content.
• Persona, Disclosure & Identity Integrity — unnatural-conversational-style, emotional-expression-overuse, rapport-absence, ai-disclosure-failures, identity-manipulation, agent-self-attribution-errors. Grouped because each concerns whether the agent’s presented identity — its style, emotional calibration, honesty about being AI, resistance to persona-jailbreaks, and accuracy about what it can actually do — stays consistent and truthful across the call.
• Call Lifecycle & Transparency — internal-process-leakage, graceless-call-ending, premature-call-ending. Grouped because all three are about the mechanics of starting, closing, or narrating the call itself — exposing backend machinery, or getting the ending wrong in either direction (too abrupt or dragged out).
• Data Capture Mechanics — slot-extraction-errors, multi-field-collection-overload, incremental-capture-failures, unnecessary-data-collection, unverified-data-usage, vague-consent-exploitation, spoken-form-failures, text-prompt-voice-failure. Grouped because all eight concern how information moves between caller and agent — what’s asked, how it’s parsed, when it’s saved, and in what form it’s spoken — as opposed to the business rules governing that information.
• Flow & Business-Logic Compliance — script-compliance-drift, qualification-flow-violations, scope-boundary-violations, unauthorized-commitments, override-signal-failures, intent-boundary-confusion, outcome-classification-errors, wrong-number-wrong-person, negative-banlist-priming, tool-description-failures. Grouped because all ten concern whether the agent follows the required sequence, boundaries, and classifications a business process depends on — step order, scope limits, promise authority, and correct tool/intent selection.
• Adaptive Responsiveness & Context Continuity — hesitation-mishandling, response-adaptation-failure, multi-turn-context-loss, multilingual-code-switching, language-barrier-failures. Grouped because each is about whether the agent actually updates its behavior in response to what the caller just did — a hedge, a question, a reference to something said earlier, or a language switch — instead of continuing a fixed plan.
• Content Delivery & Verbosity — verbosity-despite-brevity-instructions, monologue-without-engagement. Grouped because both concern the sheer amount of content delivered per turn exceeding what a listener (as opposed to a reader) can retain, independent of the timing or business-logic layers above.

When Conversation Flow Matters

• The agent operates a multi-step qualification, booking, or data-collection flow where step order, required fields, and outcome classification feed a downstream CRM or business process
• The deployment is a phone-based, full-duplex voice agent where users can and do interrupt, pause to think, hesitate, or ask off-script questions mid-call
• A pipeline owner is deciding between fixing a symptom with better prompting versus fixing it architecturally — many patterns here (banlist priming, prompt bloat, multi-field collection) are prompt-design anti-patterns fixable without touching the underlying model

Cross-Pattern Insight

The 44 conversation-flow patterns share a recurring root cause: an agent optimized to complete a plan (a script, a turn-taking heuristic, a fixed silence threshold, a list of things to say) rather than to continuously reconcile that plan against what is actually happening in the conversation. Turn-taking failures happen because a silence timer doesn’t check whether the sentence was semantically complete. Flow-compliance failures happen because the agent advances to the next scripted step without validating prerequisites. Data-capture failures happen because the agent waits for a complete record instead of saving incrementally. Responsiveness failures happen because the agent generates its next line before checking whether the caller just asked a question it needs to answer first. The fix pattern that recurs across nearly every one of the 44 files is the same: insert an explicit check — a classifier, a state-machine gate, a validation step — between “caller said X” and “agent’s planned next action,” so the plan can be preempted, corrected, or delayed rather than executed blindly.

Frequently Asked Questions

What makes conversation flow have so many more patterns than other speech-and-audio goals?

Conversation flow sits at the intersection of real-time audio mechanics (turn-taking, latency, silence handling — problems that don’t exist in text chatbots) and business-logic dialog management (step sequencing, data capture, compliance, scope boundaries — problems that exist in any structured conversation, voice or text). Both problem classes are large on their own; conversation flow is where they combine, which is why it accounts for 44 of the repository’s speech-and-audio patterns versus 6-8 for the other three goals.

Which conversation-flow patterns are actually audio/timing problems versus dialog-logic problems?

Ten patterns — barge-in, end-of-turn detection, interruption mishandling, turn-taking errors, silence misinterpretation, backchannel timing, response latency, slow-tool silence, opening timing mismatch, and prompt-bloat latency — are governed by audio pipeline mechanics (VAD, endpointing, TTS streaming). The remaining patterns are dialog-logic problems that would also occur in a text-based multi-turn agent, just without the millisecond stakes.

Can better prompting alone fix conversation-flow failures?

For some patterns, yes and the fix is well-documented — negative-banlist-priming and prompt-bloat-latency are pure prompt-design anti-patterns with a straightforward rewrite. For most others — turn-taking, slot extraction, qualification-flow sequencing, scope boundaries — the documented mitigation is architectural: a state machine, a validation gate, or a classifier layer sitting between the caller’s utterance and the agent’s next response, not a prompt tweak alone.

What’s the single most common root cause across the 44 patterns?

The agent executing its next planned action (a script step, a silence-based turn transition, a scripted question) without first checking whether the caller’s most recent input changes what that next action should be — whether that input is an interruption, a question, a hesitation, an implicit correction, or an override signal like “I’m driving.”

How does conversation flow relate to the audio-handling and speech-recognition goals?

Conversation flow assumes speech has already been captured (audio-handling) and transcribed (speech-recognition) — its failures are about what the agent does with a correctly or approximately-correctly recognized utterance: when to respond, what to say, what data to extract, and which business rule to apply. A garbled transcript from a speech-recognition failure can still trigger a conversation-flow failure (e.g., wrong intent classification), but the two are different bugs with different fixes.

Patterns

Total: 44 patterns

Related Goals

• Speech Recognition — upstream transcription errors that can trigger conversation-flow failures like intent misclassification
• Voice Synthesis — the output-rendering layer conversation-flow content ultimately passes through
• Audio Handling — signal-level and session-lifecycle problems that occur below conversation-flow’s dialog-management layer