A chatbot with high CSAT and a low resolution rate is still a cost centre. Resolution requires the system to actually complete the user's intent: retrieve the right policy, update the right record, escalate to the right human with full context. We design for resolution from the start, not CSAT — and measure both.
The chatbot optimization problem has always been resolution rate, not response quality — but most chatbot implementations optimize for response quality because it is easier to measure. A chatbot that gives confident-sounding wrong answers scores well on CSAT if it sounds authoritative and the user does not immediately realize the answer was wrong. Resolution rate — did the user actually accomplish what they came to do — is harder to measure and usually much lower than CSAT suggests.
The conversational AI landscape has also moved. Voice agents using ElevenLabs or PlayHT streaming TTS have crossed a latency threshold where conversations feel natural rather than halting. Multimodal inputs — users sharing screenshots of error messages, photos of products, or documents — are now first-class in GPT-4o and Claude 3.5 Sonnet. Personality design has emerged as a genuine discipline: the "uncanny valley" of AI conversations — where responses are fluent but feel robotic in ways users cannot articulate — is a solvable design problem, not an inherent limitation of LLMs.
We design conversational systems starting from the intent taxonomy — the structured map of user goals the system needs to handle. Each intent has a defined resolution path: what information is needed, what action or answer resolves it, and what escalation triggers apply. This prevents the common failure mode of an LLM that generates fluent responses to intents it cannot actually resolve.
For voice agents, we integrate ElevenLabs or PlayHT streaming TTS with Whisper or platform ASR to produce conversational latency. Voice personality design — the tone, pacing, and conversational characteristics — is treated as a first-class design concern, not an afterthought. For multimodal conversations, we build processing pipelines that handle image and document inputs appropriately and pass the extracted context to the conversation model.
Analyze existing support tickets, chat logs, or product questions to build a data-driven intent taxonomy. Cover the top intents by volume and critical intents by business impact. Design resolution paths for each intent category.
Ingest support documentation, product content, and policies into a retrieval system. Configure chunking, embedding model selection, and retrieval quality. Establish process for keeping the knowledge base current when content changes.
ElevenLabs or PlayHT streaming TTS with latency profiling. Whisper or platform ASR for voice input. Conversation state management for multi-turn voice interactions. Personality design for the voice agent persona.
GPT-4o or Claude vision for image understanding. Document extraction pipeline for shared files. Context injection from multimodal inputs into the conversation state.
Integrate with Zendesk, Intercom, or Freshdesk. Define escalation triggers. Pass conversation context, resolution attempts, and multimodal inputs to human agents at handoff. No escalation should start from scratch.
We configure ElevenLabs and PlayHT streaming TTS pipelines targeting first-chunk latency under 500ms — the threshold where voice conversations stop feeling synthetic. Personality design is a product requirement: tone, pacing, and speaking style are spec'd against your brand and user context before a word gets recorded.
GPT-4o and Claude 3.5 Sonnet accept image and document inputs natively. We build processing pipelines that ingest user-submitted screenshots, product photos, and PDFs — extract structured context, and inject it into conversation state so the agent responds to what the user is actually showing, not just what they typed.
Responses are grounded in your product documentation, support policies, and knowledge base — not the LLM's general training. The retrieval layer uses embedding search over chunked, indexed source content. Knowledge updates propagate automatically when source documents change: no retraining, no redeployment.
We define resolution states per intent — confirmation signals, follow-up absence windows, explicit user signals — and track whether conversations reach them. Unresolved conversations trigger escalation before users ask for a human. Resolution rate is logged, dashboarded, and improvable through intent coverage expansion.
Robotic-feeling agents are a design failure, not an LLM ceiling. We spec conversational characteristics explicitly: response framing, acknowledgment patterns, uncertainty expressions, and recovery behavior when the agent doesn't have an answer. Persona consistency is enforced through system prompt architecture, not post-hoc prompt patching.
A well-scoped system typically handles 40–60% of tier-1 support queries without human escalation. The 40–60% it escalates are the ones that needed a human anyway — with context the agent has already assembled, not cold handoffs where the human starts from scratch.
Anonymized engagements with real metrics — no client names per NDA.
68%
Query Deflection
4.1s
Avg Response Time
3.2x
Advisor Throughput
“The query volume our advisors were handling manually dropped within the first month. The system handles the routine questions correctly, escalates when it should, and our compliance team signed off on every response template before it went live.”
— Head of Advisory Operations, Wealth Management Firm
2.1x
Time on Page
38%
More Qualified Leads
52%
Fewer No-Shows
“The no-show reduction was the metric our agents cared about most. The buyers who book visits after exploring the virtual tour have already self-selected — they know the property and they are serious.”
— Head of Digital, Real Estate Platform
Yes, for the right use cases. ElevenLabs and PlayHT streaming TTS produces conversational latency — first audio chunk in under 500ms — that works for support and service conversations. The applicable use cases are constrained: scripted or semi-scripted conversations, simple support flows, appointment scheduling. Open-ended complex reasoning conversations still expose latency that breaks the conversational feel. We assess viability for your specific use case before recommending voice.
No-code platforms (Intercom Fin, Zendesk AI, Drift) are appropriate when your intent coverage is narrow, your knowledge base is simple, and you do not need custom integration or complex resolution logic. Custom implementation is appropriate when you need deep integration with internal systems, complex multi-step resolution flows, voice agent capabilities, or multimodal input handling that platforms cannot support.
Sensitive topic detection runs on every message — not just at explicit escalation requests. Detection of distress signals, escalating complaints, or sensitive topics (medical, legal, financial advice) triggers immediate human handoff with priority routing. This is a mandatory escalation path that cannot be overridden by intent classification.
Personality design is the deliberate crafting of conversational characteristics that determine how the agent feels to interact with: response framing, acknowledgment patterns, how it expresses uncertainty, how it handles topic boundaries, and how consistent its persona is across conversation turns. The "uncanny valley" problem — fluent but robotic — is usually a personality design failure, not an LLM capability failure. Well-designed conversational personas feel helpful; poorly designed ones feel like a FAQ search with extra steps.
Tell us what you are building. We will scope it, price it honestly, and give you a clear plan.
Free 30-min scoping call
Fordel Studios