LLM ENGINEERING
Conversational AI development for assistants that finish the job. They hold context across a conversation, retrieve from your own content, take real actions in your systems, and hand over to a person cleanly - with a transcript you can audit afterward.
What separates a conversational agent from a chatbot
A conversational AI agent is an assistant that holds state across a conversation, retrieves grounded information, calls tools to take real action, and knows when to stop and pass to a person. A chatbot matches an utterance to a scripted reply. The gap between them is not fluency; it is whether anything gets done.
Voice adds constraints rather than features. Latency becomes a design budget rather than a metric, because a pause that reads as thoughtful in chat reads as broken on a call. Transcription errors compound into misunderstandings. The agent cannot show a list, so it has to hold options in a shape a person can follow by ear, and confirm anything consequential back before acting.
The measure we build toward is task completion, not containment. An assistant that keeps someone in the conversation without resolving anything is optimizing the wrong number, and customers experience it as being trapped.
You probably need this if you recognize these
- Your current bot handles the greeting and hands over everything that matters.
- Customers repeat their account details to a bot, then again to a person.
- The assistant answers questions but cannot change anything.
- Call volume is predictable, out of hours, and repetitive.
- Nobody can tell you what the assistant actually resolved last month.
What you get
Conversation design
What the assistant handles, what it refuses, how it recovers from misunderstanding, and how it hands over. Written before any implementation.
Grounded answers
Retrieval over your own content with citations in chat, and spoken source attribution where it matters in voice.
Tool access
Real actions in your systems: look up an order, change an address, issue a credit within a limit, book a slot.
Identity and context
Who is speaking, what they have already told you, and what happened last time.
The voice layer
Speech recognition and synthesis tuned for your vocabulary, with a latency budget and barge-in handling.
Escalation
Warm handover with the transcript, the intent and the recommended next action attached.
Transcripts and evals
Every conversation recorded, scored against a suite, and reviewable.
How a conversation reaches an outcome
Conversational AI development and voice agents
Identify
The assistant establishes who it is speaking to and pulls their context.
Self-healed - retried with fallback tool. Human not required.
Understand
It works out the actual goal, which is often not the first thing said.
Ground
It retrieves from your content and answers from what it found, not from memory.
Confirm
Anything consequential is stated back and confirmed before it happens.
Act
Within its limit, it does the thing. Outside it, it prepares and escalates.
Close or hand over
It confirms the outcome, or passes to a person with the full transcript and context.
Conversations that finish the task
Ticket Triage Agent
The conversational front door to the support queue, resolving or routing with context intact.
Knowledge-Base Agent
Answers from documentation with citations, and says plainly when the documentation does not cover it.
Order & Logistics Agent
The clearest voice case: “where is my order”, answered and acted on end to end.
Where conversational agents run
Voice usually means integrating with the contact center you already run rather than replacing it, which keeps routing, reporting and compliance recording where they are.
- Live chat and messaging
- Telephony and contact center platforms
- Helpdesk and ticketing
- CRM
- Order management
- Knowledge bases
- Identity and authentication
- Speech services
Platform names are shown as examples of the categories agents connect to. They are not partnerships or endorsements.
What the assistant may and may not do
Autonomy boundary
Answer, look up, and act up to a value or risk limit you set. Anything above it is prepared and handed over.
Approval gates
Refunds above the limit, contractual commitments, account closures and anything irreversible all require a person.
What stays human
Distress, complaints, anything legal, repeated misunderstanding, and any explicit request for a person, which is always honored immediately.
Logging
Full transcripts with retrieved sources, tool calls and outcomes, retained under your policy and available for review.
How the work runs
Typical ranges from our engagement model (doc 04 §5), not a quote.
| Stage | Typical | What happens |
|---|---|---|
| Pilot | 3-10 days audit, then 2-4 weeks | Conversation design plus one narrow intent, built and scored against real transcripts. |
| Build | 3-8 weeks | Retrieval, tool access, identity, escalation, the voice layer and the eval suite. |
| Release | 1-2 weeks | One channel and a limited intent set first, widened as task completion holds. |
| Managed | ongoing, optional | Transcripts sampled and scored, intents extended, latency and cost watched. |
What we measure
- Task completion
- The primary metric - not containment, which rewards trapping people (Target)
- Handover quality
- Share of escalations arriving with full context attached (Target)
- Latency budget
- Set per channel and enforced, because voice fails on delay before it fails on accuracy (Target)
- Your baseline
- Resolution rate and handling time today, measured before anything changes (Yours)
“ Until then these are design targets and measurement commitments, not results.
What this looks like in practice
Support triage hive
Reference scenario · SaaS. Three channels, one queue, and a first response in minutes rather than hours.
Reference scenario - a composite build illustrating our method. Figures are modeled and the model is shown.
Frequently asked questions
Will customers know they are speaking to an AI?
Yes. The assistant identifies itself, and a request for a person is honored immediately rather than deflected once more. In several jurisdictions disclosure is a legal requirement for voice; everywhere else it is simply the right call, because customers work it out and resent the pretense more than the automation.
Is voice actually ready, or is this a chat product with speech bolted on?
Voice is viable for narrow, well-defined tasks and still difficult for open-ended ones. The constraints are real: latency budgets, transcription errors on names and product codes, and the fact that you cannot show a list. We scope voice to intents that survive those constraints, and we will tell you when a task belongs in chat instead.
How does it hand over to a human?
Warm, with the full transcript, the identified intent, what was already tried and a recommended next action. A handover that makes the customer start again is worse than never engaging them. In voice this means transferring context alongside the call rather than just the call.
Can it actually do things, or only answer?
It acts, within a limit you set - order lookups, address changes, appointment booking, credits up to a value. Everything above the limit is prepared and passed to a person. Assistants that can only answer are why most first-generation deployments disappointed.
What about accents, noise and our product names?
Speech recognition is tuned on your vocabulary - product names, SKUs and the terms customers actually use - and the assistant confirms anything consequential back before acting. Recognition quality varies by accent and environment, so evaluation includes real recordings rather than clean studio samples.
How do we stop it going off-topic or being manipulated?
Scope is enforced in the system rather than requested in a prompt: declared tools, validated outputs, refusal behavior tested as a feature, and injection attempts in the eval suite. Conversations are sampled and scored continuously, so drift shows up in a report rather than on social media.
Related services
LLM Engineering
The layer underneath every assistant.
RAG & Knowledge Systems
Grounded answers with citations.
Customer Support Agents
The queue behind the conversation.