AI chatbot & voice agent development

A chatbot that answers confidently and wrongly is worse than no chatbot. The difference between the two outcomes is not the model, it is whether the assistant is made to look something up before it speaks.

Retrieval-augmented generation is that mechanism: your handbook, price list, service pages or policies get chunked into a vector store, the question finds the relevant passages, and the model answers from those passages with a clear instruction to say it does not know when they do not cover the question. The same grounding applies on the phone, where a wrong answer costs an appointment.

Who this is for

• Clinics, salons and practices missing calls — after hours, during treatments, or whenever the front desk is with someone.
• Companies with a document nobody reads — an HR handbook, a policy library, a product catalogue that generates the same twenty questions weekly.
• Businesses running paid ads where a caller who hits voicemail is a lead already paid for and then lost.
• Teams doing manual call notes — where a transcript and a CRM update after every call is currently someone's job.

What is included

Chatbot and voice builds share the same three layers: the knowledge, the conversation design, and the write-back into your systems. Only the interface differs.

Layer	What gets built
Knowledge base	Document ingestion, chunking and embedding into a Supabase vector store, with a re-index path for updates
Grounding	Retrieval step, system prompt with scope limits, and an explicit refusal behaviour for unknowns
Conversation design	Greeting, qualifying questions, objection paths, and a defined handoff to a human
Voice layer	Vapi agent with phone number, voice selection, call recording, and after-hours routing rules
Booking	Live calendar availability check and appointment creation inside the conversation
Post-call analysis	Transcript parsed for intent and outcome, then tagged and routed to the right pipeline stage
CRM write-back	Contact created or matched, transcript logged as a note, tags and opportunity set
Handover	Prompt documentation, escalation rules, and instructions for updating the knowledge base yourself

How the build runs

1. Scoping call. Which questions the agent must answer, which it must never attempt, and what a successful conversation ends with.
2. Knowledge audit. We look at the documents you have. If the answers do not exist in writing anywhere, that gets fixed before the bot gets built.
3. Retrieval build. Ingestion, chunking and embedding, then testing retrieval alone against real questions before any conversation layer exists.
4. Conversation and voice. Prompt, flow, escalation and booking wired up, then tested by call.
5. Adversarial testing. Deliberate out-of-scope, ambiguous and hostile questions, checking the agent declines rather than invents.
6. Handover and tuning. Two weeks of prompt tuning against real transcripts after go-live, because the first week of real users always finds something.

Proof

Scope and timeline

Note that AI agents carry running costs beyond the build — model usage, voice minutes, and phone numbers. Those are yours directly, on your own accounts, and get estimated during scoping. The maths on whether a voice agent pays for itself walks through the missed-call arithmetic.