AI reception: one window for inbound requests across every channel

Context

Inbound requests arrive through a site form, Telegram, email and the phone — and each channel has its own owner, its own delay and its own way of losing a lead. The client repeats themselves in every channel; the qualified request waits behind the unqualified one; nobody owns the whole picture.

The service collapses those channels into one window with a single behavior: understand, answer, qualify, route.

Approach

The core design decision is normalization before intelligence. Every inbound message — a form submission, a Telegram chat, an email, a call transcript, a CRM event — is reduced to one canonical request shape and placed in a single queue. Everything downstream operates on that shape, so adding a sixth channel means writing one adapter, not rethinking the pipeline.

Only after normalization does the LLM read intent, answer what it can from the knowledge base, and score qualification before a human ever sees the request.

How it works

Each request moves through four stages. Understanding: the LLM classifies what the client actually wants, independent of how it was phrased. Answering: routine questions are resolved directly from the knowledge base, with the source material as ground truth. Qualification: the request is scored against the client's criteria — budget signals, urgency, fit. Routing: the request lands with the right employee, carrying the full conversation context, and a notification fires so it cannot sit unseen.

Humans enter the loop exactly once, at the point where judgment is actually required.

Key engineering details

The knowledge base is the accountability layer: automated answers are grounded in curated company content rather than the model's general knowledge, which keeps the first line auditable and safe to put in front of clients. Where the base has no answer, the request escalates instead of the model guessing.

Routing carries context deliberately — the receiving employee sees the original message, the automated exchange and the qualification result, so the handoff costs the client zero repetition. CRM integration writes the request into the pipeline the sales team already uses.

Outcome & what shipped

What shipped is a pilot-ready product line: the channel adapters, the normalized queue, the four-stage LLM pipeline, knowledge-base answering, qualification scoring and context-preserving routing with notifications.

It deploys as an AI reception, a first line of support, a lead-qualification layer or a single inbound window — wherever requests currently scatter across channels and go cold. Planned next steps: piloting on a live multi-channel inbox, tuning qualification rules to the client's funnel, and SLA timers with escalation.

What we built

• Channel intake

  Site, Telegram, email, CRM and telephony normalized into one canonical request shape and a single queue.

• Intent understanding

  The LLM classifies what the client actually wants before anyone triages the request.

• Knowledge-base answers

  Routine questions answered from grounded company knowledge; no answer means escalation, not improvisation.

• Qualify & route

  Requests scored against client criteria and handed to a named employee with full context and a notification.