Wizbit← Back to home

Five systems you can
open right now.

Not case studies. Not screenshots of somebody else's results. An AI voice receptionist you can talk to, a dashboard you can open, and a workflow you can download and run yourself. No signup, no form, no email.

Running nowLive

AI VOICE RECEPTIONIST

A receptionist that answers the phone

Speak to it in the browser. It hears you, works out what you asked for, prices it against a real treatment list, checks a real diary for free slots and books you in. When it is not sure, it hands over to a human instead of guessing.

~0.7s
to first sound
4.9-5.7s
to a full spoken answer
Deepgram Aura-1n8nAnthropicWeb Speech APIGoogle Sheets
Call it now

Latency measured on our own build, CLAUDE.md 2026-08-26. Aura-1 was chosen over Aura-2 because 2.0-2.6s beat 4.6-5.2s for the same sentence.

MED-SPA DASHBOARD

The screen a clinic actually opens

Speed-to-lead, no-show rate, revenue booked, and a short queue of the few things that genuinely need a person. Running on sample clinic data against a live n8n webhook, not a screenshot.

n8nGoogle SheetsWebhooks
Open the dashboard

The four metrics named are the ones public/medspa-demo.html renders.

WHATSAPP BOOKING AGENT

A booking agent that lives in WhatsApp

Qualifies the enquiry, answers treatment and price questions from a fixed list rather than from a model's memory, offers real slots by day, and writes the confirmed booking into the diary. Clinical questions get refused, on purpose.

WhatsApp Business APIn8nAnthropicGoogle Sheets
Watch it book a client

The refusal behaviour and the fixed-facts answering are in automation/build-medspa-06.mjs.

REAL-ESTATE LEAD FOLLOW-UP

The lead that gets answered at 11pm

Portal enquiry lands, gets logged with both timestamps, and gets a reply on the official WhatsApp Business API in under sixty seconds. Follow-ups queue at 24 and 48 hours and stop dead the moment the lead replies.

The n8n canvas for the real-estate lead follow-up workflow
<60s
first reply, any hour
24h / 48h
follow-ups, then stop
n8nWhatsApp Business APIGoogle Sheets
See how it works

Three real screenshots of this system running are on the homepage: the n8n canvas, the sheet filling up, and the WhatsApp thread as the buyer sees it.

OPEN SOURCE

The voice workflow, yours to take

The whole n8n AI voice receptionist as importable JSON, plus the TTS proxy, MIT licensed. No email, no form, no signup. If you never speak to us and just run it yourself, that is a fine outcome.

n8nMIT
Download the workflow

public/templates/ - the JSON, the proxy and the LICENSE are in this repo.

Patterns we build

These six are not client work and are not presented as it. They are the shapes we build most often, described by how they run and by the part of each one that is genuinely difficult.

Lead capture

Lead-to-CRM Engine

  1. Web / portal / DM
  2. Parse, dedupe, enrich
  3. CRM + owner alert

Enquiries arrive across web forms, portal emails and DMs, and get keyed into the CRM by hand - so they arrive late, arrive twice, or do not arrive.

The hard part

Deduplication. The same person enquiring twice from two sources is one lead, and deciding that reliably - across different payload shapes, with names spelled differently - is most of the build.

n8nHubSpotClearbit

Support triage

Support AI Triage

  1. Ticket in
  2. Classify + route
  3. Draft reply, human sends

Every ticket lands in one queue and a human reads all of them to find the few that are urgent, so first response time is set by queue depth rather than severity.

The hard part

Knowing when not to answer. A confident wrong reply to an angry customer costs more than a slow right one, so the classifier is tuned to escalate on uncertainty rather than guess.

OpenAIIntercomSlack

Reporting

Ops Reporting Bot

  1. Pull 6 sources
  2. Reconcile + aggregate
  3. Post to Slack

The Monday report is assembled by hand from six systems, which means it is stale before it is read and nobody rebuilds it when a source changes.

The hard part

Disagreeing sources. Two systems reporting different totals for the same week is normal, and the report has to say which one it trusted and why instead of quietly picking one.

n8nGoogle SheetsSlack

Document intake

Invoice Processing

  1. Email / PDF in
  2. Extract + validate
  3. Ledger, held for approval

Invoices arrive as PDFs in an inbox and someone retypes the totals into the ledger, which is slow and produces errors nobody finds until reconciliation.

The hard part

Never posting unattended. Extraction is confident and occasionally wrong, so nothing reaches the ledger without a human approving it, and every run keeps the source document attached.

MakeOpenAIQuickBooks

Inventory sync

Multi-Channel Order Sync

  1. 3 sales channels
  2. Reconcile stock
  3. Warehouse + fulfil

Stock lives in three storefronts and a warehouse system that do not talk, so the same unit gets sold twice and someone finds out at fulfilment.

The hard part

Ordering. Two channels selling the last unit in the same second is the case that matters, so writes are queued and idempotent rather than fired as they arrive.

n8nShopifyAmazon API

Recruitment

Recruitment Pipeline

  1. Applications in
  2. Screen + score
  3. Schedule + notify

Applications pile up, screening happens in batches when someone has time, and good candidates take another offer while waiting to hear back.

The hard part

Scoring you can defend. A rejection has to be explainable, so the model returns its reasons against stated criteria and a human owns the decision rather than the score.

MakeOpenAICalendly