How the AI works

How TabNudge's AI handles your customer replies.

What it reads, what it drafts, what it sees, what it doesn't see, where it runs, and how the human-in-the-loop guarantee works. No marketing language.

01
What the AI actually does

Reads, classifies, drafts, picks a tone, knows when to stop.

Reads the reply
Reads incoming replies from your customer.
Classifies the reply
Buckets it into paid / disputing / asking for time / silent.
Drafts a response
Writes a response that matches the situation.
Picks a tone
Uses the tone preference selected for the workspace — tradie, professional, or firm.
Stops if needed
Pauses the sequence if the conversation has moved on.
02
What data the AI sees

Just the reply, the invoice, the thread, and the tone.

Reply text
The text of the customer reply.
Invoice metadata
Invoice number, amount, and due date.
Thread history
Previous follow-ups in that email thread.
Tone preference
The selected tone for the workspace.
03
What data the AI does not see

Other customers, banking, anything off-thread.

Other invoices
Other customers on the workspace.
Banking
Bank or payment details — TabNudge never touches these.
Off-thread data
Anything outside the specific email thread.
04
Where the AI runs

OpenAI for drafts. Vercel for the app. Supabase (Sydney) for data.

AI provider
TabNudge uses OpenAI for draft generation.
Lifetime
Customer data is passed to OpenAI via API for the duration of the request only.
Training
Per OpenAI's API terms, data submitted via the API is not used to train OpenAI's models by default.
Application infra
The TabNudge application runs on Vercel infrastructure.
Database
Hosted on Supabase in the Sydney (AU) region.
05
The human-in-the-loop guarantee

Nothing sends without your say-so.

Always reviewed
Every draft is reviewed before sending.
Approval required
Nothing leaves your account without your approval.
Editable
Any draft can be edited before send.
Pause anytime
A sequence can be paused or stopped at any time.

For the connection, encryption, and compliance detail, see /security.