Configuring the AI.

Out of the box the AI sounds like a competent, professional Australian voice that already knows your Xero data. These are the four places you shape that into your business: the persona, your vocabulary, how the tone and cadence work per stage, and the billing knowledge base the AI draws on.

Where to find these settings

Dashboard → Persona for prompt, voice, and vocabulary. Dashboard → Cadence for timing and channels. Dashboard → Knowledge base for your terms and standard answers. Changes are live within seconds. No redeploy.

1. Persona prompt

The persona prompt sets the AI's tone and the rules it follows in every conversation. You start from a preset (Warm and professional / Friendly and casual / Firm and businesslike / Brief and to the point) and edit from there.

What you can change:tone of voice, the AI's name, opening style, sign-off phrasing, what kinds of questions it should defer to a human instead of answering, business-specific behaviour ("always offer a 7-day extension before escalating").

What you can't change: the disclosure opener (AI identity + recording line), the call-recording consent, or the DNCR check on outbound calls. Those are compliance invariants: they ship with every tenant.

Versioning. Every save creates a new prompt_version_id. Calls log which version they ran. If you change something and quality regresses, the dashboard shows you exactly which version of the prompt was in play, and you can roll back.

2. Tenant vocabulary

Vocabulary is the layer between speech-to-text and the persona. It catches the words your business uses that a generic model wouldn't.

Preferred phrases

Words you want the AI to use. Useful for terms your customers expect: "the pay-link" not "the payment portal"; "your account" not "your debt".

Banned phrases

Words the AI must not say. Common bans we see: "definitely", "absolutely", "great question", robotic "Goodbye", and anything that sounds like a threat. We ship a few opinionated bans by default (the words that consistently make AI-ish calls sound AI-ish, including a hard ban on "sweet" as filler); you can add more.

Pronunciations

For the speech engine. Customer and business surnames, suburb names, and brand names the model mispronounces. "Indooroopilly" → in-droo-ROO-pilly. A tricky company name → the phonetic spelling you'd write for a new staff member.

How to grow it:when reviewing a call, transcript words you wouldn't use → click Add to banned. Mispronunciations → click Add pronunciation. The vocabulary grows from real calls, not from imagination.

3. Tone per stage & cadence

The same engine runs the whole get-paid cycle, but the toneis not the same at every stage, and that's the point. A quote is a courtship; an invoice is a debt. Office Voice knows a quote is a nudge and an invoice is a chase, and never points the wrong one at a customer.

What you set

Quote follow-up timing: how long after a quote is sent in Xero before the warm nudge goes out (default ~24 hours).
Invoice cadence: the age thresholds at which the chase steps up (a gentle first touch, then firmer, then a final one, e.g. around 14 / 30 / 60 days overdue). The tone firms as the invoice ages; it never threatens.
Channels: voice, SMS, and email run as one cadence. Turn any channel on or off; voice leads and the others fill in around it.

The guardrails run underneath. The moment Xero marks an invoice paid the cadence stops, and disputes or payment plans go on hold with a resume date, so tuning the cadence harder never risks double-chasing a paid or disputed account. See Workflows.

4. Billing knowledge base

The knowledge base is the structured data the AI uses to talk about how you get paid: your terms and the answers to the questions customers actually ask on a collection call. It sits in your tenant config as a set of fields the composed prompt reads.

What's in it

payment_terms: Your standard terms (e.g. "due on receipt", "14 days"). Lets the AI answer "when was this due?" accurately.
how_to_pay: Always the secure Xero pay-link, sent by SMS or email. The AI never takes card details by phone: this is a hard rule, not a configurable one.
standard_answers: Your approved responses to common asks: "what's this invoice for?", "can I have a few more days?", "can I pay in instalments?".
extension_policy: Whether the AI can offer a short extension or a simple payment plan on its own, and up to what limit, before it needs your approval.
escalation_triggers: What always goes to a human: a genuine dispute, financial hardship, or anything off-script. When in doubt, it escalates.

What it's for

When a customer asks something on the list, the AI answers from your approved wording and your real terms. If they ask about something noton the list (or push into a dispute) the AI captures it and hands off to a human. It doesn't improvise answers about money it wasn't briefed on.

Defaults.New tenants ship with sensible AR defaults: pay-link-only payment, a short standard-answer set, and conservative escalation triggers. Edit them to match your terms; it's your data.

Composition order (for the curious)

When the AI runs, your settings are composed into a single system prompt in a fixed order:

Layer	What it contributes
1. Base	Compliance, banned phrases, opener structure: the invariants that ship with every tenant.
2. Tenant persona	Your persona prompt, AI name, voice selection.
3. Tenant vocabulary	Preferred phrases, banned phrases, pronunciations.
4. Knowledge base	Your payment terms, standard answers, extension and escalation rules.
5. Customer context	Pre-warmed from Xero on phone-number match: their open invoices, sent quotes, payment history, and any promise they made last time.
6. Workflow	The goal, tools, and tone for the stage the teammate is running: a nudge for a quote, a chase for an invoice.

You don't edit this order (it's structural) but knowing it helps when debugging: "why did the AI say X?" traces to one of these six layers, and the conversation review page tags the prompt version it ran.

Common configuration mistakes

Over-scripting answers it should defer.The knowledge base is for the things you're happy for the AI to answer. Disputes and hardship should escalate: don't write clever comebacks the AI will use to argue.
Banning words you don't actually want banned. Bans cascade: the AI will pick more stilted alternatives. Only ban things that genuinely jar when you hear them.
Setting the invoice cadence too aggressive. Firm is fine; threatening isn't (and isn't compliant). The default thresholds are conservative for a reason: tighten them gradually.
Leaving your terms blank. If payment_termsisn't set, the AI can't answer "when was this due?" and will escalate calls it could have handled.

Workflows (how each stage of the get-paid cycle uses your config).
Approval mode & graduation (review prompt changes before they go live for everyone).
Onboarding guide (the first-time setup walkthrough).