InvoiceGrid is coming soon. We're building it for Australian business owners who send many invoices to other businesses. The aim is to give owners time back while looking after the customers who receive those invoices. Reminders are one part of the invoice work we're planning to take on.
Building InvoiceGrid to work within your limits
Owners choose which work to delegate
You keep raising invoices in Xero. We're building InvoiceGrid to bring new ones in, check their progress, manage follow-up and return customer answers to your records. We want to bring you the decisions that need judgement, with the information to make them.
The owner view is being designed to show what has been delegated. On the customer side, we're including ways to explain a problem or report a payment. That gives the business something to respond to when a reminder alone would not help.
Our plan sets limits on AI's role
Our design starts with reminders waiting for approval. You choose which routine actions may run within your rules, and can keep that approval step in place. Business decisions that need human judgement stay with an authorised person.
In the design, AI helps with wording. Amounts, due dates and payment details are supplied from the business's records. AI is not being given authority to change bank details, move money or decide a dispute.
We're designing checks to run again before an action, so a changed invoice can be picked up.
Records to help explain the invoice work
The record is being designed to distinguish what was checked, proposed, approved and actually sent. An owner could then see whether an action is waiting for approval, sent or held for review.
The design holds an affected action when information is stale. A draft that fails its checks is rejected or brought to a person for review. Disputes return to the owner. A customer's payment report stays separate from a payment matched in Xero.
The OAIC's guidance on using AI products describes the risk of inaccurate outputs and the value of oversight and checks. We want our controls to make problems easier to spot, contain and correct.
Our test plan includes disputes and unclear replies
Our evaluation plan includes ambiguous replies, disputed invoices and attempts to get AI to act outside its limits. We're researching how to make the work safe, accurate and understandable, with Australian legal requirements shaping development.
The National AI Centre's implementation guidance covers documented testing, ongoing monitoring and human intervention. These guide work that is still under development.
When assessing an invoice service, we'd ask for a demonstration of a stale record, a disputed invoice and a failed check. That would give us a concrete basis for discussing what work could be left in its hands.