Issue the invoice
From contract terms, milestones, approved time, or usage — so the invoice matches what was agreed rather than what someone remembered to bill.
AI agents · finance
It issues the invoice, applies the cash, watches the ageing, and drafts the chase — after checking whether the customer is mid-renewal, has an open dispute, or paid yesterday. That context is the difference between collections that work and collections that cost you accounts.
What it does
Each of these is work a person does today. The agent does them in sequence and stops at the first thing it is not confident about.
From contract terms, milestones, approved time, or usage — so the invoice matches what was agreed rather than what someone remembered to bill.
Matches payments to open items including partials, short payments, and lump sums covering several invoices. Deductions are classified rather than left as a variance.
By recoverability rather than by days outstanding, using payment behaviour, amount, contract status, and dispute state together.
A message pitched to the relationship and the history, ready for a person to release. It knows the difference between a habitually slow payer and a new problem.
A customer whose payment pattern is slipping surfaces before it becomes a write-off conversation, with the change in behaviour quantified.
Active renewal, open escalation, disputed line, or unapplied payment all suppress dunning automatically rather than requiring someone to remember.
Authority
The agent prepares invoices and chase messages; a person releases them. Every new customer starts here for every account.
Enabled per segment and per threshold — typically friendly reminders under a set amount, once the tone and suppression rules have been reviewed for a quarter.
Final demands, referrals to collections, and bad-debt write-offs require a named human approver. These affect a relationship and a balance, and are never delegated.
Dunning automation has a bad reputation among finance teams, and it is deserved. Almost everyone who has run it has a story about a demand letter that went to a customer mid-renewal, a reminder to an account with an open escalation, or a chase to someone who paid four days ago and whose payment had not been applied.
None of those are AI failures. They are consequences of a collections tool that can see an ageing report and nothing else. The information that would have prevented each one lived in the CRM, the support system, or the unapplied cash queue.
Because contracts, tickets, projects, and payments are objects on one graph here, the agent can check all of it before drafting. Suppression for renewals, disputes, and recent payments is a default rather than a rule you have to remember to write.
Most AR teams spend more hours applying cash than chasing it, and nobody puts that on a slide because it is unglamorous. A lump-sum wire covering eleven invoices with a short payment on one is a twenty-minute puzzle, done several times a week, by someone whose time is worth more than that.
The agent handles exact, split, and partial matches, classifies deductions — short payment, discount taken, disputed line — and holds the genuinely ambiguous. What is left is a real exception list rather than a pile, and unapplied cash stops being a month-end project.
It will not send a formal demand, refer an account to collections, or write off a balance without a named human approving it. Those actions change a commercial relationship or a reported number, and both belong to a person regardless of how confident the model is.
Questions
Send an ageing export and we will rank it by recoverability, with the reasoning attached.