"AI billing" marketing usually means one of three unrelated things: a chat box on top of a rating engine, a model that guesses invoice lines, or a slide that says "agentic" fourteen times.
Finance teams need a narrower question: where does judgement actually live in contract-to-cash — and where must the math stay deterministic?
Judgement steps vs ledger steps
| Step | Judgement? | Sonic approach |
|---|---|---|
| Read a signed PDF | Yes | Structurer + operator review before approve |
| Suggest collections next action | Yes | Watchtower suggestions; human sends |
| Interpret a messy bank line | Partial | Match candidates + operator confirm |
| Compute seat proration | No | Calendar units of unit_price_basis |
| Rate a shipment row | No | Contract formulas + published indexes |
| Add lines to sent invoice | Never | Immutability; credit note or new invoice |
AI belongs on the left column. It does not belong on inventing amounts after send. The schedule and rate card are the source of truth for money; the model is not a second general ledger.
Contract parsing — draft, not decree
The contract pipeline produces a reviewable draft: customer, products, phases, one-time fees, deferred instalments, payment gates, volume companions, renewal clauses bracketed by org parsing context.
Operators fix ambiguity on the review screen because legal language is ambiguous — especially:
- Whether a setup fee gates subscription start vs merely invoices on its own date.
- Which products sit in a minimum commitment group.
- Whether billing day follows contract start when the PDF is silent.
Approve materialises catalog entities and the schedule. That gate is the control, not the model's confidence score.
Collections — accelerate tone, not relationships
Watchtower evaluates sent unpaid invoices. AI refresh can propose the next email or task. Operators approve, edit, or send from a person when the account needs a person.
Failure mode to avoid: automation that chases harder while bank rec still shows an unmatched deposit. Matching cash stops the cadence; smarter copy does not.
Bank and statement lines — candidates, not verdicts
Bank reconciliation compares amount, reference, and payer name (customer or shipper) to open invoices. Near-matches surface as variance — fees, FX, partials.
AI may help interpret descriptions; it does not mark paid without a human-confirmed match on the rules you care about. Payment processor webhooks mark card payments; wires still need the queue.
What we will not claim
- ASC 606 / Ind AS 115 certification.
- Autonomous GL posting without your chart mapping.
- Replacing Stripe, Razorpay, or your bank.
- Correcting sent invoices by regeneration magic.
Those boundaries are product features. They keep the chain trustworthy.
Watch out for
- Review before approve on every new contract shape. The parser learns org context; it does not read minds.
- Do not auto-send dunning until contacts validate. Broken merge fields scale embarrassingly.
- Train evaluators on immutability. If the demo promise is "AI fixes invoices," you are evaluating the wrong product.
- Freight AI is still mostly mapping + formulas. The judgement is shipper identity, not poetic invoice prose.
- Suggested next steps are drafts. Approve them like you would a junior analyst's email.