What it can genuinely do
- Pull approved scope, change orders and completed work into line items.
- Rewrite technical field notes into wording a customer understands.
- Apply deposits and progress payments already recorded against the job.
- Flag completed jobs with no invoice yet.
- Point out an invoice total below the approved amount plus change orders.
- Prepare the delivery message alongside the document.
What it should never do alone
- Invent or infer amounts the customer never approved.
- Issue or send an invoice without a human approving it.
- Decide tax treatment — that belongs to your accountant's rules.
- Apply discounts, waive fees or agree to payment plans.
- Write off a balance or edit a delivered document.
Why the record quality decides the result
An AI drafting an invoice is reading whatever the business wrote down. If change orders were verbal and job notes say "did the extra bit", the draft will be wrong in exactly the same way a human's would be — just faster.
The prerequisite for useful drafting is structured job data: approved amounts, documented changes, recorded completion. Businesses that fix that get value from drafting; businesses that skip it get confidently formatted mistakes.
How to evaluate the claim
- Ask what records the draft is built from, and whether they are shown next to each line.
- Ask whether issuing requires human approval, and whether that can be turned off.
- Ask what happens when data is missing — a good system stops, a bad one guesses.
- Ask whether the draft is stored as a revision with an audit trail.
- Test it on a messy job, not a clean demo one.