The worry is fair.
37% of NZ small businesses in Xero’s survey said they don’t trust AI output (Xero, May 2026). AI tools can state something wrong with total confidence. The answer is not to avoid them, it’s to decide where they’re allowed to act alone.
Where AI is reliable, and where it isn’t.
- Good at: first drafts, summaries, sorting and tagging, reading a messy document into fields a person confirms.
- Risky at: numbers, dates, prices, legal or medical statements, anything a customer acts on without checking.
- Wrong tool for: anything that must follow the same rule every time. Use plain automation for that; it doesn’t guess.
Keep a person in the loop.
Design the process so AI prepares and a person approves: the reminder email is drafted, you press send; the invoice is read, you confirm the total. In the systems we build, that approval step is part of the workflow, with a record of who checked what.
What that looks like.
Invoices drafts the reminder to a late payer, and a person decides to send it. Hero summarises a meeting into decisions and actions you can correct before anything goes out.