Google posts ran on autopilot. Review replies drifted off-brand.
A Charlotte founder building AI for Google Business Profiles, an Australian SEO agency co-founder and a UK invoicing founder all hand AI the repetitive work, then keep a person between the draft and the customer.
Google Business posts ran on AI with no human input after setup.
Review replies on autopilot drifted off-brand, and a few customers said so.
Two more founders, in Australia and the UK, draw the line in the same place: before a customer sees it.
We asked which parts of a job AI has actually taken, and which still need a person. These three founders, working in local search, SEO and invoicing, kept landing on the same checkpoint: the send button. Drafting is cheap now. Sending is where the cost lives.
Posts ran themselves. Replies needed an owner's eyes.
Amin Said, Founder of Pure Technology Consulting LLC (Charlotte, United States)
“In building AI Local Boost I saw AI take over post generation and directory sync for Google Business Profiles without any human input after setup.”
“The clearest mistake came from letting AI run review replies on autopilot before the business owner reviewed the first batch. Tone drifted and a few customers flagged the answers as off-brand, which slowed profile momentum until the loop was closed.”
Posts and directory sync could run alone, but replies that answer a real customer needed the owner first. The lesson is about order: review first, autopilot second.
SEO data went to AI. Client emails didn't.
Harvie Williams, Co-Founder of Enkindle Group (Craigie, Australia)
“An agent can draft a client email in seconds, but it can't tell the tone is wrong for someone who's had a rough month, or that a line about the client's business was never confirmed by them.”
“If a mistake can be fixed quietly, the agent does it. If it reaches a client or the public, it's both. The agent prepares, a person approves.”
Weekly Search Console and GA4 reporting moved to AI completely. Anything a client reads is a different job, and the line follows who sees the work, not how hard it is.
Right on the page, wrong on the bill.
Kamal Leadbitter, Founder of SaidToPaid (Elton, United Kingdom)
“An invoice can look right and still have the wrong charge.”
“A quick check can catch something that would otherwise cost you money or your customer’s trust.”
Turning job notes into an invoice draft is automated at SaidToPaid. Human approval before the customer sees it is not, and the case for it is simple: the check is cheap next to what it can catch.
Where they split
Where they agree: nobody wants the person gone, and nobody wants the draft written by hand. Everyone keeps the speed.
Where they differ: what triggers the check. For Said it was the first batch, for Williams it's reach, for Leadbitter it's the invoice itself.
My take
AI rarely takes a whole job. It takes the repeatable output, then hands the risky moment back to a person. Here the risky moment is easy to spot: it's whatever a customer will read. Automate everything behind that line, and put a human check on whatever crosses it, at least until the first batch proves the tone holds. Then ask of every task: who sees the output, and who answers if it's wrong?
Your turn
Which task in your company runs with no human input today, and which one would you never let reach a customer unchecked?
Next: a skills-data founder, a social-scheduler founder and a fintech CEO on moving the question from who does the task to who answers for it.
