AI reads the press requests. A human sends the email to the journalist.
A Brisbane founder, a Tallinn founder and a US founder hand AI the reading and the drafting, then keep the send button for a person.
An AI reads every press request. A person writes to the journalist.
That is the split in three very different companies: one in Brisbane, one in Tallinn, one in the US.
Nobody here is anti-AI. They hand over the reading, the filtering and the drafting without a fuss. Those parts work, and nobody is pretending otherwise.
The send button is another matter. That is where the machine stops and a person signs.
We asked which parts of a job AI has actually taken, and which still need a person. Three founders drew the line in almost the same place. It matters now because every company is deciding, role by role, what to hand over. Here is where they put the line, and what each one calls it.
The reading went to the bot.
Jennifer Grehan, Founder of The Core Collab (Brisbane, Australia)
“Every morning AI reads every press request we receive, filters out what doesn't fit, and puts a shortlist in my calendar.”
“Give AI the work that's about information. Give people the work that's about trust.”
Reading, filtering and shortlisting are information work, and that is exactly what AI took. The second quote is her rule for everything else, short enough to fit on a sticky note.
Anything that commits the company stays human.
Christian Rios, Founder of Unfortunately Co. (Tallinn, Estonia)
“Work that would have been a developer, a marketer and an analyst is now mostly agent work.”
“What stayed human is anything that commits the company: sending an email to a journalist, approving a print order, paying a fee, accepting terms, pushing a release live.”
A developer, a marketer and an analyst is a lot of headcount to fold into one kind of work. Notice the list that stays human is not about difficulty. It is about commitment: once it is sent, paid or live, someone has to answer for it.
The agent drafts. It cannot send.
Tom Pinder, Founder of StartVest (Hampstead, United States)
“In PRAPI, the agent drafts every pitch but can’t send one. That’s deliberate: a human approves each send, because a wrong claim to a journalist costs more than the time saved.”
“Companies automate output, then forget review, provenance, and refusal rules.”
The drafting is the visible task. The approval step is where the risk lives, and his warning is about companies that automate the first and forget the second.
Where they split
Where they agree: AI owns the reading, the filtering and the drafting. Grehan's shortlist, Rios's agent work and Pinder's pitches all point the same way.
Where they differ is the test. Grehan asks whether the work is about information or about trust.
Rios asks whether it commits the company. Pinder asks whether being wrong costs more than the labor saved.
My take
The line doesn't run between jobs. It runs through the middle of them. Before you automate a role, find its send button and decide who gets to press it. Then ask what it costs if the wrong thing goes out, because that tells you who should hold the button.
Your turn
Where is the send button in your job: the step AI can draft but a person still has to approve?
Next: three founders who automated the visible front end and found the gap behind it: a sale that never became an order, a message sent that should not have been, and a smart cooler nobody restocks.
