Three experts approve AI output. One also guards what the provider sees
A New York founder and CEO, an Australian founder and an AI automation engineer all keep a human sign-off, but the engineer pushes the question further, to what the provider can see.
Three experts. One shared habit: a person approves before anything goes out.
Nils Abegg says AI can produce "convincing campaigns that feel completely wrong for the company." Melanie Laas has Claude draft every post and still approves each one.
Then Abegg adds the twist. Output is not the only risk. There is also "the knowledge that gives a company its competitive advantage."
Same checkpoint. Different fear.
Claude drafts every post. A human clicks yes.
Melanie Laas, Founder of EMME Influence (South Mackay, Australia)
“Claude drafts every social post for my business, and I approve each one.”
“In my view, most roles mix work an agent can do with decisions a person has to own. Leaders should map those decisions first, then build each role around them.”
Melanie Laas turns the headcount debate into a design problem: decisions first, roles second. Her own workflow is the proof of concept: the agent drafts, the founder decides.
Can AI answer? Wrong test. Can you defend the answer?
Chirag Kulkarni, Founder & CEO of Taco (New York City, United States)
“The key is not whether AI can produce an answer but whether we can explain and support it when needed.”
Chirag Kulkarni lets AI start the work: research, sorting, summarizing. A person owns the finish, because someone has to stand behind the answer.
Approve the output. Then ask what you fed the model.
Nils Abegg, AI Automation Engineer at einfachAI
“AI can produce convincing campaigns that feel completely wrong for the company. I see this myself, and recommend human approval before publication.”
“My biggest concern is exposing the knowledge that gives a company its competitive advantage. A no-training promise still requires trusting the provider to honour it. For information whose disclosure could seriously damage the business, my default is to keep it out of external AI services and use a controlled, self-hosted environment.”
Nils Abegg takes the sign-off for granted and asks the next question: where does company knowledge live? His worry is trust in the provider, not the quality of the draft.
Where they split
Where they differ on why: Kulkarni points to explanation and responsibility. Laas points to decisions a person has to own. Abegg points to fit with the company.
Where they really split: Abegg adds a risk upstream of approval, the information going into the provider.
His default for information whose disclosure could seriously damage the business: keep it out of external AI services and use a controlled, self-hosted environment.
My take
Approval is the easy part to agree on. The harder call is what the agent gets to see before it ever drafts a word. If a role is a bundle of tasks, a sensible split might be this: the agent does the work, the person owns the decision, and the company owns its crown jewels. Before picking a tool, write two lists: what the agent may touch, and what it never sees.
Your turn
Who approves your AI's output today, and what is the one piece of company knowledge you would keep out of an external AI service?
Next: a solo founder runs a team of agents but still had to push back on ten hook lines, while German and UAE implementers say the win is less busywork, not fewer roles.
