Every: what AI clones of decision-makers are actually good at
What the article is about
Mike Taylor, in Every’s Also True for Humans newsletter, documents a series of experiments building AI clones of decision-makers — his direct manager Natalia, GitHub’s COO Kyle Daigle, and CEO Dan Shipper — and testing whether those clones could simulate how each person would evaluate a draft or proposal. The question he is trying to answer is practical: can an AI version of a reviewer tell you what will get approved before you spend time on a real review?
What the clones were good at
The clones performed well at anticipating objections. When Taylor submitted a draft to the AI version of a reviewer, it reliably surfaced the kinds of concerns that person had raised before — skepticism about scope, questions about the framing, preferences about structure. Used as an early filter, the clones helped identify problems worth fixing before a real review rather than during it.
They were also useful for understanding how a specific person’s priorities shape what they pay attention to. Each clone developed a recognizable pattern of response that matched what Taylor had observed in the real person over time.
Where the clones fell short
The clones overused public frameworks. When uncertain, they fell back on widely cited mental models or consulting-style structures that the real people used rarely or not at all. They also drew more heavily on written sources — published posts, recorded talks, documents — than on spoken patterns from meetings or informal conversations, which meant the clone’s voice was more formal and more hedged than the original.
The deeper limitation was that the clones could identify when an objection existed but not decide whether it was worth overriding. Real approval decisions often turn on judgment calls that depend on context the clone did not have — a relationship, a conversation from last week, a strategic shift that had not yet been written down anywhere.
What Taylor recommends
Use AI review as a first pass, not a final sign-off. The clones are good at catching the predictable objections early, which makes the real review more efficient. They are not a substitute for the judgment that determines whether an edge case matters in a given situation.
He also recommends watching for drafts that come back unusually formal or heavily padded — a signal that the clone is compensating for uncertainty by leaning on written sources rather than genuine calibration.
Who it is useful for
Writers and editors who work in environments with multiple stakeholders and long approval cycles, and who want to shorten feedback loops without skipping real review. The article is also useful for teams exploring AI personas or simulated review as part of their editorial workflow.