The argument in brief· 3–5 min read· full book 59 pp
What a new engineer actually needs, and what a chatbot cannot give them.
The mistake is not adopting AI. It is adopting autonomy. Onboarding’s binding constraint is relational context, not retrievable information — and the instrument the industry uses to evaluate the change is biased toward the automated option.
Onboarding a new engineer quietly became a link to a knowledge base and a link to an assistant: ask it anything. The trouble is that the thing onboarding is for was never mainly information. Gallup’s framing of the failure is unusually blunt about what gets lost:
“This failure gets in the way of the formation of an emotional bond between the new hire and the company – a connection that can make or break retention.” Gallup, on bad onboarding
And the largest multiplier in that analysis is not manager involvement. It is the least writable thing in any company — the question a document cannot answer:
“Do you understand how things get done here?”
That is tacit, relational and local. It lives in who winces when you propose something, which service nobody touches on a Friday, and which decision was already litigated two years ago. A retrieval system can tell you what the code does. It cannot tell you what the room knows.
So does AI-mediated onboarding actually under-serve people? The honest answer is that the evidence is thin in a specific, correctable way — and that is the finding rather than a hedge. The only direct controlled comparison is small enough that its own authors say so plainly:
“We recruited seven participants from Beko: two domain experts and five developers onboarding onto Bankhet.” The only direct comparison
Seven people, and a result reported as “approximately 10 minutes faster (~25 vs ~35 min)”. That is a real measurement of a real thing. It is not a basis for restructuring how a company brings people in.
The deeper problem is the shape of the evidence rather than its quantity. Two unrelated studies found that what people report about an AI tool comes apart from what measurement shows they learned, and the error runs in a known direction: it favours whatever felt smoothest. Meanwhile the studies that test understanding run for an hour, the studies that run for weeks count merged code, and nothing occupies the space between — which is exactly the span onboarding actually takes.
Adopt the tool. Keep the human in the loop that forms judgement, and measure understanding over the weeks it really takes rather than the hour that is easy to instrument.
The thesis is defensible and undertested, and the book says so rather than overclaiming. Two other small studies show AI-mediated and text-based onboarding performing well on their own measures. What is decisive is not the disagreement but the missing middle.
59 pages. 22 sources captured and hashed, 54 claims bound to exact quotations, each checked against a stored copy before the sentence citing it was written.