Microsoft and multiple other companies are making calculated decisions to sacrifice institutional knowledge — and entire business units — on the altar of future AI supremacy.
Maybe that bet pays off. But the organizations that survive this period are going to be the ones that knew the difference between what AI can replace and what it genuinely cannot.
(Inspired by a LinkedIn post that crystallised something I've been thinking about for a while.)
The Gap Between AI Knowledge and Human Experience
Consider these three scenarios:
AI can explain a design pattern. An experienced architect knows when not to use it — because they've seen it fail.
AI can generate a SQL query. A veteran DBA knows that query will deadlock every Friday at month-end because of a reporting job introduced eight years ago.
AI can summarize an incident. The engineer who has handled similar incidents over a decade recognizes subtle warning signs that never make it into the incident report.
These are not edge cases. They are the situations that define whether a system stays alive or quietly degrades.
Where It Matters Most
These differences matter most in situations involving ambiguity, risk, and accountability.
AI works brilliantly when the problem is well-defined, the context is complete, and the cost of being wrong is low. Institutional knowledge carries its weight precisely in the opposite conditions — when the problem is messy, context is missing, and someone has to own the outcome.
Three Likely Outcomes
As companies make these bets, I expect outcomes to fall into three buckets:
a) Overestimation — and a painful correction. Some companies will discover, after losing experienced people, that they cannot maintain quality or innovate effectively. The institutional knowledge walks out the door and doesn't come back.
b) Successful redesign. Some companies will successfully redesign work so that fewer experienced engineers, supported by AI, produce comparable or better results. This is the best case — and it requires deliberate effort, not just headcount cuts.
c) Mixed results — the most likely outcome. Most companies will land somewhere in between. Outcomes will vary significantly depending on the domain, the quality of AI adoption, and how well execution is managed.
My Take
The danger is not AI itself. The danger is the assumption that AI can absorb institutional knowledge simply by processing documentation and tickets.
It cannot. Institutional knowledge lives in judgment — in the thousands of small decisions people make based on things they learned the hard way. You cannot prompt-engineer that into existence.
The companies that figure this out early will have an enduring advantage. The rest will spend years rebuilding what they discarded.