There is a pattern emerging inside organisations that is worth paying attention to.

AI is being used to make knowledge work faster, but also thinner.

A report gets written with AI assistance. It gets summarised by another AI prompt. That summary is pasted into an email. Someone else runs it through a chatbot for “key points”. By the time it reaches a decision-maker, the original material has effectively been abstracted multiple times over.

What arrives is fluent, structured, and increasingly detached from the source.

It resembles an internal game of telephone - except every step has been optimised for speed, which makes the degradation harder to see.

The issue is not the tools. It is what gets lost between transformations.
Knowledge work is not just information processing. It depends on edge-case detail: the caveat that changes a recommendation, the assumption buried in a footnote, the constraint that invalidates a neat conclusion, the line that says “except in this jurisdiction”.

Multi-stage summarisation systematically strips these away first. What remains is coherence without fidelity.

And coherence is dangerous when it is mistaken for completeness.
The organisational risk is subtle. You end up with documentation that reads well, circulates easily, and steadily diverges from reality. Shared understanding weakens even as output volume increases.

The fix is not to reduce AI usage. It is to change how it is used.

Treat summarisation as a diagnostic step, not a compression step. Explicitly interrogate what has been removed at each stage. Ask what a domain expert would object to. Calibrate output depth to the materiality of the decision being made.

Over time, this becomes a design choice: whether the organisation optimises for readable artefacts or for retained meaning.

The difference compounds.