A March 2026 Harvard Business Review article has a useful name for a familiar failure: “trendslop.” The authors looked at what happens when large language models are asked for strategic advice. The result, as the article frames it, is advice that sounds like strategy because it compresses visible patterns into fluent executive language, but lacks the local constraints and tacit knowledge that make strategy real.
That word matters because it points at the wrong test.
Most companies still talk about AI risk as if the problem is bad facts. Hallucination. Fabricated citations. Wrong numbers. Those are real problems, but they are also the easiest ones to notice. A bogus source can be checked. A wrong date can be corrected. A fake customer quote leaves a trail.
Trendslop is harder. It can be factually clean and still useless.
The failure is not that the model invents the world. The failure is that it knows the map of how people talk about the world. It knows the public vocabulary of competitive advantage, market entry, platform shifts, customer segmentation, and operational focus. It can produce the document shape that organizations already reward: crisp bullets, balanced tradeoffs, a confident recommendation, enough caveats to feel mature.
That is exactly why it is dangerous.
The more polished the output, the easier it is for an organization to mistake completed language for completed thinking. Strategy is not the prose artifact at the end. Strategy is the collision between a claim and a specific operating reality: incentives, distribution, customer trust, internal politics, capital constraints, timing, and the ugly facts that rarely appear in public summaries.
The counterargument is fair: plenty of human strategy work is trendslop too. Consultants and executives have been producing fluent abstractions long before ChatGPT. If AI makes cheap generic advice cheaper, maybe that is clarifying. Maybe it forces the human work to move up a level.
I think that is only true if the organization has a way to measure the difference.
If the review process rewards polish, AI wins immediately. If the review process asks who has talked to customers, which constraint would kill the plan, what would change the recommendation, and what evidence would make the team abandon it, AI becomes useful in a narrower role. It can draft the map. It cannot certify contact with the territory.
This is the measurement problem under AI in miniature. The old proxy was output: pages written, decks produced, recommendations delivered. AI makes that proxy collapse. It can generate the surface area of strategic work faster than the organization can develop judgment about whether the work is any good.
So the real governance question is not “Should we use LLMs for strategy?” That question is already being answered in practice because the tools are too useful to ignore.
The better question is: what must remain non-delegable?
My answer is the part where a claim gets attached to local reality. The tool can summarize the market. It can propose options. It can surface standard risks. But someone has to own the act of saying, “This is true here, for us, under these constraints, and we are willing to bet on it.”
That sentence is strategy. Everything before it is drafting.