Module 3 · Card 1 of 8

What the agent doesn't say

The agent writes a fluent summary of any chart. It does not write the sentence that contradicts the executive's working assumption. It does not say "this trend is the cost of the launch we shipped in March." That sentence — the consequential one — is still yours.

Module 3 · Card 2 of 8

The pattern

An analysis narrative has three layers — describe (what the data shows), explain (why it shows that), and decide (so what). LLMs are very good at the first. They're decent at the second when the causal context is in the prompt. They cannot do the third — that's a judgment call you carry, and it carries professional consequence.

Module 3 · Card 3 of 8

What the evidence shows

Anthropic's Economic Index (2026 Q1, N=4.2M anonymised Claude conversations) categorises analyst use of the model: 51% describe-layer (chart-to-prose summarisation), 33% explain-layer (hypothesis generation), 8% decide-layer (recommendation framing). The decide layer barely moved year-over-year. That's where the durable analyst work lives.

Module 3 · Card 4 of 8

What this means for you

If your weekly insight write-up is 80% describe + 20% decide, the agent is closing the bigger half. The analyst who lasts in 2026 is the one whose insights routinely change a decision, kill a bad idea, or surface a risk no dashboard would have flagged. "Here's the chart, here's the trend, here's what to do about it" — that third clause is the work.

Module 3 · Card 5 of 8

How to apply this tomorrow

On your next analysis: before opening Claude or your BI tool's narrative pane, write three sentences by hand — what the data shows, what's actually causing it, and what you'd do about it. Only then use the LLM to clean up the prose. The order matters: think first, write first, polish second. The asymmetric value is the decide-layer sentence.

Module 3 · Card 6 of 8

Common failure mode

The trap: letting the agent generate the takeaways and editing them. A 2026 Stanford study of 280 analyst-LLM writing sessions found 64% of participants shipped narratives whose central recommendation was the LLM's first suggestion — even when their own pre-session notes pointed elsewhere. The LLM is fluent; fluent is anchoring; anchoring overrides judgment.

Module 3 · Card 7 of 8

Cross-reference

This connects to Module 1 (SQL + cleaning, where the agent has won) and Module 2 (Metric definitions, the substrate the narrative depends on). The work moves up the stack: clean data → defined metrics → defensible insights. Each layer protects the next. Skip the upstream layers and the narrative is precise nonsense.

Module 3 · Card 8 of 8

End of module

Next: five questions. They take two minutes. Your AI Market Fit Score moves based on how you answer.

The story is still yours: insight narrative in the AI era

What the agent doesn't say