Not a replacement, a lever
A few friends of mine do research at brokerages and private funds. They don’t use LLMs to “have the AI write the report”. They use them to cut the collation work down to a fifth of the time, and spend what’s left thinking.
Typical uses
1. Earnings summaries and YoY/QoQ tables
Dump the full text of a listed company’s earnings report in and have the model pull out:
- Key financial metrics
- YoY / QoQ changes
- Subtle shifts in management’s wording (this is the valuable one)
Two hours by hand, five minutes by model.
2. Merging views across research notes
Ten sell-side notes → have the model extract each shop’s core view, price target, and points of disagreement, then output one comparison table.
| Broker | Rating | Price target | Core disagreement |
| --- | --- | --- | --- |3. Structuring interview notes
A one-hour expert call, a 10,000-character transcript. Have the model sort it into four buckets: industry state, competitive landscape, company differentiation, risks.
What it can’t replace
Judgment and accountability.
A model can tell you what’s in the filing. It can’t tell you whether the company is worth buying. The analyst is still the one signing the investment memo, and a model that gets it wrong can’t carry the compliance liability.
One hard rule
Every number a model produces must trace back to its original source. If the note says “revenue of 3.2 billion” and the model supplied it, you have to be able to find it on a specific page and paragraph of the source PDF. Otherwise it’s a time bomb.