| Dimension | Score |
|---|---|
| Output reliability | 3/5 |
| Workflow fit | 3/5 |
| Context handling | 2/5 |
| Learning curve vs. payoff | 4/5 |
| Failure transparency | 2/5 |
| Overall | 2.8/5 |
What AI does well
First-draft company summaries. ChatGPT and Claude turn a 10-K and a set of earnings call transcripts into a readable one-page brief in minutes. I timed this against a junior analyst doing it by hand: the AI draft took 3 minutes, the human draft took 40. The catch is the AI drops nuance from management’s tone on the call, so you still need to skim the transcript yourself for anything that sounds evasive.
Fast comps pulls and sanity checks. Perplexity is good at pulling recent multiples for a peer set and citing where each number came from, which matters when a manager asks “where did that 8.2x come from.” It saves the browser-tab hunt across four data providers. It still misses companies that changed their ticker or got acquired in the last quarter, so you have to verify the peer list is current.
Formula and formatting help inside the model. Excel Copilot writes a XLOOKUP or a nested IF faster than most analysts type it, and it can explain what an inherited formula does before you touch it. On a messy legacy model, that explanation step alone saves real time. It does not know your model’s conventions, so it will happily suggest a formula that breaks your existing named-range structure.
Market and macro summaries. ChatGPT and Claude condense a week of Fed commentary or sector news into a paragraph you can drop into a morning note. This is genuinely faster than reading five newsletters. The summary reads confidently even when the underlying sources disagree, so you lose the disagreement unless you ask for it explicitly.
Where it fails
Model auditing. I asked Claude to review a three-tab DCF model for circular references and sign errors. It caught one obvious formatting issue and missed a hardcoded growth rate that should have flowed from an assumption cell. The model looked reviewed. It was not. If you ship that model to a client believing it was checked, that is a real error, not a stylistic one.
Numbers with no source attached. ChatGPT will state a company’s EBITDA margin as fact without telling you if it is trailing twelve months, last fiscal year, or a stale training-data figure. In finance, the period matters as much as the number. You end up re-pulling every figure from the source filing anyway, which erases a chunk of the time saved on the summary.
Cross-session memory of your assumptions. None of these tools remember that your model used a 9% discount rate last week, so every new session starts from zero context. You re-explain your assumptions every time, which is the opposite of how analysts actually build on prior work.
My take
My take (August 2026, Bernat Sampera)
AI earns its place in the research and drafting stage of an analyst’s day. ChatGPT and Perplexity genuinely cut the time to a first draft, and Excel Copilot is a real help on formula-heavy grunt work. That is where it stops. None of these tools will audit your model the way a second analyst would, and none of them flag when a cited number is stale or from the wrong period. I compared this closely against Perplexity vs ChatGPT for sourcing accuracy, and Perplexity wins on citations but still misses recent corporate actions. The honest read: use AI to get to a first draft faster, then check every number the way you always did. The tools do not replace verification. They just move it later in the process.
What I would build
The real gap is model memory. An analyst’s context is not just the current spreadsheet, it is the assumptions, conventions, and prior review comments attached to that specific model across weeks of iteration. A useful system would attach persistent context to the model file itself: every assumption change, every prior reviewer note, every source citation, retrievable the next time anyone opens that workbook. Right now every AI session starts blind and treats a model you have worked on for a month the same as a file it has never seen. Fixing that means the tool stops re-asking what your discount rate is and starts catching when this week’s number contradicts last week’s.
Verdict (August 2026, Bernat Sampera): AI tools cut the drafting time on comps memos and market summaries by half, but none of them will catch a broken formula reference in your own model, so you still check every number by hand. Overall: 2.8/5.