You have 12 to 24 months of revenue, cost, and margin numbers, and a meeting on Thursday. The job is to find the two or three real movements buried in the noise — not anecdotes from one bad month — and turn each into a hypothesis you can defend. This page is the analyst-side workflow: what to feed the model, the exact prompt, and the validation step that stops you from quoting a number the AI invented.
TL;DR
- Use AI to cluster monthly variance into themes and generate hypotheses fast, not to produce final numbers.
- Feed it a clean monthly aggregate table (12 months minimum), tagged one-time events, and the specific hypothesis you want pressure-tested.
- Re-derive every number the model quotes from the source spreadsheet. On the Vals AI Finance Agent v2 benchmark (June 2026), no frontier model clears 52% on entry-level analyst work — treat output as a draft, never a fact.
- For files larger than ~200 rows, push the data through a code tool (ChatGPT Advanced Data Analysis or Claude’s Analysis tool) so the math runs in Python, not in the model’s head.
When AI earns its place here
AI is genuinely good at three things in this workflow:
- Theming variance. Given a year of monthly numbers, it groups swings into candidate causes — seasonality, product-mix shift, a cost step-change after a vendor switch — faster than a human scanning a sheet.
- Surfacing anomalies you’ve tuned out. The 6% margin dip you stopped noticing three quarters ago gets flagged.
- Killing your false confidence. Asked to, it will name the “trend” you keep repeating in meetings that the data does not actually support. That single output is often the most valuable thing it produces.
Where it fails, and why
A language model does not do arithmetic — it predicts text. When it “calculates” a growth rate inline, it is pattern-matching, and it will state a wrong figure with full confidence. This is not a small effect. On Vals AI’s Finance Agent v2 benchmark, which scores models on the work of entry-level financial analysts answering questions on real public filings, the top three as of June 2026 sit within a point of each other — GPT-5.5 at 51.76%, Claude Opus 4.7 at 51.51%, and Claude Sonnet 4.6 at 51.03%. Categories like earnings retrieval clear 70%, but anything requiring multi-step modeling tops out near 23%.
The practical rule: AI cannot separate signal from noise on a short window. Give it a 3-month view and it will call any 8% swing a “trend.” Always provide at least one full year of history, and never let a model-quoted number reach a slide without re-deriving it yourself.
What to feed the model
| Input | Why it matters |
|---|---|
| Monthly metric table — revenue, COGS, gross margin, operating margin | Pre-aggregated so the model reasons over ~12-24 rows, not thousands |
| 12 months minimum, 24 if you have it | One year is the floor for distinguishing seasonality from a real shift |
| One-time events tagged by month (price change, channel launch, layoff) | Stops the model from inventing a “mystery” cause for a known event |
| The specific hypothesis you want pressure-tested | Forces a falsifiable answer instead of a vague summary |
Do not paste raw transactional rows. Models lose precision on long tables and will silently mis-sum them. If your data lives in line items, aggregate to monthly first — or hand the file to a code tool (next section).
The prompt
Analyze trends in this monthly financial data. Do not estimate or round —
if a calculation is needed, show the arithmetic so I can verify it.
Data: [paste full monthly table]
Tagged events: [paste event list with months]
Window: [number of months]
Hypothesis to test: [your hypothesis]
Return:
1. Three trends, each with the specific months that drive them.
2. Three anomalies, each with a proposed cause.
3. Three falsifiable hypotheses, ranked by how easy they are to verify.
4. One "trend" the data does NOT support that I should stop quoting, and why.The “show the arithmetic” instruction matters more than it looks. When the model writes out the steps, a wrong number becomes catchable in seconds instead of slipping through as a clean-looking total.
Use a code tool for anything non-trivial
For files over ~200 rows, or whenever the answer depends on a precise total, route the data through a tool that runs real Python instead of guessing:
- ChatGPT Advanced Data Analysis (Plus $20/mo, Pro $100 or $200/mo as of June 2026) executes pandas in-session — it loads your CSV, computes the aggregates, and you can ask to see the code.
- Claude’s Analysis tool does the same; the chat interface accepts CSV/XLSX uploads up to 30MB and runs the math in a sandbox. Sonnet 4.6 (Pro $20/mo) is the cost-sensible workhorse here; Opus 4.7 (Max plans) when reasoning depth matters.
In both, the model writes and runs code, so the numbers come from an interpreter, not from token prediction. You still verify — but the failure rate drops sharply.
How to validate the output
- Pick one trend and re-derive it from the source spreadsheet by hand or formula. If the model’s figure is off by more than ~1%, assume it hallucinated the calculation — discard the whole output and re-run with cleaner, pre-aggregated data.
- Check every “anomaly” against a known event before repeating it. Half of what models flag as anomalies are the one-time events you forgot to tag.
- Confirm with the finance lead before any AI-surfaced claim becomes a meeting talking point.
Common mistakes
- Pasting raw transactional rows and trusting the totals — aggregate to monthly, or use a code tool.
- Repeating a surfaced “anomaly” without manual verification.
- Asking for trends on a 3-month window — too short to mean anything.
- Skipping the “trend you should stop quoting” output — that line is where bad meeting conclusions die.
- Pasting customer-identifying rows into a consumer chat tier.
Keep improving the workflow
Save the prompts that work as templates per metric type. Each quarter, append the new quarter’s actuals as one more column and rerun. Over a few cycles you build a versioned trend log that catches reversals before they become surprises in front of leadership.
FAQ
- Can I just paste a CSV? For roughly 200 rows or fewer, yes. Beyond that, either aggregate to monthly first or upload the file to a code tool (ChatGPT Advanced Data Analysis or Claude’s Analysis tool) so the math runs in Python rather than in the model’s head.
- Can AI replace a financial analyst? Not as of June 2026. On the Vals AI Finance Agent v2 benchmark, no frontier model clears 52% on entry-level analyst tasks. AI compresses the grind — theming, drafting, anomaly-spotting — but a human still has to sign off on every number.
- What about sensitive data? OpenAI does not train on ChatGPT Enterprise, Team, or API inputs by default; the same no-training default applies to Claude’s Team and Enterprise plans. For consumer tiers, use Temporary Chat or disable model training in settings, and never paste card numbers, bank details, tax IDs, or customer-identifying rows.
- How do I phrase a falsifiable hypothesis? Combine a metric, a direction, and a window: “If Product B drove the lift, week-4 retention should be above 60%.” Then go check it against the data.
- Which model is best for this? They’re effectively tied on finance accuracy (within one point on Finance Agent v2), so choose on workflow: GPT-5.5 if you live in ChatGPT, Sonnet 4.6 for low-cost runs, Opus 4.7 when you want the deepest reasoning on a messy table.
Related
External reference: the Vals AI Finance Agent v2 leaderboard is worth bookmarking — it tracks how current models actually perform on real analyst tasks.
Tags: #Data analysis #Workflow #Research