ZeroTwo for analysts

AI for analysts: check the data before you chart it

Most of an analyst's day is the mechanical part of thinking: pulling, cleaning, cross-checking and formatting. This page shows how to hand that part to ZeroTwo and keep the method, the arithmetic and the interpretation. It includes a worked example you can redo by hand and 12 prompts to run on sanitised data.

Have one dataset to analyse rather than an evidence trail to build? Start with AI agent for data analysis.

What the first prompt asks for

A report on the file, before any chart

  • Row count
  • Missing values by column
  • Duplicate rows
  • Outliers
  • Base size of every segment

Then you decide whether the file is fit to analyse. No conclusion is drawn first.

Worked example

Explain a revenue change you can check on a napkin

The variance prompt asks for the decomposition and the arithmetic. Here is what that should look like, on made-up data small enough to redo by hand.

Revenue by segment in two periods, split into volume and price effects. Synthetic figures for illustration.
SegmentUnits (before to after)Price (before to after)Revenue changeVolume effectPrice effect
Segment A400 to 380$50 to $52−$240−$1,000+$760
Segment B200 to 260$80 to $80+$4,800+$4,800$0
Segment C100 to 40$120 to $125−$7,000−$7,200+$200
Total$48,000 to $45,560mixed−$2,440−$3,400+$960

Synthetic data we made up to show the method. Not a ZeroTwo output and not a real result. Check: −$3,400 + +$960 = −$2,440, and $45,560 − $48,000 = −$2,440.

Read it the way a reviewer would: Segment C's lost volume is the largest single effect, and Segment B's growth only partly offsets it. The table cannot say why C's volume fell, whether demand, stock or a change in how it is counted. That is the part the prompt must flag instead of narrating.

  1. Split the movement by segment

    Take each segment's change in revenue on its own. Nothing is explained until the pieces add back to the total.

  2. Separate volume from price

    Volume effect = change in units x last period's price. Price effect = change in price x this period's units. The two always sum to the segment's change.

  3. Rank, then name what is unattributed

    Order the effects by size. Then say what the numbers cannot tell you. Here they show that Segment C's volume fell, not why.

Small bases

Thirty is a flag, not a rule

The prompts ask ZeroTwo to flag segments whose base is too small to support the comparison, and to say which threshold it used. The right cutoff depends on your method and the precision you need. Use the flag to decide where to look, then apply your own standard.

Approximate 95 percent margin of error for a proportion near 50 percent at four sample sizes.
Respondents in the segmentMargin of error on a 50% result
30±17.9 points
100±9.8 points
400±4.9 points
1,000±3.1 points

Textbook formula, 1.96 × √(0.25 / n), for a simple random sample at 95% confidence. Weighting and survey design make real margins wider. At a base of 30 a 50% result could plausibly sit anywhere from about 32% to 68%.

12 prompts

Prompts in the order the work happens

Each card opens in ZeroTwo with the prompt already written. File names are placeholders; attach your own sanitised files.

Data quality

Run before anything is charted or concluded.

How do I check a dataset before I analyse it?

Find the gaps, duplicates and small bases first, before any conclusion is built on them.

The prompt

Before any analysis, profile this export: row count, missing values by column, duplicates, outliers, and any segment whose base is too small to support the comparison (say which threshold you used and why). List what you find before drawing a single conclusion.

Typically attach: Export_Q3.xlsx (placeholder name, use your own file)

Open this prompt in ZeroTwo

Trust

Challenge the claim and the sources before they reach a slide.

Does my data actually support the claim I want to make?

Ask directly, and ask for the counter-case. A dataset that only confirms you is usually being read selectively.

The prompt

I want to argue that [claim]. Using only this dataset, tell me honestly whether it supports that, what the strongest counter-reading is, and what would have to be true for my claim to hold.

Typically attach: Dataset.csv (placeholder name, use your own file)

Open this prompt in ZeroTwo

How do I check the sources in an AI-written summary?

Verify each citation resolves to something retrievable. Check all of them the first time, then sample.

The prompt

Go through this summary and check every citation points to a real, retrievable source. List the ones that do not, and do not silently replace them.

Typically attach: Summary_draft.docx (placeholder name, use your own file)

Open this prompt in ZeroTwo

What do I do when two sources give different numbers?

Show the disagreement and judge which is more credible. Averaging produces a figure nobody will defend.

The prompt

These sources report different figures for the same measure. Show the discrepancy, explain the likely reason for it, and say which is more credible and why. Do not average them.

Open this prompt in ZeroTwo

Analysis

Arithmetic and synthesis that someone else can re-run.

How do I explain a change in the numbers?

Decompose the movement before narrating it, so the explanation follows the arithmetic rather than the other way round.

The prompt

This metric moved between the two periods. Decompose the change into its contributing factors, show the arithmetic, and rank the factors by how much of the movement each explains. Flag anything the data cannot attribute.

Typically attach: Period_compare.xlsx (placeholder name, use your own file)

Open this prompt in ZeroTwo

How do I sanity-check a model someone else built?

Trace the assumptions and find the ones the output is most sensitive to.

The prompt

Read this model. List its assumptions, say which cells drive the output most, and identify anything hard-coded that looks like it should be an input. Do not change the file.

Typically attach: Model_v6.xlsx (placeholder name, use your own file)

Open this prompt in ZeroTwo

How do I synthesise a pile of research quickly?

Theme it with sources attached, and keep the outliers rather than smoothing them away.

The prompt

Synthesise these documents into themes. Attach the source and date to every claim, keep contradictions visible as their own list, and flag outlier findings rather than dropping them.

Typically attach: Research_pack.pdf (placeholder name, use your own file)

Open this prompt in ZeroTwo

Exhibits

Charts that state their base and admit what they cannot show.

What belongs in a chart footnote?

Source, base, date and method, plus a label on any segment too small to stand alone.

The prompt

Turn this into a chart. State the base for every figure, label any segment too small to support the comparison as indicative (say which threshold you used and why), and write the footnote with source, base, date and method.

Typically attach: Segments.xlsx (placeholder name, use your own file)

Open this prompt in ZeroTwo

Which chart type should I use for this?

Pick from what the data can support, not from what looks impressive.

The prompt

Given this data and the point I need to make, recommend a chart type and say what it would mislead about. If the data cannot support the comparison I want, say so instead of drawing it.

Open this prompt in ZeroTwo

Reporting

Write it up so it can be challenged, and re-run next cycle.

How do I write a brief a stakeholder can challenge?

Separate what you know, what you infer, what you do not know, and what you recommend.

The prompt

Write this up as a brief structured into what we know, what we think, what we do not know, and what we recommend, kept strictly separate. Every factual claim carries an inline source and date.

Open this prompt in ZeroTwo

How do I rerun the same analysis next month?

Save the method, not just the output, so the second run costs less than the first.

The prompt

Write down the method used here as a repeatable procedure: inputs required, steps, checks, and the format of the output. I will run this again next cycle against new data.

Open this prompt in ZeroTwo

How do I explain this analysis to a non-technical audience?

Translate the finding without softening the caveats that make it defensible.

The prompt

Rewrite this finding for an audience with no statistical background. Keep every caveat that affects how the result should be used, and do not round away uncertainty.

Typically attach: Findings.docx (placeholder name, use your own file)

Open this prompt in ZeroTwo

Division of work

ZeroTwo stops at the judgment

The assembly goes to ZeroTwo. The parts you would have to defend in the room stay yours.

ZeroTwo does

  • Gather and organise the inputs
  • Extract, theme and cross-reference
  • Keep each claim tied to its source
  • Surface gaps and contradictions
  • Format to your conventions

You keep

  • Choose the method and the threshold
  • Verify the arithmetic
  • Decide what the numbers mean
  • Write the recommendation
  • Put your name on it

It is not the right tool when the output has to leave without a human reading it, when your data-handling terms do not cover AI processing, or when the job is building a financial model from scratch. ZeroTwo can review a model's assumptions, but it does not replace the spreadsheet.

Which page is yours

Pick the starting point that matches the job

Turning findings into a document? The AI report generator and document analysis pages cover those tasks.

FAQ

Questions analysts ask first

Will it check my data or just chart it?

Only if you ask it to check first, and the first prompt on this page does. It asks for row count, missing values, duplicates, outliers and segment bases before drawing any conclusion, and you decide whether to proceed. Read that report against the file yourself before you trust the chart.

Can I trust the numbers it produces?

Treat every figure as needing your verification. The prompts ask for the calculation to be shown and the base to be stated so the arithmetic is checkable. Recompute at least one figure by hand before it goes into an exhibit.

How is this different from an AI data analysis tool?

A task-level tool analyses a dataset you hand it. This page is organised around the analyst's whole evidence trail: profiling inputs, synthesising research, building the exhibit and writing it up so someone else can challenge it. For the single task itself, see AI agent for data analysis.

Do I have to upload client or company data?

No. Run the first pass on public, synthetic or sanitised data to judge whether the output meets your standard. Check your employer's or client's data-handling terms before real data goes anywhere.

Will it keep our conventions between runs?

The recurring-analysis prompt writes the method down as a procedure with inputs, steps, checks and output format. Keep that with your files in a ZeroTwo project and run it again against new data. Treat the saved procedure as a document to review, not as something that updates itself.

Does this replace the analyst?

It takes on the extraction and formatting that delay the thinking. Method choice, interpretation and anything you would have to defend stay with you.

Start with the last dataset you charted too quickly.

Run the check on sanitised data first. The Free plan includes 100 free credits to start, and the pricing page lists what paid plans add.

Open ZeroTwo free