AI agent for data analysis
Analyze data with visible methods and checkable results
Good analysis starts with the question and the meaning of each row. Provide a bounded dataset, a data dictionary and a decision to support. Then inspect the transformations, denominators and calculations behind the findings before using them to choose an action.
- 01QuestionThe decision the numbers must inform.
- 02GrainWhat exactly one row represents.
- 03DenominatorWhich rows count, and which are excluded.
- 04MethodThe formula or code behind each figure.
- 05LimitsWhat this data cannot establish.
Define the dataset, grain and decision.
Most wrong numbers come from a misread row, not from bad arithmetic. Three inputs prevent most of it.
A bounded dataset
Attach the relevant CSV or supported table. Remove columns the question does not need and anything you are not authorized to share.
A data dictionary
Define the row grain, each field, how missing values are marked, the time window and the denominators you expect.
A decision question
State the comparison or action the analysis should inform, so the agent can pick calculations that match it.
What data can this workflow analyze?
A CSV or other supported table is the starting point. This page does not publish row limits, file-size limits or the calculation engine behind a given task, so start with a small extract and confirm the behavior on your own file.
Files and plans
| Plan | Work agent | Files |
|---|---|---|
| Free$0/mo | No | 1 per message |
| Plus$14.99/mo | Yes | Multiple per task |
| Pro$29.99/mo | Yes | Multiple per task |
The Free plan starts with 100 free credits and adds 20 bonus credits on each day you log in. See pricing for the full comparison.
For work inside a workbook, such as formulas, tabs and cleanup, the spreadsheet-specific analysis page is the better fit. If the data lives in a database instead, see the database analysis workflow for what that path supports.
What this workflow is for
Describe
Counts, rates, segments, distributions and trends, with their calculation basis shown.
Exploratory only
Predict
Projections extend patterns in the past. Treat them as hypotheses and validate them against held-out data.
Not from rows alone
Explain why
Observed association is not cause. A causal claim needs a designed comparison, not a table of existing rows.
Worked setup: explain a change in retention
This is a synthetic setup for showing what to define before the first calculation, not a recorded run. No results are shown because none were produced here. The letters in the ledger stand for counts you would fill in from your own file.
| Field | Meaning | Trap to settle first |
|---|---|---|
| user_id | One person. Can repeat across rows. | Counting rows instead of people |
| signup_date | Date the account was created. | Blank for some rows |
| last_active_date | Most recent activity in the window. | Blank can mean never active or not recorded |
| plan | Plan at the time of the snapshot. | Plan may have changed since signup |
| region | Self-reported region. | Free text, inconsistent spellings |
| Step | Rows | Rule to record |
|---|---|---|
| Rows in the file | A | Counted before anything is touched |
| Exact duplicate rows | − B | Removed and counted |
| Repeated user_id, different rows | − C | Keep one per person; state which and why |
| Outside the date window | − D | Window stated before the analysis starts |
| Missing signup_date | − E | Excluded or filled? Say which |
| Rows analyzed | A − B − C − D − E | The denominator. It must reconcile |
The last line is the point. If A minus every exclusion does not equal the rows analyzed, some rows were dropped or double-counted without being reported, and every rate built on that denominator is suspect.
Inspect the analysis output
Every important number should lead back to a calculation. These are the pieces to expect, and a three-step check you can run on any of them.
Result table
Each figure next to its numerator, denominator and the formula that produced it.
Chart you can read
Labeled axes, a stated time window, and values that match the table underneath.
Narrative kept separate
Interpretation sits apart from the numbers, so you can disagree with one without losing the other.
Action table
Each recommended action tied to a rule you can state and challenge.
Open questions
Exclusions, assumptions and anything the data could not settle.
Recompute one number yourself
- Pick one number that would change your decision.
- Recompute it yourself from the same rows, in a spreadsheet or a short script.
- Compare it to the reported figure. If they differ, the method is the first suspect.
Choose the workflow shape that matches the assignment
What matters is whether you can see the method, repeat it and audit it. These are general approaches, and each can be the right one.
| What you need | Narrated summary only | Spreadsheet by hand | Agent that runs and shows calculations |
|---|---|---|---|
| Method visible | Usually not. The text describes results, not steps | Yes. Every formula is in a cell | Should be. Ask for formulas or code per figure |
| Recalculation | Re-ask and hope the answer matches | Change an input and the sheet updates | Re-run on a changed input and compare |
| Duplicates and missing values | Easy to skip without saying so | Handled if you remember to handle them | Ask for counts and a stated rule for each |
| Charts and export | Description of a chart, not the chart | Built-in charts, full control | Depends on the task. Confirm the output before relying on it |
| Speed on a new question | Fast | Slow for unfamiliar data | Faster than by hand, if you still check it |
| Best for | A first look at what is in a file | Small, stable models you will maintain | Ad hoc questions where the denominator matters |
Our take: ZeroTwo's Work agent is built for the third shape, and ZeroTwo publishes this page. We have not published a measured comparison, so use the recompute check above to judge any tool on your own data.
Where the analysis remains uncertain
A careful analysis says where it stops. If a result reads more certain than the data allows, push back.
Small samples
A rate from a handful of rows moves a lot with one row. Ask for the count behind every rate.
Selection bias
If the rows are not a fair sample of the people you care about, the result describes only the rows.
Missing data
Dropping blanks changes who is counted. The result should say how many rows were dropped and why.
Association is not cause
Two things moving together can share a third driver. Expect the analysis to say so rather than recommend as if it were proven.
Reuse the method on your own data
Send the smallest file that answers the question, and keep the definitions with the result.
Input checklist
- One row grain, stated in a sentence.
- Only the columns the question needs.
- A date window and the rule for blanks.
- The decision the result will inform.
Keep the upload small
Remove personal identifiers and anything you are not authorized to share, and mask what the analysis does not need. Read the privacy policy before uploading sensitive data.
What people ask
Does the agent execute calculations or only describe the data?
A narrated summary and an executed calculation are different things, and this page does not publish which calculation tools run for a given task. Treat any figure you plan to act on as unverified until the agent shows the formula or code behind it and you have recomputed one number yourself, for example in a spreadsheet.
How are duplicate and missing rows handled?
They should be handled by a rule you can read, not silently. Ask for the count of exact duplicates, the rule for repeated IDs, the date window, and what happens to rows with missing key fields, and check that the ledger of rows reconciles from the file total to the analyzed total. If a rule was not stated, ask for it.
Can I inspect formulas or code?
Ask for them. For every number that matters, request the formula or code, the input rows it used and its denominator. If the agent cannot show how a figure was produced, do not use that figure for a decision.
Which file sizes and formats are supported?
The spreadsheet workflow takes a CSV or other supported table, and the Free plan allows one file upload per message while paid plans allow multiple files per task. This page does not publish row limits, file-size limits or a full format list, so check docs.zerotwo.ai or try a small extract first.
Can the output establish causation?
No. Rows from an existing dataset can show association, not cause. The agent can describe what moved together and flag confounders, but a causal claim needs a designed comparison. For consequential decisions, have a qualified analyst review the work.
Explore related workflows
When the work is inside a workbook: formulas, tabs and cleanup.
When the data lives in a database, not a file.
After the calculations are checked and the story is settled.
The wider set of tasks analysts bring to ZeroTwo.
For repeatable metric reports instead of one-off analysis.
Put one real question to a bounded dataset
The Work agent is included on paid plans, starting at $14.99/month for Plus. The Free plan does not include it.