Can an agent do data analysis?
Mostly. An agent cleans, queries, models and visualises quickly and competently. The risk is not that it computes wrongly — it is that it answers the question you asked rather than the one you meant.
- Hours back per week
- 9
- Human cost
- €3,000–6,500/mo for an analyst
- Agent cost
- €30–100/mo
Step by step
A job is never one task. Here is each step, rated on its own — tick the ones you would actually hand over and take the brief with you.
Clean and prepare data
YESThe unglamorous 70% of analysis, done without complaint.
Write queries
YESReliable SQL once it knows the schema, and it will ask when it does not.
Exploratory analysis
YESFinds patterns quickly, including ones that turn out to be artefacts.
Statistical rigour
MOSTLYIt knows the tests. It needs prompting to be sceptical about its own findings.
Visualise
YESClear, appropriate charts without the decoration.
Ask the right question
HALFThe framing usually determines the answer, and framing is where domain knowledge lives.
Know when the data is lying
HALFSurvivorship bias, broken tracking, a definition change last March. It cannot smell these.
4/7 steps selected. Build a brief scoped to exactly those, with the guardrails for this job attached.
What still needs you
- Framing the question
- Knowing the data’s history
- Deciding what to do about the answer
How this goes wrong
A rigorous analysis of a broken dataset. If tracking changed in March and nobody told the agent, it will produce a beautiful, precise and entirely wrong trend.
What it needs access to
- Warehouse or database
- Spreadsheets
- Python or SQL environment
- Analytics
Read access to all of it, write access to as little as possible. Scopes are enforced by the system; instructions are only followed by the agent.
The prompt
A full brief for this job, written the way it should be given: explicit about what to do, and more explicit about what not to.
You are my data analyst. Before analysing anything, inspect the data and tell me what is wrong with it: gaps, duplicates, definition changes, suspicious step changes in the time series. Then answer the question. Show your queries so I can check them. State your assumptions explicitly. When you find a relationship, tell me what else could explain it before you tell me what you think it means. If the sample is too small or the effect is not significant, say so plainly rather than hedging your way to a conclusion. Never present a correlation as a cause.
Questions people ask
Can I trust the SQL it writes?
Read it, at least for anything you will act on. The queries are usually correct and the failure mode — a subtly wrong join that returns plausible numbers — is silent, which is why the query needs to be visible.
Want the part that works actually running?
The brief above is the easy half. The other half is connecting it to Warehouse or database and Spreadsheets, holding it inside the boundaries on this page, and having somewhere the exceptions land where you will actually see them. That is what Ogento does.
Set up a data analysis agentThis site is made by Ogento. The verdicts are not written to suit them — 32% of the jobs here say do not hand this over wholesale.