Analyst tasks you should not hand to an LLM, and why

The four-question test

Run any task through these questions before you give it to a model. One "no" is enough to keep it with you.

1. 1Can I detect an error quickly? If a wrong output looks exactly like a right one, you are relying on luck.
2. 2Is the data allowed to go there? Your company's approved-tool list and data classification decide this, not convenience.
3. 3Would I be comfortable saying "the AI decided" to my manager? If not, the decision is yours.
4. 4Is the cost of an error small? A clumsy email draft is cheap. A wrong revenue figure in a board pack is not.

Seven red lines, with the failure each prevents

1. Producing the final number

A model can write the query. It should not be the thing that adds up the column you report. Language models generate text token by token; when asked to total a pasted table they can return a figure that is close but wrong, with no warning. Run the calculation in SQL, Excel or Python, where the arithmetic is deterministic, and let the model help only with the code.

A quick demonstration you can run shows why "close" is not acceptable. The data is synthetic:

from decimal import Decimal

# Synthetic invoice lines, in rupees
lines = ["18450.50", "7299.00", "112000.75", "3499.99", "64820.10"]
exact = sum(Decimal(x) for x in lines)
plausible_wrong = Decimal("206070.43")  # digits transposed: the kind of near-miss a model can produce

print("Exact total:", exact)                        # 206070.34
print("Difference:", plausible_wrong - exact)       # 0.09
assert exact == Decimal("206070.34")
assert exact != plausible_wrong

The near-miss differs by nine paise. Nobody would spot it by eye, and it would still fail a reconciliation.

2. Anything with customer or employee personal data in an unapproved tool

Pasting a CRM export with names and phone numbers into a public chatbot moves personal data outside your organisation's control. India's Digital Personal Data Protection Act, 2023 places obligations on the organisation processing such data, and your employer's policy will reflect that. Mask, aggregate, or use only the tool your company has approved. See DPDP Act basics for data analysts for what this means in daily work.

3. Defining a metric

"Active user", "churned customer" and "on-time delivery" are agreements between teams. A model will happily produce a definition, and it may be reasonable, but it will not be your organisation's definition, and nobody will have agreed to it. When the finance and sales numbers disagree next quarter, a model-authored definition has no owner.

4. Claiming a cause

Ask a model why sales dropped and it will give fluent reasons. It has not seen your promotions calendar, your stock-outs or the new competitor store. Use its list as brainstorming, then check each item against data. Never paste its explanation into a report as a finding.

5. Signing off a report

Sign-off means "I have checked this and I stand behind it." That cannot be delegated to a tool, because the tool bears no consequence. If you used a model in the process, your sign-off covers its output too.

6. Deciding what to leave out

Choosing which caveats to include and which segment to exclude as an outlier are judgement calls with ethical weight. A model optimising for a clean narrative may drop the awkward caveat. That caveat is often the most important sentence.

7. The difficult conversation

Telling a sales head that their region's numbers were overstated for three months is a trust conversation. A model can help you rehearse it. It should not write the message you send, because the recipient will read tone, and generic tone reads as evasive.

What you can safely delegate instead

The same tasks have delegable pieces:

Say this in an interview

When an interviewer asks how you use AI, a strong answer names a red line. Something like: "I use it to draft queries and documentation, but the final numbers always come from a query I've read and reconciled, and I never paste customer data into a tool that isn't approved." That tells them you understand accountability, which is what they are really checking.

If you are still sorting which tasks go where, the companion article on how GenAI changes an analyst's day sorts a full week into three buckets. Verification before trust is built into the Data Analyst programme's GenAI modules.