The Grandmother Test

Should I let an AI answer my team's data questions?

For curiosity, yes — and it will save your analyst a great deal of time. For anything someone will act on, only with guardrails: written definitions it must follow rather than guess, a known-answer test set you re-run monthly, and a rule about which questions must reach a human first. Today's data agents have senior-level technical execution and junior-level business judgment.

Gergana Tyaneva · 21 September 2026 · 12 years in product and marketing analytics

The grandmother version

A brilliant new assistant starts on Monday. Very fast, very confident, writes beautifully.

They have never worked here. They don't know that the good plates are in the second cupboard, that the scales in the kitchen read heavy, or that the tin marked "sugar" has been salt since 2019.

You'd happily let them answer the phone. You wouldn't let them send out the invoices in week one without someone checking — not because they're careless, but because being confident and being right are different things.

Where to point it

Nobody is deciding anything. Curiosity, learning, exploring — let people answer their own questions instead of queueing behind your analyst. This is where most of the value is.

The metric and the decision are already settled. The agent does the build, a human signs off.

Two definitions are both defensible. Have it compute each one and use the outputs to force the alignment conversation that should have happened months ago.

A data question arrivesWill anyone act on it?noLet the agent answerthis is most of the valueyesIs the definition settled?Agent builds, human signs offknown-answer test set, monthlynoHuman firstsettle it, then automateThe model is not what makes this safe. The definitions and the test set are.

Where not to

The metric measures nothing anyone acts on. Automating work that didn't matter produces more work that doesn't matter, faster.

People will act and the definition is disputed. A human goes first, always.

It would publish numbers or rewrite shared definitions unreviewed. A wrong number adopted here cascades into every analysis downstream, and nobody remembers where it came from.

What actually makes it safe

Not the model. Four things: your definitions written as rules it must follow rather than infer; your vocabulary, including which tables are dead; hard limits — read-only by default, no publishing, no writes to shared definitions; and a known-answer test set of 20–30 questions with verified answers, re-run monthly to prove it is still right.

Plus one written escalation rule: which question types must reach a human before anyone acts. Anything touching money, headcount, public numbers or a live experiment.

The short version

The technical work is solved. The judgment is not, and that is the whole job.

If your metric definitions do not exist yet, the honest first project is writing them — not buying an agent that will guess at them very quickly.

AI Analytics Setup — from €12,000, after a €1,500 assessment credited against the build.

See the flagship →

More answers

Attribution — Which channel actually pays?Plumbing — Our numbers live in eight tools and no two agreeTime — Month-end reporting eats three days and nobody reads itMoney — We know people churn. We don't know who, when, or what it's worth

One question before you read on. I'd like to count visits with Google Analytics, which sets cookies and sends your IP to Google. Nothing has loaded yet, and the site works exactly the same either way. What's collected.